Assessing Game Balance With Alphazero: Exploring Alternative Rule Sets in Chess

Assessing Game Balance with AlphaZero:
Exploring Alternative Rule Sets in Chess
Nenad Tomašev* Ulrich Paquet* Demis Hassabis Vladimir Kramnik

DeepMind DeepMind DeepMind World Chess Champion
2000–2007§
arXiv:2009.04374v1 [cs.AI] 9 Sep 2020
Abstract 1. Introduction
Rule design is a critical part of game development, and
It is non-trivial to design engaging and balanced small alterations to game rules can have a large effect on a
sets of game rules. Modern chess has evolved over game’s overall playability and the resulting game dynam-
centuries, but without a similar recourse to history, ics. Fine-tuning and balancing rule sets in games is often
the consequences of rule changes to game dynam- a laborious and time-consuming process. Automating the
ics are difficult to predict. AlphaZero provides an balancing process is an open area of research (Jaffe et al.,
alternative in silico means of game balance assess- 2012; de Mesentier Silva et al., 2017), and machine learn-
ment. It is a system that can learn near-optimal ing and evolutionary methods have recently been used to
strategies for any rule set from scratch, without help game designers balance games more efficiently (An-
any human supervision, by continually learning drade et al., 2005; Leigh et al., 2008; Halim et al., 2014;
from its own experience. In this study we use Grau-Moya et al., 2018). Here we examine the potential of
AlphaZero to creatively explore and design new AlphaZero (Silver et al., 2018) to be used as an exploration
chess variants. There is growing interest in chess tool for investigating game balance and game dynamics un-
variants like Fischer Random Chess, because of der different rule sets in board games, taking chess as an
classical chess’s voluminous opening theory, the example use case.
high percentage of draws in professional play, Popular games often evolve over time and modern-day chess
and the non-negligible number of games that end is no exception. The original game of chess is thought to
while both players are still in their home prepara- have been conceived in India in the 6th century, from where
tion. We compare nine other variants that involve it initially spread to Persia, then the Muslim world and later
atomic changes to the rules of chess. The changes to Europe and globally. In medieval times, European chess
allow for novel strategic and tactical patterns to was still largely based on Shatranj, an early variant orig-
emerge, while keeping the games close to the inating from the Sasanian Empire that was based on the
original. By learning near-optimal strategies for Indian Chaturaṅga (Murray, 1913). Notably, the queen and
each variant with AlphaZero, we determine what the bishop (alfin) moves were much more restricted, and
games between strong human players might look the pieces were not as powerful as those in modern chess.
like if these variants were adopted. Qualitatively, Castling did not exist, but the king’s leap and the queen’s
several variants are very dynamic. An analytic leap existed instead as special first king and queen moves.
comparison show that pieces are valued differ- Apart from checkmate, it was also possible to win by baring
ently between variants, and that some variants are the opposite king, leaving the piece isolated with the entirety
more decisive than classical chess. Our findings of its army having been captured. In Shatranj, stalemate was
demonstrate the rich possibilities that lie beyond considered a win, whereas these days it is considered a draw.
the rules of modern chess. The evolution of chess variants over the centuries can be
viewed through the lens of changes in search space complex-
ity and the expected final outcome uncertainty throughout
the game, the latter being emphasized by modern rules and
seen as important for the overall entertainment value (Cin-
*
Equal contribution cotti et al., 2007). Modern chess was introduced in the
§
Classical (2000–2006); FIDE and Undisputed (2006–2007) 15th century, and is one of the most popular games to date,
Assessing Game Balance with AlphaZero
captivating the imagination of players around the world. Shane and Gawain Jones played the first-ever grandmaster
No-castling match during the London Chess Classic. This
The interest in further development of chess has not sub-
was followed up by the very first No-castling chess tourna-
sided, especially considering a decreasing number of de-
ment in Chennai in January 2020, which resulted in 89%
cisive games in professional chess and an increasing re-
decisive games (Shah, 2020).
liance on theory and home preparation with chess engines.
This trend, coupled with curiosity and desire to tinker with
such an inspiring game, has given rise to many variants of 2. Methods
chess that have been proposed over the years (Gollon, 1968;
In this section we motivate nine alterations to the modern
Pritchard, 1994; Wikipedia, 2019). These variants involve
chess rules, describe the key components of AlphaZero
alterations to the board, the piece placement, or the rules,
that are used in the analysis in Section 3, and outline how
to offer players “something subtle, sparkling, or amusing
AlphaZero was trained for Classical chess and each of the
which cannot be done in ordinary chess” (Beasly, 1998).
nine variants.
Probably the most well-known and popular chess variant is
the so-called Chess960 or Fischer Random Chess, where
pieces on the first rank are placed in one of 960 random 2.1. Rule Alterations
permutations, making theoretical preparation infeasible. There are many ways in which the rules of chess could be
Chess and artificial intelligence are inextricably linked. Tur- altered and in this work we limit ourselves to considering
ing (1953) asked, “Could one make a machine to play chess, atomic changes that keep the game as close as possible to
and to improve its play, game by game, profiting from its classical chess. In some cases, secondary changes needed
experience?” While computer chess has progressed steadily to be made to the 50-move rule to avoid potentially infinite
since the 1950s, the second part of Alan Turing’s question games. The idea was to try to preserve the symmetry and
was realised in full only recently. AlphaZero (Silver et al., the aesthetic appeal of the original game, while hoping to
2018) demonstrated state-of-the-art results in playing Go, uncover dynamic variants with new opening, middlegame or
chess, and shogi. It achieved its skill without any human endgame patterns and a novel body of opening theory. With
supervision by continuously improving its play by learn- that in mind, we did not consider any alterations involving
ing from self-play games. In doing so, it showed a unique changes to the board itself, the number of pieces, or their
playing style, later analysed in Game Changer (Sadler & arrangement. Such changes were outside of the scope of
Regan, 2019). This in turn gave rise to new projects like this initial exploration. Rule alterations that we examine are
Leela Chess Zero (Lc0, 2018) and improvements in exist- listed in Table 1. The variants in Table 1 are by no means
ing chess engines. CrazyAra (Czech et al., 2019) employs new to this paper, and many are guised under other names:
a related approach for playing the Crazyhouse chess vari- Self-capture is sometimes referred to as “Reform Chess” or
ant, although it involved pre-training from existing human “Free Capture Chess”, while Pawn-back is called “Wren’s
games. A model-based extension of the original AlphaZero Game” by Pritchard (1994). None have yet come under
system was shown to generalise to domains like Atari, while intense scrutiny, and the impact of counting stalemate as a
maintaining its performance on chess even without an exact win is a lingering open question in the chess community.
environment simulator (Schrittwieser et al., 2019). Alp- Each of the hypothetical rule alterations listed in Table 1
haZero has also shown promise beyond game environments, could potentially affect the game either in desired or unde-
as a recent application of the model to global optimisation sired ways. As an example, consider No-castling chess. One
of quantum dynamics suggests (Dalgaard et al., 2020). possible outcome of disallowing castling is that it would
AlphaZero lends itself naturally to the problem of finding result in an aggressive playing style and attacking games,
appealing and well-balanced rule sets, as no prior game given that the kings are more exposed during the game and
knowledge is needed when training AlphaZero on any par- it takes time to get them to safety. Yet, the inability to easily
ticular game. Therefore, we can rapidly explore different safeguard one’s own king might make attacking itself a poor
rule sets and characterise the arising style of play through choice, due to the counterattacking opportunities that open
quantitative and qualitative comparisons. Here we examine up for the defending side. In Classical chess, players usually
several hypothetical alterations to the rules of chess through castle prior to launching an attack. Therefore, such a change
the lens of AlphaZero, highlighting variants of the game that could alternatively be seen as leading to unenterprising play
could be of potential interest for the chess community. One and a much more restrained approach to the game.
such variant that we have examined with AlphaZero, No- Historically, the only way to assess such ideas would have
castling chess, has been publicly championed by Vladimir been for a large number of human players to play the game
Kramnik (Kramnik, 2019), and has already had its moment over a long period of time, until enough experience and
in professional play on 19 December 2019, when Luke Mc- understanding has been accumulated. Not only is this a long
2
Variant Primary rule change Secondary rule change

Castling is disallowed
No-castling -
throughout the game
Castling is disallowed
No-castling (10) -
for the first 10 moves (20 plies)
Pawn one square Pawns can only move by one square -
Forcing stalemate is a win
Stalemate=win -
rather than a draw
Pawns can move by 1 or 2 squares
Torpedo anywhere on the board. En passant can -
consequently happen anywhere on the board.
Pawns can move by two square
Semi-torpedo -
both from the 2nd and the 3rd rank
Pawns can move backwards
Pawn moves do not count
Pawn-back by one square, but only back to the
towards the 50 move rule
2nd/7th rank for White/Black
Pawns can also move laterally
Sideway pawn moves do not
Pawn-sideways by one square. Captures are
count towards the 50 move rule
unchanged, diagonally upwards
It is possible to capture
Self-capture -
one’s own pieces
Table 1. A list of considered alterations to the rules of chess.
process, but it also requires the support of a large number tionally outputs a scalar value v ∈ (−1, 1) which estimates
of players to begin with. With AlphaZero, we can automate the expected outcome of the game from position s.
this process and simulate the equivalent of decades of human
The two predictions in (1) are used in Monte Carlo tree
play within a day, allowing us to test these hypotheses in
search (MCTS) to refine the assessment of a board position.
silico and observe the emerging patterns and theory for each
The prior network p assigns weights to candidate moves at a
of the considered variations of the game.
“first glance” of the board, yielding an order in which moves
Figure 1 illustrates each of the variants with an example are searched with MCTS. The output v can be viewed as
position. a neural network evaluation function for position s. The
statistical estimates of the game outcomes after each move
2.2. Key components of AlphaZero are refined through MCTS, which runs repeated simulations
of how the game might unfold up to a certain ply depth.
AlphaZero is an adaptive learning system that improves In each MCTS simulation, fθ is recursively applied to a
through many rounds of self-play (Silver et al., 2018). It sequence of positions (or nodes) up to a certain ply depth
consists of a deep neural network fθ with weights θ that if they have not been processed in an earlier simulation. At
compute maximum ply depth, the position is evaluated with (1), and
(p, v) = fθ (s) (1) that evaluation is “backed up” to the root, for each node
adjusting its “action selection rule” to alter which moves
for a given position or state s. The network outputs a vec- will be selected and expanded in the next MCTS simulation.
tor of move probabilities p with elements p(s0 |s) as prior After a number of such MCTS simulations, the root move
probabilities for considering each move and hence each next that was visited (or expanded) most is played.
state s0 .1 If we denote game outcome numerically by +1,
for a win, 0 for a draw and −1 for a loss, the network addi-
2.3. Training and evaluation
1
We’ve suppressed notation somewhat; the probabilities are
technically over actions or moves a in state s, but as each action a We trained AlphaZero from scratch for each of the rule
deterministically leads to a separate next position s0 , we use the alterations in Table 1, with the same set of model hyperpa-
concise p(s0 |s) in this paper.
3
8
0Zrlka0s 8
rZ0ZkZ0s
7
o0Z0Zpo0 7
apo0lpop
6
0ZnZbZ0o 6
pZnobm0Z
5
ZpopZ0Z0 5
O0Z0o0Z0
4
0Z0O0ZnO 4
0OBZPZ0Z
3
Z0M0ANO0 3
Z0OPZNZ0
2
PO0LPOBZ 2
0ZNZQOPO
1
S0Z0ZKZR 1
S0A0J0ZR
a b c d e f g h a b c d e f g h
(a) An example from No-castling chess: This is a typical po- (b) An example from No-castling(10) chess: The play tends to
sition where both kings haven’t found immediate safety and be slower and more strategic, to allow for later castling. Here,
remain exposed into the middlegame. on the 11th move, Black castles at the very first opportunity
and White castles immediately after as well.
8
rZ0lkZ0s 8
0Z0Z0ZNZ
7
obo0Zpa0 7
Z0Z0Z0Z0
6
0o0o0Zpo 6
0Z0Z0Z0Z
5
m0m0o0Z0 5
Z0Z0ZKZk
4
0ZPZPZ0Z 4
0Z0Z0Z0Z
3
O0ZPANO0 3
Z0Z0ZNZ0
2
0OQZNOBO 2
0Z0Z0Z0Z
1
S0Z0J0ZR 1
Z0Z0Z0Z0
(c) An example from Pawn-one-square chess: Black just moved (d) An example from Stalemate=win chess: An endgame posi-
the knight to a5. In Classical chess this would seem counter- tion that would have been a draw in Classical chess is now a
intuitive due to the potential of playing the pawn to b4, forking win instead.
the knights. Here, however, the pawn cannot move to that
square in a single move, justifying the manoeuvre.
Figure 1. Examples of new strategic and tactical themes that arise in the explored chess variants. Figure 1e continues on the following
page.
4
8
0Z0Z0Z0Z 8
0ZrlrZ0j
7
Z0ZkZ0Zr 7
Z0ZnZ0a0
6
0ZpOpZ0Z 6
0o0o0Zpo
5
Z0O0Z0Z0 5
o0Z0oPZn
4
0ObO0Z0Z 4
PZ0ZNZ0Z
3
Z0A0ZpZ0 3
ZQZBA0OP
2
pZ0Z0JpZ 2
0O0Z0O0Z
1
Z0Z0S0Z0 1
Z0ZRJ0ZR
(e) An example from Torpedo chess: White needs to generate (f) An example from Semi-torpedo chess: The ability to rapidly
rapid counterplay, and does so with a torpedo move: b4-b6. advance pawns from the 3rd/6th rank enables Black the fol-
Black responds with Rh1, to which White promotes to a queen lowing energetic option: d6-d4, resulting in a forced tactical
with yet another torpedo move, b6-b8=Q. sequence. See Game AZ-19 in Appendix B.6 for details.
8
rZ0lka0s 8
rZbl0skZ
7
ZbZnZpop 7
ZpZ0Zpap
6
pZnZpZ0Z 6
0Z0MpmpZ
5
ZpopO0Z0 5
ZNZpA0Z0
4
0Z0O0O0Z 4
0ZPZ0Z0Z
3
O0O0ANZ0 3
Z0Z0Z0ZP
2
0O0ZNZPO 2
PO0Z0JPZ
1
S0ZQJBZR 1
S0ZQSBZ0
(g) An example from Pawn-back chess: Here, Black uses this (h) An example from Pawn-sideways chess: After sacrificing
possibility to challenge White’s central pawns, while opening the knight on f2 the previous move, Black utilises a sideways
up the diagonal for the b7 bishop, by a pawn-back move d5-d6. pawn move f7-e7 for tactical purposes, opening the f-file to-
wards the White king, while attacking the knight on d6.
8
rZbZ0akZ
7
opZ0ZpZ0
6
0ZnZrm0Z
5
l0Zpo0A0
4
0ZPZ0Z0O
3
O0Z0O0Z0
2
0OQM0OPZ
1
Z0JRZBZR
a b c d e f g h
(i) An example from Self-capture chess: a self-capture move
Rxh4 generates threats against the Black king.
Figure 1. (Continued from previous page.) Examples of new strategic and tactical themes that arise in the explored chess variants.
5
rameters. The models were trained for 1 million training which we construct from AlphaZero’s network priors (1) for
steps, with a batch size of 4096 and allowing for an average the first couple of moves and show that the breadth of open-
0.12 samples per position from self-play games. In order ing possibilities in each of these chess variants seems to be
to encourage exploration during training, a small amount inversely related to their relative decisiveness (Section 3.5).
of noise was injected in the prior move probabilities (1) be- Sections 3.6 and 3.7 highlight the difference in opening play
fore search, sampled from a Dirichlet Dir(0.3) distribution, according to the prior distributions of the variants. Rule ad-
followed by a renormalization step (Silver et al., 2018). Fur- justments, especially those affecting piece mobility, are also
ther diversity was promoted by stochastic move selection in expected to affect the relative material value of the pieces.
the first 30 plies of each of the training self-play games, by Finally, Section 3.8 provides approximations for piece val-
selecting the final moves proportionally to the softmax of ues in each of the variants, computed from a sample of
the MCTS visit counts. The remaining game moves from 10,000 fast-play AlphaZero games.
ply 31 onwards were selected as top moves based on MCTS.
Training self-play games were generated using 800 MCTS 3.1. Self-play games
simulations per move.
For each chess variant, we generated a diverse set of
The absence of baselines makes it hard to formally assess N = 10,000 AlphaZero self-play games at 1 second per
the strength of each model, which is why it was important move, and N = 1,000 games at 1 minute per move. The
to couple the quantitative analysis and metrics observed outcomes of the fast self-play games are presented in Figure
at training and test time with a qualitative assessment in 2a; the longer games follow in Figure 2b. As AlphaZero is
collaboration with Vladimir Kramnik, a renowned chess approximately deterministic given the same MCTS depth
grandmaster and former world chess champion. As the and number of rollouts, we promote diversity in games by
rule changes that are considered in this study are mostly sampling the first 20 plies in each game proportional to the
minor in practical terms, it is reasonable to assume that the softmax of the MCTS visit counts, followed by playing the
trained models are of similar strength, although it is equally top moves for the rest of the game.
reasonable to expect that some of them could be further fine-
tuned to account for the differences in game length and the In addition to that, we generated a set of N = 1,000 fast-
average number of legal moves that need to be considered play games from fixed starting positions arising from the
at each position. Given the nature of the study, the high Dutch Defence, Chigorin Defence, Alekhine Defence and
level of observed play in trained models, and the number of King’s Gambit for each of the variants, as further discussed
rule alterations considered, we decided not to pursue such in Section 3.3.
a potentially laborious process, as it would not alter any of The two sets of diverse self-play games are used in Section
the high-level conclusions that we present and discuss. 3.2 to compare the decisiveness of each variant, in Section
3.4 to analyse how many special moves are used, and in
3. Quantitative assessment Section 3.8 to estimate piece values across variants.
There are marked differences between the styles of chess A selection of these games is presented in Appendix B.
that arises from each of the rule alterations Aesthetically,
each variant has its own appeal, and we highlight them fur- 3.2. Expected scores and draw rates
ther in Section 4. Here we provide a quantitative comparison It is widely hypothesised that classical chess is theoretically
between variants, to complement the qualitative observa- drawn; that the odds π = (πwin , πdraw , πlose ) of white win-
tions. Using a large quantity of self-play games, we infer ning, drawing and losing are (0, 1, 0) at optimal play. We
the expected draw rate and first-move advantage for each determine how favourable for white or how “drawish” differ-
variant, expressed as the expected score for White (Section ent variants are by estimating the expected scores and draw
3.2). We then illustrate how the same opening can lead to rates at non-optimal play under the same conditions. We
vastly different outcomes under different chess variants in keep the conditions that chess variants are played against
Section 3.3, and that these opening-specific differences can themselves with AlphaZero fixed, like the move selection
differ from the aggregate differences across all openings. criteria or Monte Carlo Tree Search (MCTS) evaluation
An analysis of the utilisation of the newly introduced op- time.
tions made possible by the new rule alterations in Section
3.4 shows that the non-classical moves are used in a large The overall decisiveness in the generated game sets depends
percentage of games, often multiple times per game, in each on the time controls involved. We see in Figures 2a and
of the variants. This suggests that the new options are in- 2b that across all variations the percentage of drawn games
deed useful, and contribute to the game. We estimate the increases with longer thinking times, and longer thinking
diversity of opening play by looking at the opening trees times also affect the expected score for White, as shown in
Table 2. This suggests that the starting position might be
6
White wins Draw Black wins White wins Draw Black wins
Classical 772 8820 409 Classical 18 979 3

No-castling 1110 8441 449 No-castling 29 968 3
No-castling (10) 604 9002 394 No-castling (10) 11 985 4
Pawn one square 709 8891 400 Pawn one square 9 988 3
Stalemate=win 1000 8606 394 Stalemate=win 25 971 4
Torpedo 2086 7191 723 Torpedo 93 894 13
Semi-torpedo 1306 8103 591 Semi-torpedo 27 964 6
Pawn-back 532 9160 308 Pawn-back 6 990 4
Pawn-sideways 872 8815 313 Pawn-sideways 15 980 5
Self-capture 871 8783 346 Self-capture 17 981 2
(a) The game outcomes of 10,000 AlphaZero games played at (b) The game outcomes of 1,000 AlphaZero games played at 1
1 second per move for each different chess variant. minute per move for each different chess variant.
Figure 2. AlphaZero self-play game outcomes under different time controls. As moves are determined in a deterministic fashion given the
same conditions, diversity was enforced by sampling the first 20 plies in each game proportional to their MCTS visit counts. Across all
variations the percentage of drawn games increases with longer thinking times. This seems to suggest that the starting position might be
theoretically drawn in these chess variants, like in Classical chess, and that some of the variants are simply harder to play, involving more
calculation and richer patterns.
Variant Training 1sec 1min 3.2.1. I NFERENCE FOR GAME ODDS

Classical 54.1% 51.8% 50.8% To compare variants, we first infer the odds of their out-
No castling 55.7% 53.3% 51.3% comes under set playing conditions. For a given variant,
No castling (10) 52.5% 51.0% 50.4% let the game outcomes G be nwin wins and nlose losses for
Pawn one square 53.5% 51.6% 50.3% white, and ndraw = N − nwin − nlose draws. If we assume
Stalemate=win 54.9% 53.0% 51.1% a uniform Dirichlet prior on π and multinomial likelihood
Torpedo 57.0% 56.8% 54.0% for winning, drawing or losing, the posterior distribution is
Semi-torpedo 54.7% 53.6% 50.9% Dirichlet,
Pawn-back 53.0% 51.1% 50.1%
Pawn-sideways 54.8% 52.8% 50.5% p(π|G) = Dir(nwin + 1, ndraw + 1, nlose + 1) . (2)
Self-capture 54.2% 52.6% 50.8%
3.2.2. D RAW RATES
Table 2. Empirical score for White under different game conditions,
for each chess variant: self-play games at the end of model training, To compare the decisiveness of chess variants, we infer
1 second per move games, and 1 minute per move games. Diversity the probability that variant A has a lower draw rate than
in 1 second per move games and 1 minute per move games was variant B, given the games played G A and G B under the
enforced by sampling the first 20 plies in each game proportional same conditions:2
to their MCTS visit counts.
A B
p(πdraw < πdraw )=
ZZ
A B
p(π A |G A ) p(π B |G B ) dπ A dπ B .

I πdraw < πdraw
(3)
theoretically drawn in these chess variants, like in Classical
chess, and that some of the variants are simply harder to The integral is not available in closed form; we evaluate it
play, involving more calculation and richer patterns. We with a Monte Carlo estimate by drawing pairs of samples
hypothesise that the relative differences in AlphaZero’s win from p(π A |G A ) and p(π B |G B ) – using (2) – and computing
A B
rates might translate to differences in human play, although the fraction of times that samples satisfy πdraw < πdraw .
this hypothesis would need to be practically validated in the Figure 3a provides an indication of the relative decisiveness
future. Yet, in absence of any existing human games, we of variants, when played by AlphaZero at approximately 1
can use these results as a preliminary guess of what those second per move, and Figure 3b provides the comparison at
results might be, assuming that what is difficult to calculate
2
for AlphaZero may be difficult for human players as well. This approach follows MacKay (2003, Chapter 37.1).
7
Torpedo 0.50 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 Torpedo 0.50 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00
Semi-torpedo 0.00 0.50 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 Semi-torpedo 0.00 0.50 0.55 0.70 0.95 0.96 0.97 0.99 1.00 1.00
No-castling 0.00 0.00 0.50 1.00 1.00 1.00 1.00 1.00 1.00 1.00 No-castling 0.00 0.45 0.50 0.65 0.93 0.95 0.96 0.99 1.00 1.00
Stalemate=win 0.00 0.00 0.00 0.50 1.00 1.00 1.00 1.00 1.00 1.00 Stalemate=win 0.00 0.30 0.35 0.50 0.87 0.90 0.92 0.98 1.00 1.00
Self-capture 0.00 0.00 0.00 0.00 0.50 0.76 0.78 0.99 1.00 1.00 Classical 0.00 0.05 0.07 0.13 0.50 0.56 0.62 0.83 0.94 0.97
Pawn-sideways 0.00 0.00 0.00 0.00 0.24 0.50 0.53 0.95 1.00 1.00 Pawn-sideways 0.00 0.04 0.05 0.10 0.44 0.50 0.56 0.79 0.91 0.96
Classical 0.00 0.00 0.00 0.00 0.22 0.46 0.50 0.94 1.00 1.00 Self-capture 0.00 0.03 0.04 0.08 0.38 0.44 0.50 0.75 0.89 0.95
Pawn one square 0.00 0.00 0.00 0.00 0.01 0.05 0.06 0.50 0.99 1.00 No-castling (10) 0.00 0.00 0.01 0.02 0.17 0.21 0.25 0.50 0.71 0.83
No-castling (10) 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.01 0.50 1.00 Pawn one square 0.00 0.00 0.00 0.00 0.06 0.09 0.11 0.29 0.50 0.65
Pawn-back 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.50 Pawn-back 0.00 0.00 0.00 0.00 0.03 0.04 0.05 0.17 0.35 0.50
o
ale ca o
Se te= g
-c e s al
o
ale ca o
at ng
d l
k
Pa lf-c win
de e
wn Cl ays
Pa C win
-c -ca ys
wn lin ure
wn are
wn 0)
e 0)
-si sica
-si tur
tli ar
ac
ac
i-t d
St - ed
a n
i-t d
St - ed
No on ssic
No Self ewa
Pa g (1
on g (1
m rpe
m tli
m rpe
m stli
as qu
Pa squ
w
Pa ast pt
-b
-b
e=
No orp
No orp
wn ap
wn las
s
a
Se To
Se To
n
Pa
row column row column

(a) A draw rate comparison p(πdraw < πdraw ) at approxi- (b) A draw rate comparison p(πdraw < πdraw ) at approx-
mately 1 seconds per move, on 10,000 AlphaZero games per imately 1 minute per move, on 1,000 AlphaZero games per
variation. variation.
Torpedo 0.50 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 Torpedo 0.50 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00
Semi-torpedo 0.00 0.50 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 No-castling 0.00 0.50 0.52 0.68 0.94 0.96 0.97 1.00 1.00 1.00
No-castling 0.00 0.00 0.50 1.00 1.00 1.00 1.00 1.00 1.00 1.00 Semi-torpedo 0.00 0.48 0.50 0.66 0.94 0.96 0.97 1.00 1.00 1.00
Stalemate=win 0.00 0.00 0.00 0.50 1.00 1.00 1.00 1.00 1.00 1.00 Stalemate=win 0.00 0.32 0.34 0.50 0.87 0.91 0.93 0.99 1.00 1.00
Self-capture 0.00 0.00 0.00 0.00 0.50 0.65 0.95 1.00 1.00 1.00 Classical 0.00 0.06 0.06 0.13 0.50 0.60 0.63 0.88 0.95 0.99
Pawn-sideways 0.00 0.00 0.00 0.00 0.35 0.50 0.90 1.00 1.00 1.00 Self-capture 0.00 0.03 0.04 0.09 0.40 0.50 0.53 0.82 0.92 0.98
Classical 0.00 0.00 0.00 0.00 0.05 0.10 0.50 0.96 1.00 1.00 Pawn-sideways 0.00 0.03 0.03 0.07 0.37 0.46 0.50 0.80 0.91 0.97
Pawn one square 0.00 0.00 0.00 0.00 0.00 0.00 0.04 0.50 1.00 1.00 No-castling (10) 0.00 0.00 0.00 0.01 0.12 0.18 0.20 0.50 0.70 0.87
No-castling (10) 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.50 1.00 Pawn one square 0.00 0.00 0.00 0.00 0.05 0.08 0.09 0.30 0.50 0.72
Pawn-back 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.50 Pawn-back 0.00 0.00 0.00 0.00 0.01 0.02 0.03 0.13 0.28 0.50
o
ale ca o
Se te= g
Se -ca do
Pa lf-c win
de e
-c e s al
wn Cl ays
ale tor g
at do
w a l
Cl win
- sid re
wn lin ays
wn are
k
wn 0)
e 0)
Pa lf-c sica
-si tur
tli ar
ac
ac
i-t d
St - ed
a n
St mi- tlin
No on ssic
No n- ptu
Pa g (1
on g (1
m rpe
m tli
m pe
as qu
Pa squ
w
Pa cast ew
-b
-b
e=
No orp
No orp
wn ap
Se as
s
s
a
Se To
T
Pa
(c) A comparison of expected scores p(erow > ecolumn ) at 1 (d) A comparison of expected scores p(erow > ecolumn ) at 1
second per move, on 10,000 games per variation. minute per move, on 1,000 games per variation.
Figure 3. A comparison of draw rates. The most decisive chess variants under both time controls are Torpedo, Semi-torpedo, No-castling
and Stalemate=win. These four variants also give White the largest first-move advantage.
1 minute per move. Under both time controls, the most deci- be more important for defending worse positions than it is
sive chess variants we explored are Torpedo, Semi-torpedo, for attacking – given that attacking tends to involve moving
No-castling and Stalemate=win. Torpedo and Semi-torpedo forward on the board.
have increased pawn mobility, allowing for faster, more dy-
namic play, leading to more decisive outcomes. There are 3.2.3. E XPECTED SCORES
also more moves to consider at each juncture. No-castling
The decisiveness of a chess variant under imperfect play
chess makes it harder to evacuate the king to safety, similarly
does not necessarily have to correspond to the first-move
affecting the draw rate. Finally, Stalemate=win removes one
advantage. In classical chess, White scores higher on aver-
important drawing resource for the weaker side, converting
age. Top-level chess players tend to press for an advantage
a number of important endgame positions from being drawn
with the White pieces and defend with the Black pieces,
to being winning for the stronger side. Under the same con-
looking for opportunities to counter-attack. The reason is
ditions of play, the slower Pawn one square chess variant
the first-move advantage; it is an initiative that, with good
and Pawn-back chess variant are the most drawish. Pawn-
play, persists throughout the opening phase of the game.
back chess incorporates additional defensive resources, and
This not a universal property that would hold in any game ,
the ability to go back to protect the weak squares seems to
8
as playing the first move might also disadvantage a player chess. We can already see in this one example that the
in some types of games. It is therefore important to estimate overall differences in decisiveness between variants are not
the effect of the rule changes on the first-move advantage equally distributed across all possible opening lines, and
in each chess variant, expressed as the expected score for that the evaluation of the difference in the expected score
White. will depend on the style of opening play.
The expected score for White is defined as: In case of Chigorin Defence in Figure 4b, Pawn-sideways
chess seems to be refuting the variation, based on our initial
e = πwin + 21 πdraw (4) findings. In a smaller sample of games played at 1 minute
for a particular set of conditions like time controls, the per move, we have seen a 100% score being achieved by
move selection criteria and the AlphaZero model playing AlphaZero in this line of Pawn-sideways chess, though these
the game. Given the game outcomes G A and G B of variants are still preliminary conclusions. To the human eye the
A and B, the probability of white having a higher first-move line does not appear to be very forcing; it is not a short
advantage in variant A is tactical refutation, but results in a fairly long-term strategic
ZZ advantage, which AlphaZero converts into a win. This line
also seems to be harder to defend in No-castling chess and
p(eA > eB ) =
A A B B
+ 12 πdraw + 12 πdraw

I πwin > πwin
Torpedo, but not in Stalemate=win chess, unlike the Dutch
Defence.
p(π A |G A ) p(π B |G B ) dπ A dπ B , (5)
The Alekhine Defence in Figure 4c seems to be less sound in
which we again evaluate with a Monte Carlo estimate. all of the variations considered, compared to Classical chess,
with a major increase in decisiveness in Pawn-sideways
White’s first-move advantage with approximately 1 second chess, No-castling chess and Torpedo chess.
and 1 minute per move in AlphaZero games is compared
in Figures 3c and 3d respectively. The relative ordering Finally, King’s Gambit in Figure 4d seems to give a substan-
of variations follows the ranking in general decisiveness, tial advantage to Black across all chess variants considered,
suggesting that the new chess variants that are more decisive although in No-castling chess and Torpedo chess, White has
in AlphaZero games are also more advantageous for White, somewhat better winning chances than in Classical chess.
possibly due to an increase in dynamic attacking options. Pawn-sideways chess, again, seems to be the worst of the
variants to consider playing this line in. Still, in our prelimi-
3.3. Differences in specific openings nary experiments with games at longer thinking times, most
games would still ultimately end in a draw. This suggests
To further illustrate how different alterations of the rule set that it is still likely a playable opening, when played at a
would require players to adjust their opening repertoires, we very high level with deep calculation.
provide a comparison of how favourable specific opening
positions are for the first player, for each of the variants pre- 3.4. Utilisation of special moves
viously introduced in Table 1. Figure 4 shows the win, draw,
and loss percentages for White under 1 second per move, for Several of the variants that are explored in this study involve
the Dutch Defence, Chigorin Defence, Alekhine Defence additional move options that are not permitted under the
and King’s Gambit, on a sample of 1000 self-play games. rules of Classical chess, like additional pawn moves and
The only variant we did not include in these comparisons self-captures. It is not clear from the outset how often these
is Pawn one square, as the lines used in the comparisons newly introduced moves would be utilised in each of the
involve the double-pawn-moves which are not legal in that variants. Will they make a difference? We use the set of
variant. 10,000 games at 1 second per move from Section 3.1 to
quantify how often the additional moves are played.
These four opening systems are not considered to be the
most principled ways of playing Classical chess. They are 3.4.1. T ORPEDO MOVES
therefore particularly interesting for establishing if a certain
rule change pushes the evaluation of each of these openings In Semi-torpedo chess, 88% of all games have at least one
from “slightly inferior” to “unsound” or “unplayable”. torpedo move, and 1.20% of all moves played in the game
are torpedo moves. In Torpedo chess, these percentages
In case of Dutch Defence in Figure 4a, we see that it is are even higher: 94% of games utilise torpedo moves and
more favourable for White in Torpedo and Stalemate=win these represent 2.40% of all moves played in the game.
chess than in Classical chess. This is in line with the over- Furthermore, 28.7% of games featured pawn promotions
all increase in decisiveness in those variations, but is not with a torpedo move, highlighting the speed at which a
more favourable in case of No-castling chess, despite No- passed pawn can be promoted to a queen.
castling chess otherwise being more decisive than Classical
9
Dutch Defence Chigorin Defence

Pawn back 123 855 22 Pawn back 231 756 13
Semi-torpedo 220 719 61 Semi-torpedo 245 716 39
Torpedo 339 595 66 Torpedo 393 569 38
Stalemate = win 282 670 48 Stalemate = win 193 780 27
No-castling 221 749 30 No-castling 359 619 22
No-castling (10) 229 752 19 No-castling (10) 254 728 18
Pawn sideways 139 835 26 Pawn sideways 720 273 7
(a) Dutch Defence (1. d4 f5) (b) Chigorin Defence (1. d4 d5 2. c4 Nc6)
Alekhine’s Defence King’s Gambit
No-castling 444 534 22 Pawn back 26 776 198
No-castling (10) 388 602 10 Semi-torpedo 66 679 255
Pawn back 261 721 18 Torpedo 99 630 271
Semi-torpedo 296 673 31 Stalemate = win 54 614 332
Torpedo 432 536 32 No-castling 118 669 213
Stalemate=win 228 735 37 No-castling (10) 29 674 297
Pawn-sideways 434 552 14 Pawn sideways 45 574 381
(c) Alekhine Defence (1. e4 Nf6) (d) King’s Gambit (1. e4 e5 2. f4)
Figure 4. The same opening position can give vastly different degrees of advantage to either play, depending on the variant under
consideration, as shown here by the number of games won, drawn and lost for AlphaZero as White when playing at approximately 1
second per move, for a sample of 1000 games, while always playing the best move without any additional noise being added for play
diversity. The stochasticity captured in the results stems from the asynchronous execution of MCTS threads during search. Therefore,
these results indicate how favorable the ’main line’ continuation is, for each of the following openings: the Dutch Defence, the Chigorin
Defence, Alekhine Defence and the King’s Gambit.
3.4.2. BACKWARDS AND LATERAL PAWN MOVES capture a queen (1%), though these were mostly unnecessary
captures in winning positions, given that AlphaZero was not
In Pawn-back chess, 96.3% of the games involved a back-
incentivised to win in the fastest possible way.
wards pawn move. In Pawn-sideways chess, 99.6% of
games features lateral pawn moves, and a total of 11.4% of
3.4.4. W INNING THROUGH STALEMATE
all moves in the game were lateral pawn moves, as the recon-
figuring of pawn formations was common in AlphaZero’s In Stalemate=win chess the percentage of all decisive games
playing style in this chess variant. that were won by stalemate rather than mate in AlphaZero
games was 37.2%, though this number is inflated due to
3.4.3. S ELF - CAPTURES the fact that AlphaZero would often stylistically stalemate
rather than mate the opponent in positions where both are
In Self-capture chess, 52.5% of games featured self-capture
possible.
moves, which represented 0.7% of all moves played. The
most common self-captures involved sacrificing a pawn The percentages listed above suggest that the rule changes
(86.9%), although sacrificing a bishop (5.3%) or a knight featured in these chess variants did indeed leave a trace
(4.5%) was not uncommon. Rook self-capture sacrifices on how the game is being played, and that they are useful
were rare (2.3%) and occasionally AlphaZero would self- additional options that can potentially change the game dy-
10
namics. Yet, it is important to note that the resulting games Variant Entropy Equivalent 20-ply games
are still of approximately similar length, as shown in Figure
No-castling 27.65 1.02 × 1012
8 in Appendix A, with some changes in the empirical dura-
Torpedo 27.89 1.30 × 1012
tion of decisive games. This means that playing a game in
Self-capture 27.94 1.36 × 1012
one of these chess variants is unlikely to prolong or shorten
No-castling (10) 27.97 1.40 × 1012
the game by a large amount, meaning that classical time con-
Classical 28.58 2.58 × 1012
trols should still be appropriate. Note that the numbers in
Stalemate=win 29.01 3.97 × 1012
Figure 8 that correspond to the number of plies in AlphaZero
Semi-torpedo 31.63 5.45 × 1013
games are an upper bound on game length, since AlphaZero
Pawn-back 32.30 1.07 × 1014
was trained without discounting, and would therefore not
Pawn-sideways 34.16 6.85 × 1014
play the fastest winning sequence in its decisive games.
Pawn one square 38.95 8.24 × 1016
3.5. Diversity Uniform random 64.96 1.63 × 1028
For a game to be appealing, it has to be rich enough in Table 3. The average information content in nats in the first 20
options that these options do not get quickly exhausted, as plies of the AlphaZero prior for each chess variant. The uniform
play would then become repetitive. We use the average random baseline assumes an equal probability for each move in
information content (entropy) of the first T = 20 plies Classical chess, and provides rough indication of the ratio between
of play from each variant’s prior as a surrogate diversity “plausible” and “possible” games according to the AlphaZero prior.
measure. The trained AlphaZero policy priors model the The uniform random baseline depends on the number of legal
move probabilities of the positions in self-play training data, moves per position, and is marginally different but of the same
magnitude for other variations.
and reflects the statistics at which opening lines appear there.
An entropy of zero corresponds to there being one and only
one forcing sequence of moves to be playable for White theoretic terms, the entropy
and Black, all other moves leading to substantially worse
positions for each side. A higher entropy implies a wider
X
H(st ) = − p(st+1 |st ) log p(st+1 |st ) (6)
and more balanced opening tree of variations, leading to a st+1
more diverse set of middlegame positions. The intuition that
there would be many more plausible opening lines in slower is a function of state st and represents the number of nats (or
variants like Pawn one square, holds true experimentally. bits, if log2 is used) that are needed to encode the weighted
In simulation, more decisive variants like Torpedo chess moves in position st .
typically have fewer plausibly playable opening lines.
If there are M (st ) legal moves in state st , then the num-
The decomposition of the entropy as a statistical expectation ber of candidate moves m(st ) – the number that a top
can help identify whether there exist defensive lines that player would realistically consider – is much smaller than
equalise the game in an almost forcing way. In Classical M (st ). In de Groot (1946)’s original framing, M (st ) is a
chess, one such defensive resource is the Berlin Defence player’s legal freedom of choice, while m(st ) is their ob-
in the Ruy Lopez, taking the sting out of 1. e4. We show jective freedomp of choice. Iida et al. (2003) hypothesise
in Section 3.5.2 that AlphaZero, when trained on Classical that m(st ) ≈ M (st ) on average. Because p(st+1 |st ) is
chess, expresses a strong preference for the Berlin Defence, a distribution on all legal moves, we define the number of
similarly to the human consensus on the solidity of the candidate moves m(st ) by
Berlin endgame. Without the option to castle, this particular
line disappears in No-castling chess. m(st ) = exp(H(st )) ; (7)
3.5.1. AVERAGE INFORMATION CONTENT it is the number of uniformly weighted moves that could be
encoded in the same number of nats as p(st+1 |st ).3
The prior network from (1) defines the probability of a
priori considering move at in state st , but as move at leads We provide insight into the diversity of the prior opening tree
to state st+1 deterministically, we shall abbreviate the prior through two quantities, the move sequence entropy H(t) at
with p(st+1 |st ). depth t from the opening position, and the average number
of candidate moves at ply t, M(t).
The prior is a weighted list of possible moves for state st that
3
are utilised in AlphaZero’s MCTS search. The weights spec- As an illustrative example, if the number of candidate moves
ify how plausible each move is before MCTS calculation; is m(st ) = 3 for some p(st+1 |st ) that might put non-zero mass
on all of its moves, then m(st ) is also equal to the number of
they specify candidates for consideration. In information- candidate moves of a probability vector p = [ 13 , 13 , 13 , 0, . . . , 0]
that puts equal non-zero mass on only three moves.
11
Move sequence entropy Let s = s1:t = [s1 , s2 , . . . st ]

be the sequence of states after t plies, starting at s0 , the
initial position. The prior probability No-castling
Q–t without search – of 0.06
Torpedo
move sequence s1:t is p(s1:t |s0 ) = τ =1 p(sτ |sτ −1 ). The
0.05
Self-capture
entropy of the move sequence is No-castling (10)
Density (histogram)
X 0.04
Classical
H(t) = − p(s1:t ) log p(s1:t ) Stalemate=win
s1:t Semi-torpedo
0.03
Pawn-back
h i
= Es1:t ∼p(s1:t ) − log p(s1:t ) , (8) Pawn-sideways
0.02
Pawn one square
where the starting position s0 is dropped from notation for 0.01
brevity. An entropy H(t) = 0 implies that, according to
the prior, one and only one reasonable opening line could 0.00
0 10 20 30 40 50 60 70 80
be considered by White and Black up to depth t, with all − log p(s) when s ∼ p(s) at ply depth 20
deviations form that line leading to substantially worse posi-
tions for the deviating side. A higher H(t) implies that we Figure 5. Histograms of − log p(s) when s ∼ p(s) for each vari-
would a priori expect a wider opening tree of variations, and ant. Following (8), the means of these distributions give the en-
consequently a more diverse set of middlegame positions. tropies in Table 3. The individual histograms are separately pre-
sented in Figure 9 in Appendix A.
Average number of candidate moves The entropy of a
chess variant’s prior opening tree is an unwieldy number that
doesn’t immediately inform us how many move options we The two variants that have the largest entropy and hence
have in each chess variant. A more naturally interpretable largest opening tree in Table 3, Pawn-sideways and Pawn
number is the expected number of (good) candidate moves one square, also happen to be among the most drawish,
at each ply as the game unfolds. The average number of according to Figures 3a and 3b. The two variants that have
candidate moves at ply t is the smallest opening trees under our analysis, No-castling
X h i and Torpedo, are also the most decisive and give White
M(t) = p(s1:t ) m(st ) = Es1:t ∼p(s1:t ) m(st ) . (9) some of the largest advantages, according to Figures 3a to
s1:t 3d. Importantly, we estimate the size of the opening trees of
these more decisive versions to still be of the same order of
Both the sums in (8) and (9) are over an exponential number
magnitude as that of Classical chess.
of move sequences. We compute Monte Carlo estimates of
H(t) and M(t) by sampling 104 sequences from p(s) and Figure 5 (a separate figure for each variant appears in Fig-
averaging the negative log probabilities of those sequences ure 9 in Appendix A) visualises the density of − log p(s)
to obtain H(t), or averaging m(st ) over all samples at depth when state sequences s are drawn from p(s). The mean
t to obtain M(t). We defer a presentation of the breakdown of each density is the entropy of (8), and an overlap in the
of the average number of candidate moves per variant to histograms of two variants implies that their opening trees
Figure 11 in Appendix A, and will encounter M(t) next in contain a similar number of lines that are considered as
Figure 6 when Classical and No-castling chess are compared candidates with similar odds. In Figure 5, a histogram that
side by side. is shifted to the left means that fewer move sequences are
considered a priori, and each has higher probability. A his-
The entropy of the AlphaZero prior opening tree is given
togram that is shifted to the right implies that a larger variety
in Table 3 for each variation. Similar to the calculation in
of move sequences are a priori considered, and each has to
(7) we give an estimate of the equivalent number of 20-ply
be considered with a smaller probability. “Uniform random”
sequences as exp(H(t)). As a baseline comparison, we
is shown in Figure 9j, and would appear as a tall narrow
take a prior distribution for Classical chess where all legal
spike centred around 64 in this figure. In the following
moves are equally playable, and estimate the entropy of
section, we shall use log probability histograms as a tool to
the “Uniform random” move selection criteria. It affords
highlight the differences between Classical and No-castling
us a crude estimate of the number of possible classical
chess.
openings, as opposed to the number of plausibly playable or
candidate openings. The estimates in Table 3 for Classical
3.5.2. C LASSICAL VS . N O - CASTLING CHESS
chess and "Uniform random Classical chess” corroborate
the claim that the number of playable opening lines – a In Classical chess AlphaZero has a strong preference for
player’s objective freedom of choice – is roughly the square playing the Berlin Defence 1. . . e5 2. Nf3 Nc6 3. Bb5 Nf6
root of the number of legal opening lines (Iida et al., 2003). in response to 1. e4, and here 4. O-O is White’s main reply,
12
Variant Entropy Equiv. 21-ply games deeper look at it was when I was preparing for the
Classical (e4) 23.72 2.00 × 1010 match with Kasparov, and I thought that the open-
Classical (Nf3) 29.54 6.75 × 1012 ing was a good choice against Kasparov’s playing
No-castling (e4) 27.42 8.10 × 1011 style. Pursuing it required a belief in instinct and
No-castling (Nf3) 28.40 2.16 × 1012 the human assessment of the position. Nowadays,
it is considered to be a very solid opening, and
Table 4. The average information content in nats of the AlphaZero modern engines assess most arising positions as
prior for Classical and No-castling chess, estimated on the 20 plies
following 1. e4 and 1. Nf3.
which is not an option in no-castling chess. Yet, castling is

also an integral part of most other lines in the Ruy Lopez, af-
being equal.
3.6. Differences between opening trees

We compare how similar opening trees are by considering
”
how likely a given sequence of moves is under two variants.
fecting each move when considering relative preferences. In
To compare, we define one variant p as the reference variant,
the absence of castling, AlphaZero does not have as strong
and generate a move sequence s according to its prior. The
a preference for a particular line for Black after 1. e4, sug-
Kullback-Leibler divergence is a measure of how likely such
gesting either that it is not as easy to fully neutralise White’s
sequences of moves are under the opening book of variant
initiative, or alternatively that there is a larger number of
q compared to that of p. Given two distributions p(s) and
promising defensive options.
q(s), the Kullback-Leibler divergence from q to p is the
To indicate the difference between Classical and No-castling relative entropy of variant p with respect to q,
chess, we compare the prior’s opening trees after 1. e4
and 1. Nf3 in Figure 6. If we examine the density of
X p(s)
DKL [pkq] = p(s) log
− log p(s2:21 |s1 ) under p(s2:21 |s1 ), where s1 is the board s
q(s)
position after either 1. e4 or 1. Nf3, we see a marked shift in h i
the characteristics of the AlphaZero prior opening trees (see = Es∼p(s) log p(s) − log q(s) . (10)
Figures 6a and 6b). Statistically, the AlphaZero prior after
1. e4 is much more forcing than after 1. Nf3 in Classical It is the expected number of extra nats (or bits if log2 is
chess. This is also evident from the average information used) that is required to compress move sequences from
content of the 20 plies after 1. e4 and 1. Nf3 in Table 4. In variant p using variant q’s opening book distribution. The
No-castling chess, 1. e4 seems as flexible as 1. Nf3, with a calculation in (10) involves a sum that is exponential in the
much wider variety of emerging preferential lines of play in length of s, and we estimate it with a Monte Carlo average
the AlphaZero model. of log p(s)/q(s) over 104 sampled sequences from p(s).
Figure 6 additionally shows the average number of candi- A legal move in variant p may be illegal in variant q, in
date moves at each ply. In Classical chess, White has more which case there is no way in which sequences in p can be
options than Black in both lines, the difference slowly di- encoded in q. The Kullback-Leibler divergence in (10) is
minishing over time as the first-move advantage decreases. then infinite. More formally, this happens when q(st+1 |st )
1. Nf3 offers more options, as it is less forcing. In No- puts zero mass on state transitions which are possible in p.
castling chess, there seems to be a higher number of effec- We therefore need to ensure that the reference variant p is
tive available moves for both sides after 1. e4 in the first chosen so that its legal moves are a subset of those of q. In
couple of plies, based on the AlphaZero model. Table 5 we show all divergences with respect to Classical
chess, and distinguish between two kinds of variants:
The Berlin Defence is a contributing factor to the narrower
opening tree footprint we see in Figure 6a. As defensive 1. variants that add moves to Classical chess, and whose
tool for Black, Vladimir Kramnik successfully used the legal moves are supersets of Classical chess;
Berlin Defence in his World Championship Match with
Garry Kasparov in 2000. He describes his choice as follows: 2. variants that remove legal moves from Classical chess,
and whose moves are subsets of Classical chess.
“
Back in the 90s, the engines of the time seemed
to think that White had the advantage in the
The legal moves of Stalemate=win correspond to that of
Berlin endgame, giving evaluations around +1
Classical chess, and it is included as both a superset and a
in White’s favour. I thought that things weren’t
subset in Table 5. The density of samples from (10) is given
as simple, given that Black’s only real problem
in Figure 10 in Appendix A. The divergence is largest for
was the loss of castling rights, and the difficulty
variants that introduce the largest number of additional pawn
of connecting rooks. The first time that I had a
moves or the most restrictions. Self-capture chess, despite
13
1... e5 2. Nf3 Nc6 3. Bb5 Nf6 4. O-O Nxe4 5. Re1 Nd6 6. Nxe5 Nxe5
7. Bf1 Be7 8. Rxe5 O-O 9. d4 Bf6 10. Re1 Re8 11. c3
0.05
0.05
Classical (e4) No-castling (e4)
Classical (Nf3) No-castling (Nf3)
0.04
0.04
Density (histogram)
Density (histogram)
0.03 0.03
0.02 0.02
0.01 0.01
0.00 0.00
0 10 20 30 40 50 60 70 0 10 20 30 40 50 60 70
− log p(s) when s ∼ p(s) at ply depth 20 − log p(s) when s ∼ p(s) at ply depth 20
(a) The density of (negative) log likelihoods for opening lines (b) The density of (negative) log likelihoods for opening lines in
in Classical chess after 1. e4 and 1. Nf3 when move sequences No-castling chess after 1. e4 and 1. Nf3 when move sequences
are sampled from the AlphaZero prior. There is a marked are sampled from the AlphaZero prior. Without the option of
difference in overlap between the histograms, suggesting that castling a king to safety, the prior opening trees after 1. e4 and
AlphaZero a priori considers “narrower” opening lines after 1. Nf3 have more similar “distributional footprints” compared
1. e4 than after 1. Nf3. We identify the samples s at the high to Classical chess in Figure 6a.
likelihood spike with a particular line in the Berlin Defence.
6.5 7.5
No-castling (e4) (w)
Average number of candidate moves
6.0 7.0 No-castling (e4) (b)

No-castling (Nf3) (w)
5.5 6.5
No-castling (Nf3) (b)
5.0 6.0
4.5 5.5
4.0
5.0
3.5 Classical (e4) (w)
4.5
Classical (e4) (b)
3.0
Classical (Nf3) (w) 4.0
2.5 Classical (Nf3) (b)
3.5
2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21
Ply t Ply t
(c) The average number of candidate moves M(t), as computed (d) The average number of candidate moves M(t), as computed
with (9), for Classical chess. with (9), for No-castling chess.
Figure 6. The diversity of responses to 1. e4 and 1. Nf3 in Classical and No-castling chess, as well as the average number of candidate
moves available for White and Black at each ply. The spike is in the classical chess 1. e4 response distribution is at 1. . . e5 2. Nf3 Nc6
3. Bb5 Nf6 4. O-O Nxe4 5. Re1 Nd6 6. Nxe5 Nxe5 7. Bf1 Be7 8. Rxe5 O-O 9. d4 Bf6 10. Re1 Re8 11. c3, a known equalising line in the
Berlin Defence, leading to drawish positions.
the plethora of additional opportunities for self-capture, is variant. How many more candidate moves should a player
statistically closer to Classical chess because of the low Q, who was trained on one variant of chess, take into consid-
frequency at which the extra moves are played. eration when wanting to play at player P’s level in another
variation? Let q(s) be the candidate prior for the variation
3.7. How much opening theory should be relearned? that player Q was trained on, and p(s) the prior for variant
P, variant that Q wants to play. We define the combination
Although the relative entropy expresses how many more
nats are required to encode prior moves of one variant given
another, it does not tell us whether one variant’s player is
considering the right candidate moves when playing another
14
3.5
Average additional candidate moves

3.0
2.5
Pawn-sideways
2.0
Self-capture
Pawn one square
1.5
Pawn-back
1.0
No-castling
Semi-torpedo
0.5 Torpedo
No-castling (10)
0.0 Stalemate=win
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
Ply t
Figure 7. The average number of additional candidate moves Aq (t) that a Classical player Q with prior q(st+1 |st ) should consider in
order to match player P’s candidate moves from prior p(s) for each of the evaluated variants; see (15). (The order of the variants in the
legend matches their ordering at ply t = 20.)
Variant p Variant q DKL [pkq] of the two priors as the normalized supremum

Classical Stalemate=win 2.59 max p(st+1 |st ), q(st+1 |st )
Classical Self-capture 5.24 r(st+1 |st ) = P 0 0
.
Supersets
s0t+1 max p(st+1 |st ), q(st+1 |st )

Classical Semi-torpedo 10.35
(11)
Classical Pawn-back 11.70
There is a particular reason behind our choice of definition
Classical Torpedo 11.89
for the combined prior in (11): The number of candidate
Classical Pawn-sideways 24.23
moves that the combination of players P and Q would con-
Stalemate=win Classical 2.50 sider, is always smaller than the sum of candidate moves
Subsets
No-castling (10) Classical 7.17 that P and Q would consider individually.

No-castling Classical 13.19
Pawn one square Classical 20.28 Put more formally, define the number of candidate moves for
the combined player as the number of uniformly weighed
Table 5. Differences in the opening tree of the new chess variants moves that could be encoded in the same number of nats as
and Classical chess. These are expressed as Kullback-Leibler r(st+1 |st ),4
(KL) divergences, the direction depending on whether a particular  
variant is a superset or a subset of Classical chess, based on the rule X
change. In all cases but Stalemate=win the reverse KL divergences mr (st ) = exp − r(st+1 |st ) log r(st+1 |st ) .
are infinite as when there are legal opening lines s in variant p that st+1
don’t exist in q, and hence for which q(s) = 0 when p(s) is not (12)
(contributing − log 0 to the divergence). For any choice of priors p and q the number of candidate
moves that are considered by the combined player in state
st is lower bounded by
mr (st ) ≤ mp (st ) + mq (st ) , (13)
which we prove in Appendix A.1.
We now define the difference
additional(st ) = mr (st ) − mq (st ) (14)
to represent the number of additional candidate moves that
player Q should consider, to play at the level of P in position
4
The perceptive reader would recognise equation (12) as equa-
tion (7). We restate it here with a subscript to indicate the explicit
dependence on the distribution.
15
st . The additional number of candidates additional(st ) is guide, we provide an experimental approximation to piece
zero when the priors match, q = p, and intuitively Q doesn’t values based on outcomes of AlphaZero games under 1
need to consider any further candidate moves. The number second per move.
of additional moves may be negative; intuitively, Q puts
We approximate piece values from the weights of a linear
enough weight on all candidates that P deems important,
model that predicts the game outcome from the difference
and doesn’t need to consider any further candidate moves.
in numbers of each piece only. As background, the real
The number of additional candidate moves and is upper
AlphaZero evaluation v in (p, v) = fθ (s) is the output of
bounded by additional(st ) ≤ mp (st ) according to (13); at
a deep neural network with weights θ. The expected game
the very worst, Q would additionally have to consider all of
outcome v is the result of a final tanh activation to ensure
P’s candidates.
an output in (−1, 1). If z ∈ {−1, 0, 1} indicates the playing
We consider positions up to ply t plies sampled from prior side’s game outcome, AlphaZero’s loss function includes the
for P, and at ply t evaluate how many additional candidate mean squared error (z − v)2 (Silver et al., 2018). We create
moves Q should consider on average: a simplified evaluation function gw (s) that only takes piece
h i counts on the board into consideration. For a position s we
Aq (t) = Es1:t ∼p(s1:t ) additional(st ) . (15) def
construct a feature vector d = [1, dp , dN , dB , dR , dQ ] that
contains the integer differences between the playing side
The expectation is estimated with a Monte Carlo average
and their opponent’s number of pawns, knights, bishops,
over 104 samples from p(s1:t ).
rooks and queens. We define gw with weights w ∈ R6 as
Figure 7 shows the average additional number of candidate
moves if Q is taken as the Classical chess prior, with P iterat- gw (s) = tanh(wT d) . (16)
ing over all other variants. From the outset, Pawn one square
places 60% of its prior mass on 1. d3, 1. e3, 1. c3 and 1. h3, When trained on the 10,000 AlphaZero self-play board po-
which together only account for 13% of Classical’s prior sitions from Section 3.1 for each variant, the piece weights
mass. As pawns are moved from the starting rank and pieces w provide an indication of their relative importance. Let
are developed, Aq (t) slowly decreases for Pawn one square. (s, z) ∼ games represent a sample of a position and final
As the opening progresses, Stalemate=win slowly drifts game outcome from a variant’s self-play games. We min-
from zero, presumably because some board configurations imise h 2 i
that would lead to drawn endgames under Classical rules `(w) = E(s,z)∼games z − gw (s) (17)
might have a different outcome. Torpedo puts 66% of its
empirically over w, and normalise weights w by wp to yield
prior mass on one move, 1. d4, whereas the Classical prior is
the relative piece values. The recovered piece values for
broader (its top move, 1. d4, occupies 38% of its prior mass).
each of the chess variants are given in Table 6.
The truncated plot value for Torpedo is Aq (1) = −1.8, sig-
nifying that the first Classical candidate moves effectively
already include those of Torpedo chess. There is a slow Variant p N B R Q
upward drift in the average number of additional candidates Classical 1 3.05 3.33 5.63 9.5
that a Classical player has to consider under Self-capture No castling 1 2.97 3.13 5.02 9.49
chess as a game progresses. We hypothesise that it can, in No castling (10) 1 3.14 3.40 5.37 9.85
part, be ascribed to the number of reasonable self-capturing Pawn one square 1 2.95 3.14 5.36 9.62
options increasing toward the middle game. Stalemate=win 1 2.95 3.13 4.76 8.96
Self-capture 1 3.10 3.22 5.34 9.42
3.8. Material Pawn-back 1 2.65 2.85 4.67 9.39
Semi-torpedo 1 2.72 2.95 4.69 8.3
Material plays an important role in chess, and is often used Torpedo 1 2.25 2.46 3.58 7.12
to assess whether a particular sequence of piece exchanges Pawn-sideways 1 1.8 1.98 2.99 5.92
and captures is favourable. Material sacrifices in chess are
made either for concrete tactical reasons, e.g. mating attacks, Table 6. Estimated piece values from AlphaZero self-play games
or to be traded off for long-term positional strengthening of for each variant.
the position. Understanding the material value of pieces in
chess helps players master the game and is one of the very In Classical chess, piece values vary based on positional
first pieces of chess knowledge taught to beginners. Changes considerations and game stage. The piece values in Table
to the rules of chess affect piece mobility, and hence also the 6 should not be taken as a gold standard, as the sample of
relative value of pieces. Without a basic estimate of what the AlphaZero games that they were estimated on does not fully
relative piece values in each variant are, it would be harder capture the diversity of human play, and the game lengths
for human players to start playing these chess variants. As a do not correspond to that of human games, which tend to be
16
shorter. For comparison, we have included the piece value trative games, is provided in the Appendix (Section B). For
estimates that we obtain by applying the same method to this analysis, we use the 1,000 1-minute per move games of
Classical chess, showing that the estimates do not deviate Section 3.1 as well as 200 1-minute per move games from a
much from the known material values. Over the years, diverse set of early opening positions that all of the major
many material systems have been proposed in chess. The opening systems. By looking at the former, we were able
most commonly used one (Capablanca & de Firmian, 2006) to assess AlphaZero’s preferred style of play in each chess
gives 3–3–5–9 for values of knights, bishops, rooks and variant, and by looking at the latter, we could assess how the
queens. Another system (Kaufman, 1999) gives 3.25–3.25– treatment of different opening lines changes and which of
5–9.75. Yet, bishops are typically considered to be more those become more or less promising under each of the rule
valuable than the knights, and there is usually an additive changes. Figure 1 shows an illustrative example position for
adjustment while in possession of a bishop pair. The rook each of the considered chess variants.
value varies between 4.5 and 5.5 depending on the system
What follows is a short summary of the main takeaways
and the queen values span from 8.5 to 10. The relative
from the qualitative analysis for each of the variants, pro-
piece values estimated on the AlphaZero game sample for
vided by GM Vladimir Kramnik.
Classical chess, 3.05–3.33–5.63–9.5, do not deviate much
from the existing systems. This suggests that the estimates No-castling chess is a potentially exciting variant, given
for the new chess variants are likely to be approximately that king safety is often compromised for both players, al-
correct as well. lowing for simultaneous attacking and counter-attacking and
the equality, when reached, tends to be dynamic in nature
We can see similar piece values estimated for No-castling,
rather than “dry”. The multitude of approaches to evacuate
No-castling(10), Pawn-one-square chess, Self-capture and
the king, and their timing, adds complexity to the opening
Stalemate=win. This is not surprising, given that these
play. No-castling (10), where castling is not permitted for
variants do not involve a major change in piece mobility.
the first 10 moves (20 plies) is a partial restriction, rather
Estimated piece values look quite different in the remain-
than an absolute one – which does not change the game
ing variations, where pawn mobility has been increased:
to the same extent. Due to castling being such a powerful
Pawn-back, Semi-torpedo, Torpedo and Pawn-sideways.
option, the lines preferred by AlphaZero all tend to involve
In Pawn-sideways chess, minor pieces seem to be worth
castling, only delayed – resulting in a preference for slower,
approximately two pawns, which is in line with our anec-
closed positions, and a less attractive style of play. Such
dotal observations when analysing AlphaZero games, as
partial castling restrictions can be considered if the desire is
such exchanges are frequently made. Like Torpedo chess,
to sidestep opening theory and preparation, but this may not
pawns become much stronger and more valuable than be-
be of interest for the wider chess audience.
fore. Changes in Pawn-back and Semi-torpedo are not as
pronounced. Pawn one square chess variant may appeal to players who
enjoy slower, strategic play – as well as a training tool for
4. Qualitative assessment understanding pawn structures, due to the transpositional
possibilities when setting up the pawns. The reduced pawn
To evaluate the differences in play between the set of chess mobility makes it harder to launch fast attacks, making the
variations considered in this study, we couple the quantita- game overall less decisive.
tive assessment of the variations with expert analysis based
Stalemate=win chess has little effect on the opening and
on a large set of representative games. While the overall
middlegame play, mostly affecting the evaluation of certain
decisiveness and opening diversity add to the appeal of any
endgames. As such, it does not increase decisiveness of the
chess variation, the subjective questions of aesthetic value
game by much, as it seems to almost always be possible to
and the types of positions, moves and patterns that arise are
defend without relying on stalemate as a drawing resource.
not possible to fully capture quantitatively. For providing
Therefore, this chess variant is not likely to be useful for
a deep qualitative assessment of the appeal of these chess
sidestepping known theory or for making the game substan-
variations, we rely on the experience of chess grandmaster
tially more decisive at the high level. The overall effect of
Vladimir Kramnik, an ex-world chess champion and an au-
the change seems to be minor.
thority on the game. By characterising typical patterns, we
hope to provide players with insights to help them judge for Torpedo and Semi-torpedo chess both make the game
themselves if they would find some of these chess variants more dynamic and more decisive, and Torpedo chess in
interesting enough to try out in practice. What we provide particular leads to new motifs and changes in all stages
here are preliminary findings. of the game. Creating passed pawns becomes very impor-
tant, as they are hard to stop. The attacking possibilities
The detailed qualitative assessment of the chess variants
make Torpedo chess quite appealing, and it is likely to be of
presented in this article, along with typical motifs and illus-
17
interest for players that enjoy tactical play. To assess the consequences of the rule changes, we coupled
the quantitative analysis of the trained model and self-play
Pawn-back chess makes it possible to regain control of the
games with a deep qualitative analysis where we identified
weakened squares in the position and remove some square
many new patterns and ideas that are not possible under
weaknesses. It also introduces additional possibilities for
the rules of classical chess. We showed that there several
opening up diagonals and making squares available for the
chess variants among those considered in this study that
pieces. Counter-intuitively, even though moving the pieces
are even more decisive than classical chess: Torpedo chess,
backwards is usually a defensive manoeuvre, this can make
Semi-torpedo chess, No-castling chess and Stalemate=win
more aggressive options possible, given that pawns can
chess.
now be pushed further earlier on, as there is always an
option of moving them back to cover the weakened squares. We additionally quantified the arising diversity of opening
AlphaZero has a strong preference for playing the French play and the intersection of opening trees between chess
defence with Black, which is particularly interesting. variations, showing how different the opening theory is for
each of the rule changes. There is a negative correlation
Pawn-sideways chess is incredibly complex, resulting in
between the overall opening diversity and decisiveness, as
patterns that are at times quite “alien” when one is used
the decisive variants likely require more precise play, with
to classical chess. The pawn structures become very fluid
fewer plausible choices per move. For each of the chess
and it is impossible to create permanent pawn weaknesses.
variants, we estimated the material value of each of the
Given how important this concept is in classical chess, this
pieces based on the results of 10,000 AlphaZero games,
chess variant requires us to rethink how we approach any
to provide insight into favourable exchange sequences and
given position, making it very concrete and relying on deep
make it easier for human players to understand the game.
calculation. Restructuring the pawn formation takes time,
and players need to use that time for creating other types of No-castling chess, being the first variant that we analysed
advantages. Many of AlphaZero games in this variant have (chronologically), has already been tried in an experimental
been quite tactical, some involving novel tactics that are not blitz grandmaster tournament in Chennai, as well as a couple
possible under classical rules. of longer grandmaster games. Our assessment suggests
that several of the assessed chess variants might be quite
Self-capture chess is quite entertaining, as it introduces ad-
appealing to interested players, and we hope that this study
ditional options for sacrificing material – and material sacri-
will prove to be a valuable resource for the wider chess
fices have a certain aesthetic appeal. Self-capture moves can
community.
feature in all stages of the game. Not every game involves
self-captures, as giving away material is not always required,
but they do feature in a substantial percentage of the games, Acknowledgements
and in some games they occur multiple times. Self-capture
We would like to thank chess grandmasters Peter Heine
moves can be used to open files and squares for the pieces
Nielsen, and Matthew Sadler for their valuable feedback
in the attack; opening up a blockade by sacrificing a pawn
on our preliminary findings and the early version of the
in the pawn chain; or in defence, while escaping the mating
manuscript. Oliver Smith and Kareem Ayoub have been
net.
of great help in managing the project. We would also like
to thank the team of Chess.com for providing us with a
5. Conclusions platform to announce and discuss No-castling chess and
present annotated games.
We have demonstrated how AlphaZero can be used for pro-
totyping board games and assessing the consequences of
rule changes in the game design process, as demonstrated on References
chess, where we have trained AlphaZero models to evaluate Andrade, G., Ramalho, G., Santana, H., and Corruble, V.
9 different chess variants, representing atomic changes to the Automatic computer game balancing: A reinforcement
rules of classical chess. Training an AlphaZero model under learning approach. In Proceedings of the Fourth Inter-
these rule changes helped us effectively simulate decades of national Joint Conference on Autonomous Agents and
human play in a matter of hours, and answer the “what if” Multiagent Systems, pp. 1111–1112, 2005.
question: what the play would potentially look like under
developed theory in each chess variant. We believe that a Beasly, J. What can we expect from a new chess variant?
similar approach could be used for auto-balancing game me- Variant Chess, 4(29):2, 1998.
chanics in other types of games, including computer games,
in cases when a sufficiently performant reinforcement learn- Capablanca, J. and de Firmian, N. Chess Fundamentals:
ing system is available. Completely Revised and Updated for the 21st Century.
Chess Series. Random House Puzzles & Games, 2006.
18
Cincotti, A., Iida, H., and Yoshimura, J. Refinement and Leigh, R., Schonfeld, J., and Louis, S. J. Using coevolution
complexity in the evolution of chess. In Wang, P. P. (ed.), to understand and validate game balance in continuous
Information Sciences, pp. 650–654, 2007. games. In Proceedings of the 10th Annual Conference on
Genetic and Evolutionary Computation, pp. 1563–1570,
Czech, J., Willig, M., Beyer, A., Kersting, K., and 2008.
Fürnkranz, J. Learning to play the chess variant crazy-
house above world champion level with deep neural net- MacKay, D. J. C. Information theory, inference and learning
works and human data, 2019. algorithms. Cambridge University Press, 2003.
Dalgaard, M., Felix Motzoi, J. J. S., and Sherson, J. Global Murray, H. J. R. The History of Chess. Oxford University
optimization of quantum dynamics with AlphaZero deep Press, 1913.
exploration. NPJ Quantum Information, 2020. Pritchard, D. B. The Classified Encyclopedia of Chess
Variants. Games and Puzzles Publications, 1994.
de Groot, A. D. Het Denken van den Schaker. (Thought and
Choice in Chess). Amsterdam University Press, 1946. Sadler, M. and Regan, N. Game Changer: AlphaZero’s
Groundbreaking Chess Strategies and the Promise of AI.
de Mesentier Silva, F., Lee, S., Togelius, J., and Nealen, A. New In Chess, 2019.
AI-based playtesting of contemporary board games. In
Proceedings of the 12th International Conference on the Schrittwieser, J., Antonoglou, I., Hubert, T., Simonyan, K.,
Foundations of Digital Games, pp. 13:1–10, 2017. Sifre, L., Schmitt, S., Guez, A., Lockhart, E., Hassabis,
D., Graepel, T., Lillicrap, T., and Silver, D. Mastering
Gollon, J. Chess Variations: Ancient, Regional and Modern. atari, Go, chess and shogi by planning with a learned
Charles E. Tuttle Company, 1968. model, 2019.
Grau-Moya, J., Leibfried, F., and Bou-Ammar, H. Balancing Shah, S. First ever “no-castling” tournament results in
two-player stochastic games with soft Q-learning. pp. 89% decisive games! en.chessbase.com/post/
268–274. International Joint Conferences on Artificial \the-first-ever-no-castling-chess-\
Intelligence Organization, 2018. tournament-results-in-89-\
decisive-games (accessed 20 January 2020),
Halim, Z., Baig, A. R., and Zafar, K. Evolutionary search 2020.
in the space of rules for creation of new two-player board
games. International Journal on Artificial Intelligence Silver, D., Hubert, T., Schrittwieser, J., Antonoglou, I., Lai,
Tools, 23(2):1350028, 2014. M., Guez, A., Lanctot, M., Sifre, L., Kumaran, D., Grae-
pel, T., Lillicrap, T., Simonyan, K., and Hassabis, D. A
Iida, H., Takeshita, N., and Yoshimura, J. A metric for en- general reinforcement learning algorithm that masters
tertainment of boardgames: Its implication for evolution chess, shogi, and Go through self-play. Science, 362
of chess variants. In Nakatsu, R. and Hoshino, J. (eds.), (6419):1140–1144, 2018.
Entertainment Computing: Technologies and Application,
Turing, A. Digital computers applied to games. In Bow-
pp. 65–72, 2003.
den, B. V. (ed.), Faster Than Thought: A Symposium
on Digital Computing Machines, pp. 286–310. Pitman
Jaffe, A., Miller, A., Andersen, E., Liu, Y.-E., Karlin, A.,
Publishing, London, 1953.
and Popović, Z. Evaluating competitive game balance
with restricted play. In Proceedings of the Eighth AAAI Wikipedia. List of chess variants. en.wikipedia.org/
Conference on Artificial Intelligence and Interactive Dig- \wiki/List_of_chess_variants (accessed 20
ital Entertainment, pp. 26–31, 2012. November 2019), 2019.
Kaufman, L. The evaluation of material imbalances. Chess
Life, 1999.
Kramnik, V. Kramnik And AlphaZero: How To Re-

think Chess. www.chess.com/article/view/
\no-castling-chess-kramnik-alphazero
(accessed 2 December 2019), 2019.
Lc0. Leela Chess Zero. https://lczero.org/ (ac-

cessed November 20, 2019), 2018.
19
A. Quantitative Appendix
A.1. Proof of equation (13) Classical Torpedo
0.012
No-castling Semi-torpedo
Let p and q be two vectors with non-negative entries that No-castling (10) Pawn-back
sum to one. Define r as a vector with elements 0.010
Pawn one square Pawn-sideways
Stalemate=win Self-capture
max(pi , qi ) 0.008
ri = P .
Density
(18)
i0 max(pi , qi )
0 0
0.006
We show below that

0.004
P P P
− ri log ri − pi log pi − qi log qi
e i ≤e i +e i . (19) 0.002
P
Let R = i max(pi , qi ) be the normalizing constant in 0.000
0 100 200 300 400 500
(18). It is bounded by 1 ≤ R ≤ 2. We write the entropy as Game length
X
− ri log ri (a) The game length distributions of the total number of plies
for all self-play games for each variant.
i
1 X
=− max(pi , qi ) log max(pi , qi ) + log R
R i
Classical Torpedo
0.010
1 X No-castling Semi-torpedo
=− max(pi log pi , qi log qi ) + log R No-castling (10) Pawn-back
R i 0.008 Pawn one square Pawn-sideways
Self-capture
X
Stalemate=win
≤− max(pi log pi , qi log qi ) + log R
Density
0.006
i
1X 1X
≤− pi log pi − qi log qi + log R (20) 0.004
2 i
2 i
0.002
where the last inequality in (20) follows from max(a, b) ≥
a+b
2 . Exponentiating (20) and applying Jensen’s inequality
0.000
yields 0 100 200 300 400 500
P Game length
e− i ri log ri
(b) The game length distributions of the total number of plies for
1 1
P P
≤ Re 2 (− i −pi log pi )+ 2 (− i qi log qi ) the subset of decisive (not drawn) self-games for each variant.

1 − Pi pi log pi 1 − Pi qi log qi Figure 8. The game length distributions of the total number of plies
≤R e + e
2 2 of AlphaZero games in each chess variant, based on a sample of
P P
10,000 games played at 1 second per move. The experimental
≤ e− i pi log pi
+ e− i qi log qi
. (21)
setup is described in Section 3.1.
The final line fools from R/2 ≤ 1 as 1 ≤ R ≤ 2. The
bound is tight at R = 1 when p and q both put probability
mass uniformly on two non-intersecting same-sized subsets
of elements.5
A.2. Additional figures
5
An example of two vectors giving a tight bound in (19) is
p = [ 21 , 12 , 0, 0, 0] and q = [0, 0, 21 , 12 , 0].
20
0.05
Classical Classical
No-castling Torpedo
0.04
0.04
Density (histogram)
Density (histogram)
0.03
0.03
0.02 0.02
0.01 0.01
0.00 0.00
0 10 20 30 40 50 60 70 0 10 20 30 40 50 60 70
(a) No-castling and Classical chess (b) Torpedo and Classical chess
0.05
Classical Classical
Self-capture No-castling (10)
0.04
0.04
Density (histogram)
Density (histogram)
0.03
0.03
0.02
0.02
0.01 0.01
0.00 0.00
0 10 20 30 40 50 60 70 80 0 10 20 30 40 50 60 70
(c) Self-capture and Classical chess (d) No-castling (10) and Classical chess
0.05
Classical 0.05 Classical
Stalemate=win Semi-torpedo
0.04
0.04
Density (histogram)
Density (histogram)
0.03
0.03
0.02
0.02
0.01 0.01
0.00 0.00
0 10 20 30 40 50 60 70 0 10 20 30 40 50 60 70
(e) Stalemate=win and Classical chess (f) Semi-torpedo and Classical chess
Figure 9. The density of (negative) log likelihoods for the prior opening lines for Classical chess and each of the variants. The mean of
each histogram gives the entropy or average information content for each variant’s prior p(s), as given in (8). The subfigures are ordered
by entropy, following Table 3. Figure 9g continues on the next page.
21
0.05 Classical 0.05 Classical

Pawn-back Pawn-sideways
0.04 0.04
Density (histogram)
Density (histogram)
0.03 0.03
0.02 0.02
0.01 0.01
0.00 0.00
0 10 20 30 40 50 60 70 0 10 20 30 40 50 60 70
(g) Pawn-back and Classical chess (h) Pawn-sideways and Classical chess
0.06 Classical Classical

0.20
Pawn one square Uniform random
0.05
Density (histogram)
Density (histogram)
0.15
0.04
0.03 0.10
0.02
0.05
0.01
0.00 0.00
0 10 20 30 40 50 60 70 0 10 20 30 40 50 60 70
(i) Pawn one square and Classical chess (j) Uniform random classical moves and Classical chess
Figure 9. (Continued from previous page.) The density of (negative) log likelihoods for the prior opening lines for Classical chess and
each of the variants. The mean of each histogram gives the entropy or average information content for each variant’s prior p(s), as given
in (8). The subfigures are ordered by entropy, following Table 3.
0.175 Stalemate=win 0.16 Stalemate=win

No-castling (10) Self-capture
0.150 0.14
No-castling Semi-torpedo
Pawn one square Pawn-back
Density (histogram)
Density (histogram)
0.12
0.125 Torpedo
0.10 Pawn-sideways
0.100
0.08
0.075
0.06
0.050
0.04
0.025 0.02
0.000 0.00
−10 0 10 20 30 40 0 20 40 60 80
log pcl (s) − log qvar (s) when s ∼ pcl (s) at ply depth 20 log pvar (s) − log qcl (s) when s ∼ pvar (s) at ply depth 20
(a) A decomposition of the entropy of subset variants of Classi- (b) A decomposition of the entropy of Classical chess relative
cal chess relative to Classical chess. to its superset variants.
Figure 10. Histograms of the density of terms log p(s) − log q(s) whose mean under p(s) is the Kullback-Leibler divergence in (10).
22
7.0
8.0 Classical (w)

Classical (b)
7.5 6.5
7.0 6.0
6.5
5.5
6.0
5.0
5.5
4.5
5.0
No-castling (w)
4.5 4.0 No-castling (b)
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
Ply t Ply t
(a) Classical chess (b) No-castling chess
9.00
11 No-castling (10) (w) Pawn one square (w)

No-castling (10) (b) 8.75 Pawn one square (b)
10
8.50
9
8.25
8
8.00
7
7.75
6
7.50
5
7.25
4
7.00
2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
Ply t Ply t
(c) No-castling (10) chess (d) Pawn one square chess
6.25
Stalemate=win (w) Torpedo (w)
8 Stalemate=win (b) 6.00 Torpedo (b)
5.75
7 5.50
5.25
6
5.00
4.75
5
4.50
4.25
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
Ply t Ply t
(e) Stalemate=win chess (f) Torpedo chess
Figure 11. The average number of candidate moves M(t) from (9) for each of the variants, as computed from their prior distributions
p(s). Figure 11g continues on the next page.
23
8.0
Semi-torpedo (w) 11 Pawn-back (w)

Semi-torpedo (b) Pawn-back (b)
7.5
10
7.0
9
6.5
8
6.0
7
5.5
6
5.0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
Ply t Ply t
(g) Semi-torpedo chess (h) Pawn-back chess
8.5 8.0
Pawn-sideways (w) Self-capture (w)

Pawn-sideways (b) 7.5 Self-capture (b)
8.0
7.0
7.5
6.5
7.0 6.0
5.5
6.5
5.0
6.0 4.5
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
Ply t Ply t
(i) Pawn-sideways chess (j) Self-capture chess
Uniform random (w)

30
Uniform random (b)
28
26
24
22
20
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
Ply t
(k) Uniform random moves under classical chess rules
Figure 11. (Continued from previous page.) The average number of candidate moves M(t) from (9) for each of the variants, as computed
from their prior distributions p(s).
24
B. Appendix B.1.2. A SSESSMENT

Here we present a selection of instructive games for each The assessment of the no-castling chess variant, as provided
of the chess variations considered in the study, along with a by Vladimir Kramnik:
detailed assessment of the variations by Vladimir Kramnik.
“
I was expecting that abandoning the castling rule
Given that different rule changes that we examined had
would make the game somewhat more favorable
led to a different degree of departure from existing chess
for White, increasing the existing opening advan-
theory and patterns, we do not present an equal amount of
tage. Statistics of AlphaZero games confirmed
instructive positions and games for each chess variation, and
this intuition, though the observed difference was
rather focus on those that have either been assessed to be of
not substantial to the point of unbalancing the
greater immediate interest or simply employ patterns that
game. Nevertheless, when considering human
are unfamiliar and novel and require more time to introduce
practice, and considering that players would find
and understand.
themselves in unknown territory at the very early
The Appendix is organised into sections corresponding to stage of the game, I would expect White to have
each of the chess variations and rule alterations examined in a higher expected score in practice than under
this study, in the following order: No-castling chess (Page regular circumstances.
25), No-castling (10) chess (Page 31), Pawn one square One of the main advantages of no-castling chess
chess (Page 34), Stalemate=win chess (Page 37), Torpedo is that it eliminates the nowadays overwhelming
(Section 40), Semi-torpedo (Page 54), Pawn-back chess importance of the opening preparation in profes-
(Page 61), Pawn-sideways chess (Page 70) and Self-capture sional chess, for years to come, and makes players
chess (Page 85). think creatively from the very beginning of each
Each of the variants-specific sections first introduces the game. This would inevitably lead to a consider-
rule change, sets out the motivation for why it seemed of ably higher amount of decisive games in chess
interest to be tried out, gives a qualitative assessment and a tournaments until the new theory develops, and
high-level conceptual overview of the dynamics of arising more creativity would be required in order to win.
play by Vladimir Kramnik and then concludes with several These factors could also increase the following
instructive games and positions, selected to illustrate the of professional chess tournaments among chess
typical motifs that arise in AlphaZero play in these varia- enthusiasts.
tions. With late middlegame and endgame patterns stay-
ing the same as in regular chess, there is a major
B.1. No-castling difference in the opening phase of a no-castling
chess game. The main conceptual rules of piece
In No-castling chess, the adjustment to the original rules
development and king safety are still valid, but
involved a full removal of castling as an option.
most concrete opening variations of regular chess
no longer apply, as castling is usually an essential
B.1.1. M OTIVATION
part of existing chess opening variations.
The motivation for the No-castling chess variant, as provided For example, possibly opening a game with 1. f4,
by Vladimir Kramnik: which is not a great idea in classical chess, might
be one of the better options already, since it might
make it easier to evacuate the king after Nf3, g3,
Bg2, Kf2, Rf1, Kg1. Some completely new pat-
terns of playing the openings start to make sense,
“
Adjustments to castling rules were chronologi-
like pushing the side pawns in order to develop
cally the first type of changes implemented and
the rooks via the “h” file or “a” file, as well as
assessed in this study. Firstly, excluding a single
“artificial castling” by means of Ke2, Re1, Kf1 and
existing rule makes it comparatively easy for hu-
others. Many new conceptual questions arise in
man players to adjust, as there is no need to learn
this chess variation.
an additional rule. Secondly, the right to castle is
relatively new in the long history of the game of For instance, one has to think about what ought
chess. Arguably, it stands out amongst the rules to be preferable: evacuating the king out of the
of chess, by providing the only legal opportunity center of the board as soon as possible or aim-
for a player to move two of their own pieces at ing to first develop all the pieces and claim space
”
the same time. and central squares. Years of practice are likely
25
required to give a clear answer on the guiding

principles of early play and best opening strate-
8
rZ0s0Z0Z
gies. Even with the help of chess engines, it would 7
Z0Z0mpj0
bl0Z0o0Z
likely take decades to develop the opening theory 6
to the same level and to the same depth as we
have in regular chess today. The engines can be
helpful with providing initial recommendations
5
Z0ZPo0Zp
of plausible opening lines of play, but the right 4
0ZBZ0Z0Z
Z0Z0Z0Z0
understanding and timing of the implementation 3
of new patterns is crucial in practical play.
Studying the numerous no-castling games played 2
PONL0OPO
Z0ZRJ0ZR
by AlphaZero, I have noticed one major concep- 1
tual change. Since both kings have a harder time
finding a safe place, the dynamic positional fac- a b c d e f g h
tors (e.g. initiative, piece activity, attack), seem to
have more importance than in regular chess. In
other words, a game becomes sharper, with both
sides attacking the opponent king at the same
time. Main line after d4 The main line of AlphaZero after 1. d4
in No-castling chess is:
I am convinced that because of the aforemen-
tioned reasons we would see many interesting 1. d4 (book) d5 2. c4 e6 3. Nc3 c5 4. cxd5 exd5 5. Nf3
games, and many more decisive games at the top Nf6 6. g3 Nc6 7. Bg2 h6 8. Kf1 Be6 9. Bf4 Rc8 10. h4 a6
level chess tournaments in case the organisers 11. Rc1 Rg8 12. a3 c4 13. Ne5 Bd6 14. e4 dxe4 15. Nxe4
decide to give it a try. Due to the simplicity of the Bxe5 16. dxe5 Nxe4 17. Bxe4 Qa5 18. Kg2 Rd8 19. Qe2
adjustment compared to regular chess, it is also Nd4 20. Qe3 Nf5
easy to implement this variation at any other level,
including the online chess playing platforms, as
it merely requires an agreement between the two
0Z0skZrZ
players not to play castling in their game.
B.1.3. M AIN L INES

Here we discuss “main lines” of AlphaZero under No-
” 8
7
6
ZpZ0Zpo0
pZ0ZbZ0o
castling chess, when playing with roughly one minute per
move from a particular fixed first move. Note that these
5
l0Z0OnZ0
are not purely deterministic, and each of the given lines is 4
0ZpZBA0O
O0Z0L0O0
merely one of several highly promising and likely options. 3
Here we give the first 20 moves in each of the main lines,
regardless of the position. 2
0O0Z0OKZ
1
Z0S0Z0ZR
a b c d e f g h
Main line after e4 The main line of AlphaZero after 1. e4 Main line after c4 The main line of AlphaZero after 1. c4
in No-castling chess is: in No-castling chess is:
1. e4 (book) c5 2. Nf3 Nc6 3. Nc3 Nf6 4. d4 cxd4 5. Nxd4 1. c4 (book) e5 2. Nc3 Nf6 3. Nf3 Nc6 4. d4 exd4 5. Nxd4
e6 6. Ndb5 d6 7. Bf4 e5 8. Bg5 a6 9. Na3 b5 10. Nd5 Bb4 6. Bf4 Bxc3 7. bxc3 d6 8. g3 Ne5 9. Bg2 Kf8 10. c5
Be7 11. Bxf6 Bxf6 12. c4 Ne7 13. Nxf6+ gxf6 14. cxb5 h5 Ng6 11. Be3 dxc5 12. Nb3 Qe8 13. h4 h5 14. Bxc5+ Kg8
15. Qd2 Kf8 16. Bc4 Kg7 17. Rd1 d5 18. exd5 Qb6 19. bxa6 15. Qc2 a5 16. Bd4 Ne7 17. Bxf6 gxf6 18. a4 Kg7 19. Nd4
Rd8 20. Nc2 Bxa6 Rb8 20. Kf1 Bd7
26
8
0s0ZqZ0s 8
0ZrZkZ0s
7
Zpobmpj0 7
Z0Z0Zpo0
6
0Z0Z0o0Z 6
plnMbZ0o
5
o0Z0Z0Zp 5
Z0apZ0Z0
4
PZ0M0Z0O 4
PZ0Z0Z0O
3
Z0O0Z0O0 3
Z0Z0LNO0
2
0ZQZPOBZ 2
0O0ZPOBZ
1
S0Z0ZKZR 1
S0Z0ZKZR
B.1.4. I NSTRUCTIVE GAMES 16. . . Ke7 17. Nxc8+ Rxc8 18. a5 Qa7 19. Qb3 Bxf2 20. Bh3
Rb8
Game AZ-1: AlphaZero No-castling vs AlphaZero No-
0s0Z0Z0Z
castling The first ten moves for White and Black have 8
been sampled randomly from AlphaZero’s opening “book”,
with the probability proportional to the time spent calculat-
ing each move. The remaining moves follow best play, at
7
l0Z0jpo0
roughly one minute per move. 6
pZnZbZ0o
1. d4 d5 2. c4 e6 3. Nc3 c5 4. cxd5 exd5 5. Nf3 Nf6 6. g3
Nc6 7. Bg2 h6 8. h4 Be6 9. Kf1 Rc8 10. Be3 Ng4 11. Qd2
5
O0ZpZ0Z0
b5
4
0Z0Z0Z0O
3
ZQZ0ZNOB
2
0O0ZPa0Z
1
S0Z0ZKZR
a b c d e f g h
8
0Zrlka0s 21. Qa3+ Bc5 22. Qd3 Nb4 23. Qh7 Qd7 24. Bxe6 Qxe6
7
o0Z0Zpo0 25. Rc1 Be3 26. Rc3 d4 27. Rc5 Kd6 28. Re5 Qg4 29. Qf5
Qxg3 30. Rh2
6
0ZnZbZ0o
5
ZpopZ0Z0 8
0s0Z0Z0Z
4
0Z0O0ZnO 7
Z0Z0Zpo0
3
Z0M0ANO0 6
pZ0j0Z0o
2
PO0LPOBZ 5
O0Z0SQZ0
1
S0Z0ZKZR 4
0m0o0Z0O
a b c d e f g h 3
Z0Z0aNl0
2
0O0ZPZ0S
1
Z0Z0ZKZ0
a b c d e f g h
12. Nxb5 Qb6 13. a4 a6 14. dxc5 Nxe3+ 15. Qxe3 Bxc5 30. . . Qg6 31. Rg2 Qxf5 32. Rxf5 Ke6 33. Rc5 Kd6 34. Rf5
16. Nd6+ Ke6 35. Re5+ Kf6 36. h5 Rc8 37. Rg4 Rc1+ 38. Kg2 Nc6
27
0Z0s0Z0Z
39. Re8 Rc2 40. Kh3 Rc5 41. Kh4 Bf2+ 42. Kh3 Be3 8
43. Rh4 Rxa5 44. Kg3 Ra1 45. Rhe4 Kf5 46. Nh4+ Kf6
47. Rc8 Ne7 48. Re8 Nc6 49. Nf3 Kf5 50. b3 Rb1 51. Nh4+
Kf6 52. Ra8 Ra1 53. Kh3 Ne5 54. Re8 Rh1+ 55. Kg3
7
opZbjpo0
Nc6 56. Ra8 Ra1 57. b4 Nxb4 58. Rd8 Rg1+ 59. Kh3 6
0Z0ZpZ0a
l0Z0Z0Mp
Rh1+ 60. Kg2 Rg1+ 61. Kh3 Rh1+ 62. Kg3 Rg1+ 63. Ng2 5
Nc2 64. Kh2 Rf1 65. Rc8 Kf5 66. Nxe3+ Nxe3 67. Rxd4
Kg5 68. Rc5+ f5 69. Kg3 Kxh5 70. Re4 Ng4 71. Kg2
Rf2+ 72. Kg1 Nf6 73. Re7 Rf4 74. Rxg7 Ng4 75. Rc3 Kh4
4
0m0Z0Z0O
76. Re7 Kg5 77. Ra3 h5 78. Rxa6 Rb4 79. Ra5 h4 80. Ra3 3
OPs0Z0ZR
0Z0ZBOPZ
Nf6 81. Rg7+ Kh5 82. Rf7 Kg5 83. Rg7+ Kh5 84. Kh1 2
Rb2 85. Ra5 Kh6 86. Rg2 Rb1+ 87. Rg1 Rxg1+ 88. Kxg1
Kg5 89. Ra8 Ne4 90. Kg2 Kg4 91. Ra4 Kg5 92. Rb4 Kg4
93. Rd4 Kh5 94. Kh3 Ng5+ 95. Kh2 Ne4 96. Kg2 Kg4
1
Z0SQZKZ0
a b c d e f g h
97. Rb4 Kg5 98. Kf3 Nd2+ 99. Ke3 Ne4 100. Rb7 Kg4
101. Rg7+ Ng5 102. Rg8 h3 103. Kf2 f4 1/2–1/2
20. Rhxc3 Bc6 21. Qe1 Qxa3 22. Kg1 g6 23. Bf1 Bg7
24. Re3 Rd6 25. Rc4 Nd5 26. Rf3 Nf6
8
0Z0Z0Z0Z
Game AZ-2: AlphaZero No-castling vs AlphaZero No- 7
opZ0jpa0
0ZbspmpZ
castling The first ten moves for White and Black have 6
5
Z0Z0Z0Mp
0ZRZ0Z0O
1. Nf3 e6 2. d4 d5 3. e3 c5 4. b3 h5 5. dxc5 Bxc5 6. Bb2
Kf8 7. c4 Nf6 8. h4 Bd7 9. Nc3 Nc6 10. Be2 Rc8 11. Rc1
3
lPZ0ZRZ0
Qa5 12. cxd5 Nxd5 13. Kf1 Bxe3
2
0Z0Z0OPZ
1
Z0Z0LBJ0
a b c d e f g h
8
0ZrZ0j0s 27. Rff4 Qxb3 28. Be2 Rd7 29. Rc1 Qb2 30. Bf3 Bxf3
31. Rxf3 Qd2 32. Qf1 Qd5 33. Qe1 Qd2 34. Qf1 Qd5
7
opZbZpo0 35. Qe2 Bh6 36. Qb2 Bxg5 37. hxg5 Ng4 38. Re1 Qd2
39. Qa3+ Rd6 40. Rb1 Kf8 41. g3 Ne5 42. Rf4 Qd3 43. Qxd3
6
0ZnZpZ0Z Nxd3 44. Ra4 Rd5 45. Rxb7 Rxg5 46. Ra3 Rd5 47. Rbxa7
5
l0ZnZ0Zp Ne5 48. R7a5 Rd1+ 49. Kg2 Ng4 50. Ra1 Rd4 51. R5a4
Rd3 52. R4a3 Rd4 53. Ra4 Rd3 54. R4a3 Rd2 55. R3a2
4
0Z0Z0Z0O Rd7 56. Ra7 Rd6 57. R7a6 Rd7 58. Ra7 Rd6 59. R7a6 Rd5
60. R6a5 Rd2 61. R5a2 Rd5 62. Ra5 Rd2 63. R5a2 1/2–1/2
3
ZPM0aNZ0
2
PA0ZBOPZ Game AZ-3: AlphaZero No-castling vs AlphaZero No-
castling The first ten moves for White and Black have
1
Z0SQZKZR been sampled randomly from AlphaZero’s opening “book”,
a b c d e f g h with the probability proportional to the time spent calculat-
roughly one minute per move.
1. c4 e5 2. Nc3 Nf6 3. Nf3 Nc6 4. d4 exd4 5. Nxd4 Bb4
14. Rc2 Bh6 15. Ng5 Ncb4 16. Rc1 Ke7 17. Rh3 Rhd8 6. g3 Ne4 7. Qd3 Nc5 8. Qe3+ Kf8 9. Bg2 Qf6 10. Ndb5
18. a3 Nxc3 19. Bxc3 Rxc3 Ne6 11. Kd1 Bc5 12. Qe4 d6 13. Nd5
28
rZbZ0j0s
8 B.1.5. H UMAN GAMES
Here we take a brief look at a couple of recently played blitz
7
opo0Zpop games between professional chess players from the tour-
6
0Znonl0Z nament that took place in Chennai in January 2020 (Shah,
2020). We focus on new motifs in the opening stage of
5
ZNaNZ0Z0 the game, and show how these might be counter-intuitive
compared to similar patterns in classical chess.
4
0ZPZQZ0Z
3
Z0Z0Z0O0 Game H-1: Arjun, Kalyan (2477) vs D. Gukesh (2522)
(blitz) 1. d4 d5 2. c4 c6 3. Nc3 Nf6 4. Nf3
2
PO0ZPOBO
1
S0AKZ0ZR 8
rmblka0s
a b c d e f g h 7
opZ0opop
13. . . Qd8 14. f4 Ned4 15. Qd3 Bf5 16. e4 Bg4+ 17. Ke1
Be2
6
0ZpZ0m0Z
5
Z0ZpZ0Z0
8
rZ0l0j0s 4
0ZPO0Z0Z
7
opo0Zpop 3
Z0M0ZNZ0
6
0Zno0Z0Z 2
PO0ZPOPO
5
ZNaNZ0Z0 1
S0AQJBZR
0ZPmPO0Z
4 a b c d e f g h
3
Z0ZQZ0O0 Interestingly, even at an early stage we can see an example
of a difference in patterns that originate in Classical chess
2
PO0ZbZBO and those that arise in No-castling chess. The positioning of
S0A0J0ZR
the knight on f3 is very natural, but is in fact an imprecision.
1
AlphaZero prefers keeping the option open of playing the
a b c d e f g h pawn to f3 instead, in order to tuck the king away to safety.
It gives the following line as its favored continuation: 4. e3
18. Qc3 Nxb5 19. cxb5 Bxb5 20. Be3 Bxe3 21. Nxe3 Ne7 Bf5 5. Bd3 g6 6. h3 e6 7. Nge2 Be7 8. f3 Bxd3 9. Qxd3 Kf8
22. Kf2 h5 23. h4 Rh6 24. Rac1 c6 25. Rhd1 Qb6 26. Rd4 10. Kf2 Bg7 11. Rd1.
a5 27. Rcd1 d5 28. exd5 cxd5 29. Qa3
8
rm0l0Z0s
8
rZ0Z0j0Z 7
opZ0apjp
7
ZpZ0mpo0 6
0ZpZpmpZ
6
0l0Z0Z0s 5
Z0ZpZ0Z0
5
obZpZ0Zp 4
0ZPO0Z0Z
4
0Z0S0O0O 3
Z0MQOPZP
3
L0Z0M0O0 2
PO0ZNJPZ
2
PO0Z0JBZ 1
S0ARZ0Z0
1
Z0ZRZ0Z0 a b c d e f g h
a b c d e f g h analysis diagram
The game soon ended in a draw. 1/2–1/2 Yet, 4. Nf3 was played in the game, which continued:
29
rZ0ZkZ0s
4. . . e6 5. e3 Nbd7 6. Qc2 Bd6 7. b3 b6 8
7
obZnlpop
8
rZblkZ0s
6
0opapm0Z
7
o0ZnZpop
5
Z0ZpZ0Z0
6
0opapm0Z
4
0ZPOPZ0Z
5
Z0ZpZ0Z0
3
ZPMBZNZ0
4
0ZPO0Z0Z
2
PAQZ0OPO
3
ZPM0ONZ0
1
S0Z0J0ZR
a b c d e f g h
2
PZQZ0OPO
1
S0A0JBZR This is another example of mistiming the evacuation of the
a b c d e f g h
king. Instead of playing 10. e4, it was the right time to move
the king to safety instead, retaining a large plus for White
after: 10. Kf1 Kf8 11. h4 h5 12. a4 Ng4 13. Rh3 Rh6
Here AlphaZero suggests that it was instead time to move
the king to safety. Deciding on when exactly to initiate the
evacuation of the king from the centre and choosing the best
way of achieving it is one of the key motifs of No-castling
chess. This decision is less clear than the decision to castle
8
rZ0Z0j0Z
in Classical chess, due to a larger number of options and
7
obZnlpo0
the fact that the sequence takes more moves that all need to
be staged accordingly. Instead of moving the pawn to b6,
6
0opapZ0s
AlphaZero suggests the following instead: 7. . . h5 8. Bb2
Kf8 9. Rd1 Kg8.
5
Z0ZpZ0Zp
4
PZPO0ZnO
3
ZPMBONZR
8
rZbl0Zks 2
0AQZ0OPZ
7
opZnZpo0 1
S0Z0ZKZ0
a b c d e f g h
6
0Zpapm0Z analysis diagram
5
Z0ZpZ0Zp
4
0ZPO0Z0Z Going back to the position after 10. e4, the game continua-
3
ZPM0ONZ0 tion goes as follows:
PAQZ0OPO
2 10. . . dxe4 11. Nxe4 (Giving away the advantage. Recap-
turing with the bishop was correct, even though it might
1
Z0ZRJBZR seem as otherwise counter-intuitive.) 11. . . Nxe4 12. Bxe4
f5. (This is looking bad for Black; 12. . . Nf6 is the pre-
a b c d e f g h
ferred move.) 13. Bd3 c5 (At this point, AlphaZero assesses
analysis diagram
the position as winning for White.) 14. Kf1 (The advantage
could have been kept with 14. d5.) 14. . . Bxf3 15. gxf3 cxd4
(15. . . Rf8 may have been equalizing) 16. Bxd4 (Gives the
Going back to the game continuation, after 7. . . b6 White advantage to Black. White ought to have captured on f5
has the upper hand. The game continued: 8. Bb2 Bb7 9. Bd3 instead. The right way to respond to the game move would
Qe7 10. e4 have been 16. . . Qh4.) 16. . . Be5 17. Bxe5 Nxe5 18. Bxf5
30
8
rZ0ZkZ0s 8
rm0lkZ0s
7
o0Z0l0op 7
obopZpo0
6
0o0ZpZ0Z 6
0o0ZpZ0Z
5
Z0Z0mBZ0 5
Z0a0Z0Zn
4
0ZPZ0Z0Z 4
0Z0Z0O0o
3
ZPZ0ZPZ0 3
ZPZBONZP
2
PZQZ0O0O 2
PAPO0JPZ
1
S0Z0ZKZR 1
SNZQZ0ZR
analysis diagram
A brilliant piece sacrifice.
18. . . exf5 19. Re1 Kd8 20. Qxf5 (20. Qd2+ may have been which would have kept a big plus for Black.)
stronger) 20. . . Re8 21. f4 Qb7 22. Rg1 Ng6 (The final
9. Re1 Bf6 10. Bxf6 (10. Nc3) 10. . . Qxf6 (10. . . gxf6 was
mistake, it appears that 22. Nf7 might hold) 23. Rd1+ Ke7
the better recapture) 11. Nc3 Ng3 12. Kg1 d6 (12. . . Ke7
24. Rg3 Qh1+ 25. Ke2 Qe4+ 26. Re3 Qxe3+ 27. fxe3 Rad8
was the correct plan) 13. Ng5 Nd7 14. Nce4 Nxe4 15. Bxe4
28. Rxd8 Rxd8 29. Qe4+ Kf8 30. Qb7 1–0
Bxe4 16. Nxe4 Qg6 17. Ng5 Ke7 18. e4 (18. Qe2) 18. . . e5
19. d4 exf4 20. Nf3 Kf8 21. Qd2 Qg3 22. Re2 Rh6
Game H-2: Gelfand, Boris vs Kramnik, Vladimir (blitz)
1. f4 h5 Already Kramnik demonstrates a motif that is quite
strong in no-castling chess, pushing one of the side pawns
early. 8
rZ0Z0j0Z
7
o0onZpo0
8
rmblkans 6
0o0o0Z0s
7
opopopo0 5
Z0Z0Z0Z0
6
0Z0Z0Z0Z 4
0Z0OPo0o
5
Z0Z0Z0Zp 3
ZPZ0ZNlP
4
0Z0Z0O0Z 2
PZPLRZPZ
3
Z0Z0Z0Z0 1
S0Z0Z0J0
2
POPOPZPO a b c d e f g h
1
SNAQJBMR
a b c d e f g h 23. Rf1 (Black gains the upper hand.) 23. . . Re6 24. Nh2
(A mistake, 24. e5 was required.) 24. . . Rae8 25. Rxf4
2. Nf3 e6 3. e3 Nf6 4. b3 (Interestingly, AlphaZero doesn’t Nf6 26. e5 dxe5 27. Rf3 (Another mistake, 27. Rxe5 was
like this very normal-looking move, giving Black a slight correct.) 27. . . Qg6 28. d5 (Taking on e5 was still a better
plus after 4. . . c5 5. Bb2 Be7 6. Be2 d5 7. Rf1 Kf8 8. Kf2 continuation.) 28. . . R6e7 29. c4 e4 30. Rc3 Nh5 31. Nf1
Nc6 9. Kg1 Kg8 10. a4 Bd7.) 4. . . b6 5. Bb2 Bb7 6. Bd3 Kg8 32. Qe1 Nf4 33. Rd2 e3 34. Rxe3 Rxe3 35. Nxe3 Qe4
(5. Be2 might have been better.) 6. . . h4 (Not the most 0–1
precise, according to AlphaZero, suggesting that 6. . . c5
7. Rf1 Be7 8. Kf2 h4 9. Ng5 Kf8 10. Kg1 Rh6 11. Be2 Nc6
B.2. No-castling (10)
was still slightly better for Black.) 7. h3 (This turns out to
be the wrong reaction, giving the advantage back to Black In the No-castling (10) variant of chess, castling is only
again.) 7. . . Nh5 8. Kf2 Be7 (Here, there was an opportunity allowed from move 11 onwards, both for the first and the
to play 8. . . Bc5 instead: second player.
31
0Z0srZkZ
B.2.1. M OTIVATION 8
When it comes to limit the impact of castling on the game, it
is possible to consider different types of partial limitations,
7
ZpoqZpo0
the easiest of which is disallowing it for a fixed number
of opening moves. In this variation, we have explored the
6
pZnZ0m0o
impact of disallowing castling for the first 10 moves, but any
other number could have been used instead. Each choice
5
O0Zpo0Z0
leads to a slightly different body of opening theory, as par-
4
0O0ZPZ0Z
ticular lines either become viable or stop being viable under
different circumstances.
3
Z0OPMNZP
B.2.2. A SSESSMENT
2
0ZQZ0OPZ
The assessment of the No-castling (10) chess variant, as
1
ZRZ0S0J0
a b c d e f g h
provided by Vladimir Kramnik:
“
The main purpose of the partial restriction to
castling, as a hypothetical adjustment to the rules Main line after d4 The main line of AlphaZero after 1. d4
of chess, would be to sidestep opening theory. As in No-castling (10) chess is:
such, it is aimed at professional chess as an op-
tion to potentially consider. The game itself does 1. d4 (book) d5 2. Nf3 Nf6 3. c4 dxc4 4. Nc3 e6 5. Qa4+
not change in other meaningful ways, and Alp- c6 6. Qxc4 b5 7. Qd3 Bb7 8. e4 b4 9. Na4 Qa5 10. b3 c5
haZero usually aims at playing slower lines where 11. Ne5 cxd4 12. Qb5+ Qxb5 13. Bxb5+ Nfd7 14. Bb2 f6
castling does indeed take place after the first 10 15. Nxd7 Nxd7 16. Bxd4 Bxe4 17. O-O Bd6 18. Rfe1 Bd5
moves. This makes sense, given that castling is a 19. Nc5 Bxc5 20. Bxc5 Rb8
fast an powerful move, so aiming to take advan-
tage of it if available makes for a good approach.
Yet, the slowing down of the game could as a
side-effect lead to an increased number of draws.
Another disadvantage is the need to count and
8
0s0ZkZ0s
keep track of the move number when considering
7
o0ZnZ0op
variations.
B.2.3. M AIN L INES

Here we discuss “main lines” of AlphaZero under No-
” 6
5
4
0Z0Zpo0Z
ZBAbZ0Z0
0o0Z0Z0Z
ZPZ0Z0Z0
castling (10) chess, when playing with roughly one minute 3
per move from a particular fixed first move. Note that these
are not purely deterministic, and each of the given lines is
merely one of several highly promising and likely options.
2
PZ0Z0OPO
regardless of the position.
1
S0Z0S0J0
a b c d e f g h
in No-castling (10) chess is: in No-castling (10) chess is:
1. e4 (book) e5 2. Bc4 Nc6 3. Nf3 Nf6 4. Qe2 Bc5 5. c3 Qe7 1. c4 (book) e5 2. Nc3 Nf6 3. Nf3 Nc6 4. e4 Bb4 5. d3 d6
6. b4 Bb6 7. a4 a6 8. a5 Ba7 9. d3 d6 10. Na3 Be6 11. Nc2 6. a3 Bc5 7. b4 Bb6 8. Be3 Bg4 9. Be2 Bxf3 10. Bxf3 Nd4
O-O 12. O-O h6 13. Be3 Qd7 14. Bxa7 Nxa7 15. Rfe1 11. Na4 Nxf3+ 12. Qxf3 Bxe3 13. fxe3 Nd7 14. O-O O-O
Nc6 16. h3 Rfe8 17. Bxe6 Qxe6 18. Ne3 d5 19. Qc2 Rad8 15. Nc3 c6 16. h3 Qb6 17. Rab1 Rae8 18. a4 Re6 19. a5
20. Rab1 Qd7 Qd8 20. Qg3 Rf6
32
8
0Z0l0skZ 8
0ZkZrZ0Z
7
opZnZpop 7
opo0m0ap
6
0Zpo0s0Z 6
0Z0Z0ZpZ
5
O0Z0o0Z0 5
lNZ0ZbZ0
4
0OPZPZ0Z 4
0m0Z0Z0L
3
Z0MPO0LP 3
Z0Z0ONZ0
2
0Z0Z0ZPZ 2
PO0s0OPO
1
ZRZ0ZRJ0 1
Z0S0JBZR
B.2.4. I NSTRUCTIVE GAMES

Game AZ-4: AlphaZero No-castling (10) vs AlphaZero The fireworks continue. . .
No-castling (10) The first ten moves for White and Black 17. Rxc7+ Qxc7 18. Nxc7 Rxb2 19. Nxe8 Rb1+
have been sampled randomly from AlphaZero’s opening
“book”, with the probability proportional to the time spent
calculating each move. The remaining moves follow best
play, at roughly one minute per move. 8
0ZkZNZ0Z
1. c4 e5 2. d4 exd4 3. Qxd4 Nc6 4. Qe3+ Nge7 5. Nf3 d5
6. cxd5 Qxd5 7. Nc3 Qa5 8. Qg5 Bf5 9. Bd2 f6 10. Qh5+
7
opZ0m0ap
g6 11. Qh4 Nb4 12. Rc1 O-O-O 13. Qxf6 Bh6 6
0Z0Z0ZpZ
5
Z0Z0ZbZ0
4
0m0Z0Z0L
8
0Zks0Z0s
3
Z0Z0ONZ0
7
opo0m0Zp
2
PZ0Z0OPO
6
0Z0Z0Lpa
1
ZrZ0JBZR
a b c d e f g h
5
l0Z0ZbZ0
4
0m0Z0Z0Z Leading to a draw by perpetual check. 1/2–1/2
3
Z0M0ZNZ0 The next game is less tactically rich, but rather interesting
2
PO0APOPO from the perspective of showcasing differences in opening
play and the overall approach, when castling is not possible
1
Z0S0JBZR in the first ten moves.
a b c d e f g h
Game AZ-5: AlphaZero No-castling (10) vs AlphaZero

No-castling (10) The first ten moves for White and Black
A stunning move, offering up a piece on h6. Accepting have been sampled randomly from AlphaZero’s opening
would be disastrous for White, as Black pieces mobilise “book”, with the probability proportional to the time spent
quickly via Ned5. The h8 rook can also potentially come to calculating each move. The remaining moves follow best
e8, and this justifies the material investment. play, at roughly one minute per move.
14. e3 Rhe8 15. Qh4 Bg7 16. Nb5 Rxd2 1. c4 e5 2. Nc3 Nf6 3. Nf3 Nc6 4. Qa4
33
8
rZblka0s 8
rZbZkZ0s
7
opopZpop 7
o0opZnop
6
0ZnZ0m0Z 6
0o0Z0o0Z
5
Z0Z0o0Z0 5
Z0Z0Z0Z0
4
QZPZ0Z0Z 4
0ZBZPO0Z
3
Z0M0ZNZ0 3
A0O0O0Z0
2
PO0OPOPO 2
PZ0Z0ZPO
1
S0A0JBZR 1
S0Z0J0ZR
18. Bd5 c6 19. Bb3 Rb8 20. Bxf7+ Kxf7 21. Bd6 Ra8 22. e5
c5 23. O-O-O Ba6 24. e4 h5
This is a slightly unusual move, showcasing that the style

8
rZ0Z0Z0s
of play in this variation of chess involves opting for moves
that do not necessarily achieve as much immediately and
7
o0ZpZko0
are somewhat less direct, potentially trying to wait for the
6
bo0A0o0Z
right time to castle, when possible. In this game, however,
castling does not end up being critical.
5
Z0o0O0Zp
4. . . e4 5. Ng5 Qe7 6. c5 e3
4
0Z0ZPO0Z
3
Z0O0Z0Z0
2
PZ0Z0ZPO
1
Z0JRZ0ZR
a b c d e f g h
8
rZbZka0s
opoplpop
And the game eventually ended in a draw. 1/2–1/2
7
6
0ZnZ0m0Z B.3. Pawn one square
5
Z0O0Z0M0 B.3.1. M OTIVATION
4
QZ0Z0Z0Z Restricting the pawn movement to one square only is in-
teresting to consider, as the double-move from the second
3
Z0M0o0Z0 (or seventh rank) seems like a “special case” and an excep-
tion from the rule that pawns otherwise only move by one
2
PO0OPOPO square. In addition, slowing down the game could make it
1
S0A0JBZR more strategic and less forcing.
a b c d e f g h B.3.2. A SSESSMENT
The assessment of the Pawn one square chess variant, as
provided by Vladimir Kramnik:
“
The basic rules and patterns are still mostly the
same as in classical chess, but the opening theory
7. dxe3 Qxc5 8. Nge4 Nxe4 9. Qxe4+ Qe5 10. Qxe5 Nxe5
changes and becomes completely different. Intu-
11. e4 Bb4 12. f4 Nc4 13. e3 Nd6 14. Bd3 Bxc3 15. bxc3 f6
itively it feels that it ought to be more difficult
16. Ba3 Nf7 17. Bc4 b6
34
for White to gain a lasting opening advantage one square. This justifies the move sequence.
and convert it into a win, but since new open- 14. Nd2 Nc6 15. b3 a6 16. Nf3 Ne6 17. h3 O-O 18. O-O
ing theory would first need to be developed, this Ncd4 19. Nfxd4 exd4 20. Bd2 c6
would not pertain to human play at first. In most
AlphaZero games one can notice the rather typi-
cal middlegame positions arise after the opening
phase.
8
rZ0l0skZ
This variation of chess can be a good pedagogical
7
ZbZ0Zpa0
tool when teaching and practicing slow, strategic
play and learning about how to set up and commit
6
poponZpo
to pawn structures. Since the pawns are unable
to advance very fast, many attacking ideas that
5
Z0Z0Z0Z0
involve rapid pawn advances are no longer rel-
4
0ZPoPZ0Z
evant, and the play is instead much slower and
ultimately more positional. Additionally, this vari-
3
OPZPZ0OP
ation of chess could simply be of interest for those
wishing for an easy way of side-stepping opening
2
0ZQANOBZ
1
S0Z0ZRJ0
theory.
B.3.3. M AIN L INES

Here we discuss “main lines” of AlphaZero under Pawn
one square chess, when playing with roughly one minute
” a b c d e f g h
Main line after d3 The main line of AlphaZero after 1. d3

in Pawn one square chess is:
per move from a particular fixed first move. Note that these
are not purely deterministic, and each of the given lines is 1. d3 (book) d6 2. e3 Nf6 3. Nd2 e6 4. Ngf3 Nbd7 5. d4
merely one of several highly promising and likely options. g6 6. Bd3 Bg7 7. O-O O-O 8. h3 e5 9. c3 Re8 10. e4 b6
Here we give the first 20 moves in each of the main lines, 11. Re1 Bb7 12. a3 a6 13. Qc2 h6 14. Nb3 a5 15. dxe5 Nxe5
regardless of the position. 16. Nxe5 dxe5 17. Be3 Nh5 18. Bb5 Rf8 19. Nd2 Qf6 20. g3
Rfd8
Main line after e3 The main line of AlphaZero after 1. e3
1. e3 (book) Nf6 2. d3 d6 3. Nf3 h6 4. e4 b6 5. c3 Bb7 6. Qc2
8
rZ0s0ZkZ
e6 7. c4 e5 8. g3 g6 9. Nc3 Bg7 10. Bg2 Nc6 11. Be3 Nd7
12. Ne2 Nc5 13. a3 Na5
7
Zbo0Zpa0
6
0o0Z0lpo
8
rZ0lkZ0s 5
oBZ0o0Zn
7
obo0Zpa0 4
0Z0ZPZ0Z
6
0o0o0Zpo 3
O0O0A0OP
5
m0m0o0Z0 2
0OQM0O0Z
4
0ZPZPZ0Z 1
S0Z0S0J0
3
O0ZPANO0 a b c d e f g h
2
0OQZNOBO
1
S0Z0J0ZR Main line after c3 The main line of AlphaZero after 1. c3
a b c d e f g h
1. c3 (book) d6 2. d3 Nf6 3. Nf3 h6 4. d4 Bf5 5. c4 e6 6. Nc3
An instructive position, as it looks optically like Black is c6 7. e3 d5 8. Bd3 dxc4 9. Bxc4 Bd6 10. O-O Nbd7 11. Re1
blundering material. In this variation of chess, however, Ne4 12. Bd3 Nxc3 13. bxc3 Bxd3 14. Qxd3 Qe7 15. c4 e5
b2-b4 is not a legal move, because pawns can only move 16. Qf5 O-O 17. Rb1 b6 18. c5 Bc7 19. Ba3 b5 20. d5 cxd5
35
8
rZ0Z0skZ 8
0ZbZ0Z0Z
7
o0anlpo0 7
Z0Z0Z0ak
6
0Z0Z0Z0o 6
0l0ZnopZ
5
ZpOpoQZ0 5
ZPo0o0Zp
4
0Z0Z0Z0Z 4
0Z0ZPm0Z
3
A0Z0ONZ0 3
Z0ZPZ0OP
2
PZ0Z0OPO 2
0Z0Z0O0Z
1
ZRZ0S0J0 1
AQZRZBJ0

38. . . c4 39. gxf4 Nxf4 40. Bc3 Bxh3 41. Bd2 Qe6 42. Bxf4
Here we present some examples of AlphaZero play in Pawn exf4 43. f3
one square chess.
0Z0Z0Z0Z
Game AZ-6: AlphaZero Pawn One Square vs Alp-
8
haZero Pawn One Square The first ten moves for White
and Black have been sampled randomly from AlphaZero’s
opening “book”, with the probability proportional to the
7
Z0Z0Z0ak
time spent calculating each move. The remaining moves 6
0Z0ZqopZ
ZPZ0Z0Zp
follow best play, at roughly one minute per move.
5
1. d3 Nf6 2. Nd2 d6 3. e3 e6 4. Ngf3 g6 5. h3 Bg7 6. c3
O-O 7. c4 Nbd7 8. Rb1 e5 9. b3 c6 10. Bb2 Qe7 11. Be2 b6 4
0ZpZPo0Z
12. b4 Bb7 13. a3 h6 14. O-O h5 15. Qc2 Rfd8 16. Rfd1 c5 3
Z0ZPZPZb
8
rZ0s0ZkZ
2
0Z0Z0Z0Z
7
obZnlpa0
1
ZQZRZBJ0
a b c d e f g h
6
0o0o0mpZ
5
Z0o0o0Zp 43. . . Bg4 44. Bg2 Bxf3 45. Bxf3 Qh3 46. dxc4 Qxf3
4
0OPZ0Z0Z 47. Qd3 Qg4+ 48. Kf2 Qh4+ 49. Ke2 Qh2+ 50. Kf1 Qh1+
51. Kf2 Qh4+ 52. Ke2 Qh2+ 53. Ke1 Bf8 54. Qf3 Bc5
3
O0ZPONZP 55. Kf1 Qg1+ 56. Ke2 Qh2+ 57. Kf1 Qg1+ 58. Ke2 Qh2+
59. Kf1 1/2–1/2
2
0AQMBOPZ
1
ZRZRZ0J0 Game AZ-7: AlphaZero Pawn One Square vs Alp-
a b c d e f g h
haZero Pawn One Square The first ten moves for White
Here we have a rather normal middlegame position. The and Black have been sampled randomly from AlphaZero’s
game continued: opening “book”, with the probability proportional to the
time spent calculating each move. The remaining moves
17. Ne4 Rac8 18. b5 d5 19. cxd5 Nxd5 20. Nfd2 f6 21. Nc4
Nf8 22. a4 Kh7 23. a5 Ne6 24. Ra1 Nb4 25. Qb1 Rb8
26. axb6 axb6 27. Nc3 Qe8 28. Ra7 Nc7 29. Na3 Rd7 1. d3 c6 2. e3 d6 3. c3 g6 4. d4 Nf6 5. Nf3 Bf5 6. Be2 e6
30. Ba1 Nbd5 31. Na4 Ne6 32. e4 Ndf4 33. Bf1 Bc8 7. O-O Nbd7 8. c4 Bg7 9. b3 O-O 10. Ba3 Ne4 11. Nfd2
34. Rxd7 Qxd7 35. Nc4 Qa7 36. Naxb6 Rxb6 37. Nxb6 c5 12. Nxe4 Bxe4 13. Nd2 Bc6 14. Rc1 Qa5 15. Bb2 cxd4
Qxb6 38. g3 16. exd4 d5
36
rZ0Z0skZ
8 Nf3+ 76. Kg2 Kh6 77. Kf1 g5 78. hxg5 Kxg5 79. Kf2 Qd2
80. Bc2 Qxe2+ 81. Kxe2 Kf5 82. Kf2 Ng5 83. Kg2 Nh7
7
opZnZpap 84. Kh3 Nf6 85. Bb3 Kg5 86. Be6 Kh5 87. Bb3 Kg5 88. Be6
Kh5 89. g4+ Kg5 90. Kg3 Nh7 91. Kh3 Nf6 92. Kg3 Nh7
6
0ZbZpZpZ 93. Kh3 Nf6 1/2–1/2
5
l0ZpZ0Z0 B.4. Stalemate=win
4
0ZPO0Z0Z In this variation of chess, achieving a stalemate position is
3
ZPZ0Z0Z0 considered a win for the attacking side, rather than a draw.
2
PA0MBOPO B.4.1. M OTIVATION
1
Z0SQZRJ0 The stalemate rule in classical chess allows for additional
drawing resources for the defending side, and has been
a b c d e f g h
a subject of debate, especially when considering ways of
This is a very normal-looking position, and one would be making the game potentially more decisive. Yet, due to its
hard-pressed to guess that it originated from a different potential effect on endgames, it was unclear whether such a
variation of chess, as it looks pretty “classical”. rule would also discourage some attacking ideas that involve
material sacrifices, if being down material in endgames ends
17. Re1 Rfe8 18. h3 Bh6 19. Bc3 Qc7 20. Bf1 b6 21. Bb2 up being more dangerous and less likely to lead to a draw
Qb7 22. a3 Rac8 23. Rc2 Bg7 24. Qc1 Bh6 25. Qd1 Bg7 than in classical chess.
26. Qc1 h6 27. c5 bxc5 28. dxc5 e5 29. Qb1 h5 30. Qa2 a6
31. b4 Ba4 32. Rcc1 Bh6 33. Ba1 e4 34. Rb1 Ne5 35. Nb3 B.4.2. A SSESSMENT
Kh7 36. Nd4 Nd3 37. Re3
The assessment of the Stalemate=win chess variant, as pro-
8
0ZrZrZ0Z vided by Vladimir Kramnik:
7
ZqZ0ZpZk
6
pZ0Z0Zpa
5
Z0OpZ0Zp
4
bO0MpZ0Z
3
O0ZnS0ZP
2
QZ0Z0OPZ
1
ARZ0ZBJ0
a b c d e f g h
A very instructive position, reminiscent of a famous clas-
“
sical game between Petrosian and Reshevsky from Zurich I was at first somewhat surprised that the decisive
in 1953, where Petrosian was playing Black. The posi- game percentage in this variation was roughly
tional exchange sacrifice allows White easy play on the dark equal to that of classical chess, with similar lev-
squares. els of performance for White and Black. I was
personally expecting the change to lead to more
37. . . Bxe3 38. fxe3 f6 39. Be2 Rc7 40. Rf1 Rf7 41. Qd2 decisive games and a higher winning percentage
Ne5 42. Qe1 Bb5 43. Nxb5 axb5 44. a4 Nd3 45. Qh4 bxa4 for White.
46. Bxh5 Re5 47. Be2+ Kg7 48. Qg3 Qc7 49. Bxe5 Qxe5
It seems that the openings and the middlegame re-
50. Qxe5 fxe5 51. c6 Rxf1+ 52. Bxf1 a3 53. c7 a2 54. c8=Q
main very similar to regular chess, with very few
a1=Q 55. Qb7+ Kh6 56. Qxd5 Qe1 57. Qf7 Qxb4 58. Qa2
exceptions, but that there is a significant differ-
Qc5 59. Qd2 Nb4 60. Kf2 Nd5 61. g3 Qf8+ 62. Kg1 Qc5
ence in endgame play since some basic endgame
63. Kf2 Qf8+ 64. Ke1 Nb4 65. Bc4 Kh7 66. Qd7+ Kh6
like K+P vs K are already winning instead of
67. Qd2 Kg7 68. Qf2 Qe7 69. Kf1 Nd3 70. Qe2 Qf6+
being drawn depending on the position.
71. Kg2 Qc6 72. Bb3 Qc5 73. h4 Qc1 74. Kh2 Ne1 75. Bd1
37
B.4.3. M AIN L INES

8
0Z0ZkZ0Z Here we discuss “main lines” of AlphaZero under Stale-
7
Z0Z0O0Z0 mate=win chess, when playing with roughly one minute per
move from a particular fixed first move. Note that these
6
0Z0Z0J0Z are not purely deterministic, and each of the given lines is
Z0Z0Z0Z0
5 merely one of several highly promising and likely options.
4
0Z0Z0Z0Z regardless of the position.
3
Z0Z0Z0Z0
2
0Z0Z0Z0Z
1
Z0Z0Z0Z0
a b c d e f g h
Main line after e4 The main line of AlphaZero after 1. e4
In the position above, with White to move, in clas- in Stalemate=win chess is:
sical chess the position would be a draw due to
1. e4 (book) e5 2. Nf3 Nc6 3. Bb5 Nf6 4. O-O Nxe4 5. Re1
stalemate after Ke6. Yet, the same move wins in
Nd6 6. Nxe5 Be7 7. Bf1 Nxe5 8. Rxe5 O-O 9. d4 Bf6
this variation of chess, so the defending side needs
10. Re1 Re8 11. c3 Rxe1 12. Qxe1 Ne8 13. Bf4 d5 14. Nd2
to steer away from these types of endgames.
Bf5 15. Qe2 Nd6 16. Re1 Qd7 17. Qd1 c6 18. Nb3 b6
Similarly, the stalemates that arise in K+N+N vs 19. Nd2 Ne4 20. Nf3 Bg4
K are now wins rather than draws, for example:
8
0Z0Z0ZNZ
7
Z0Z0Z0Z0
6
0Z0Z0Z0Z
8
rZ0Z0ZkZ
5
Z0Z0ZKZk
7
o0ZqZpop
4
0Z0Z0Z0Z
6
0opZ0a0Z
3
Z0Z0ZNZ0
5
Z0ZpZ0Z0
2
0Z0Z0Z0Z
4
0Z0OnAbZ
1
Z0Z0Z0Z0
3
Z0O0ZNZ0
a b c d e f g h 2
PO0Z0OPO
Looking at the games of AlphaZero, it seems that
there are enough defensive resources in most mid-
1
Z0ZQSBJ0
a b c d e f g h
dlegame positions that certain types of inferior
endgame positions, now possible under this rule
chance, could be avoided and defended. A strong
player can in principle learn to navigate to these
positions to take advantage of them, or find ways
to escape them.
In terms of the anticipated effect on human play, Main line after d4 The main line of AlphaZero after 1. d4
I would still expect this rule change to lead to a in Stalemate=win chess is:
higher percentage of wins in endgames where one
1. d4 (book) Nf6 2. c4 e6 3. g3 Bb4+ 4. Bd2 Be7 5. Qc2
side has a clear advantage, but probably not as
c6 6. Bg2 d5 7. Nf3 b6 8. O-O O-O 9. Bf4 Bb7 10. Rd1
much as one would otherwise have been expecting.
Nbd7 11. Ne5 Nh5 12. Bd2 Nhf6 13. cxd5 cxd5 14. Nc6
This may be a nice variation of chess for chess
Qe8 15. Nxe7+ Qxe7 16. Qc7 Ba6 17. Nc3 Rfc8 18. Qf4
enthusiasts with an interest in endgame patterns.
”
38
Nf8 19. Be1 h6 20. Qd2 Ng6
rZrZ0ZkZ
8 11. O-O-O a5 12. Qc2 Rxh2
7
o0Z0lpo0 8
rZ0Zka0Z
6
bo0Zpmno 7
ZplnZpo0
5
Z0ZpZ0Z0 6
0ZpZ0mpZ
4
0Z0O0Z0Z 5
o0ZpZ0Z0
3
Z0M0Z0O0 4
0Z0O0Z0Z
2
PO0LPOBO 3
Z0M0O0Z0
1
S0ZRA0J0 2
POQA0OPs
a b c d e f g h
Main line after c4 The main line of AlphaZero after 1. c4

1
Z0JRZBZR
a b c d e f g h
in Stalemate=win chess is:
13. Rxh2 Qxh2 14. a3 Nb6 15. Kb1 Qc7 16. Bc1 Bd6 17. f3
1. c4 (book) e5 2. g3 Nf6 3. Bg2 Bc5 4. d3 d5 5. cxd5 Nxd5
O-O-O 18. Bd2 Kb8 19. Na2 Re8 20. Nc1 Nbd7 21. Bd3
6. Nf3 Nc6 7. O-O O-O 8. Nc3 Nxc3 9. bxc3 Rb8 10. Bb2
Qb6 22. Ka2 c5 23. Bf1 Rc8 24. Qa4 c4
Re8 11. d4 Bd6 12. e4 Bg4 13. h3 Bh5 14. Qc2 f6 15. d5
Na5 16. c4 b6 17. Nh4 Nb7 18. Rae1 Rf8 19. f4 Nc5 20. Re3
Qd7 8
0jrZ0Z0Z
8
0s0Z0skZ
7
ZpZnZpo0
7
o0oqZ0op
6
0l0a0mpZ
6
0o0a0o0Z
5
o0ZpZ0Z0
5
Z0mPo0Zb
4
QZpO0Z0Z
4
0ZPZPO0M
3
O0Z0OPZ0
3
Z0Z0S0OP
2
KO0A0ZPZ
2
PAQZ0ZBZ
1
Z0MRZBZ0
a b c d e f g h
1
Z0Z0ZRJ0 25. Ne2 Qc6 26. Qxa5 Bc7 27. Qc3 Bd6 28. Qc2 b5 29. Kb1
a b c d e f g h
Nb6 30. Nc1 b4
The games in Stalemate=win chess are at the first glance
8
0jrZ0Z0Z
almost indistinguishable from those of classical chess, as
the lines are merely a subset of the lines otherwise playable
7
Z0Z0Zpo0
and plausible under classical rules. 6
0mqa0mpZ
Game AZ-8: AlphaZero Stalemate=win vs AlphaZero
5
Z0ZpZ0Z0
Stalemate=win The first ten moves for White and Black
4
0opO0Z0Z
“book”, with the probability proportional to the time spent 3
O0Z0OPZ0
0OQA0ZPZ
2
play, at roughly one minute per move.
1. d4 d5 2. c4 c6 3. Nf3 Nf6 4. e3 Bf5 5. Nc3 e6 6. Nh4 1
ZKMRZBZ0
Bg6 7. Qb3 Qc7 8. Nxg6 hxg6 9. Bd2 Nbd7 10. cxd5 exd5 a b c d e f g h
39
0Z0Z0Z0Z
31. axb4 Ne8 32. e4 Qd7 33. Na2 Nc7 34. Nc3 Bxb4 35. Rc1 8
Bd6 36. Qd1 g5 37. g4 f6 38. exd5 Bb4 39. d6 Bxd6 40. Ne4
Ncd5 41. Qe1 Qc6 42. Bd3 Bf8 43. Ba5 Qb7 44. Bxb6
Nxb6 45. Bf1 Qd5 46. Qd2 Rd8 47. Qc2 Rc8 48. Qh2+ Rc7
7
Z0j0Z0Z0
49. Qd2 Rd7 50. Be2 Be7 51. Qc2 Rc7 52. Nc3 Qd7 53. Qe4 6
0ZBZ0Z0Z
ZPZ0Z0Z0
Bb4 54. Rc2 Rc8 55. d5 Qe7 56. Qe6 Kb7 57. Nb5 Qc5 5
4
0ZKZ0ZpZ
8
0ZrZ0Z0Z 3
Z0Z0ZPZ0
7
ZkZ0Z0o0 2
0Z0Z0ZnZ
6
0m0ZQo0Z 1
Z0Z0Z0Z0
5
ZNlPZ0o0 a b c d e f g h
4
0apZ0ZPZ With a draw soon after. 1/2–1/2
3
Z0Z0ZPZ0 B.5. Torpedo
2
0ORZBZ0Z In the variation of chess that we’ve named Torpedo chess,
1
ZKZ0Z0Z0 the pawns can move by either one or two squares forward
from anywhere on the board rather than just from the initial
a b c d e f g h
squares, which is the case in Classical chess. We will refer
to the pawn moves that involve advancing them by two
58. Nc3 Bxc3 59. Rxc3 Qxd5 60. Qxd5 Nxd5 61. Rxc4 Rc6 squares as “torpedo” moves.
62. Bd3 Rxc4 63. Bxc4 Nb4 We have also looked at a Semi-torpedo variant in our experi-
ments, where we only add a partial extension to the original
rule and have the pawns be able to move by two squares
8
0Z0Z0Z0Z from the 2nd/3rd and 6th/7th rank for White and Black re-
ZkZ0Z0o0
7 spectively. In this section we will focus on the universal
motifs of full Torpedo chess, and cover the sub-motifs and
6
0Z0Z0o0Z sub-patterns that correspond to Semi-torpedo chess in its
own dedicated section in Appendix B.6.
5
Z0Z0Z0o0
0mBZ0ZPZ
4 B.5.1. M OTIVATION
In a sense, having the pawns always be able to move by one
3
Z0Z0ZPZ0 or two squares makes the pawn movement more consistent,
2
0O0Z0Z0Z as it removes a “special case” of them only being able to
do the “double move” from their initial position. Increasing
1
ZKZ0Z0Z0 pawn mobility has the potential of speeding up all stages of
the game. It adds additional attacking motifs to the openings
a b c d e f g h
and changes opening theory, it makes middlegames more
complicated, and changes endgame theory in cases where
Eventually, an equal endgame arises. White is the one pawns are involved.
pushing, due to the passed pawn, but it is not enough to
make progress. B.5.2. A SSESSMENT
64. Kc1 Kb6 65. Kd2 Kc5 66. Kc3 Nc6 67. Bd3 Ne5 68. b4+ The assessment of the Torpedo chess variant, as provided
Kd6 69. Be2 Kd5 70. Bd1 Kc6 71. Kb3 Nd3 72. Ka3 Kb6 by Vladimir Kramnik:
73. Be2 Nf4 74. Bc4 Ng6 75. Kb3 Kc6 76. Bd3 Ne5 77. Be2
“
Kb6 78. Kc3 Kb7 79. Kd4 Kc6 80. b5+ Kd6 81. Bd1 Nd7 The pawns become quite powerful in Torpedo
82. Kc4 Ne5+ 83. Kb4 Kc7 84. Kc5 Nd3+ 85. Kd4 Nf4 chess. Passed pawns are in particular a very
86. Kc4 Kb6 87. Bc2 g6 88. Be4 f5 89. gxf5 gxf5 90. Bxf5 strong asset and the value of pawns changes based
Ng2 91. Bd7 Kc7 92. Bc6 g4 on the circumstances and closer to the endgame.
40
All of the attacking opportunities increase and

this strongly favours the side with the initiative,
8
0Z0Z0Z0s
which makes taking initiative a crucial part of the 7
ZpZ0lpj0
pZrobmpZ
game. Pawns are very fast, so less of a strategical 6
asset and much more tactical instead. The game
becomes more tactical and calculative compared
to standard chess.
5
Z0Z0o0Z0
There is a lot of prophylactic play, which is why
4
0ZPZPZ0o
some games don’t feature many “torpedo” moves
– “torpedo” moves are simply quite powerful and
3
ZPM0Z0O0
the play often proceeds in a way where each 2
PZ0L0OBO
Z0S0ZRJ0
player positions their pawn structure so as to dis- 1
incentivise “torpedo” moves, either by the virtue
of directly blocking their advance, or by placing a b c d e f g h
their own pawns on squares that would be able to
capture “en passant” if “torpedo” moves were to
occur.
This seems to favour the “classical” style of play Main line after d4 The main line of AlphaZero after 1. d4
in classical chess, which advocates for strong in Torpedo chess is:
central control rather than conceding space to
later attack the center once established. It seems 1. d4 (book) d5 2. c4 e6 3. Nf3 Nf6 4. Nc3 a6 5. e3 b6
like it is more difficult to play openings like the 6. Bd3 Bb7 7. O-O Bd6 8. cxd5 exd5 9. Ne5 O-O 10. a3
Grunfeld or the King’s Indian defence. Nbd7 11. f4 Ne4 12. Bd2 c5 13. Be1 cxd4 14. exd4 b5
15. h3 Rc8 16. Qe2 Ndf6 17. a4 b4 18. Nxe4 dxe4 19. Bxa6
In summary, this is an interesting chess variant,
Bxa6 20. Qxa6 Bxe5
leading to lots of decisive games and a potentially
high entertainment value, involving lots of tactical
play.
B.5.3. M AIN L INES

Here we discuss “main lines” of AlphaZero under Torpedo
” 8
7
0Zrl0skZ
Z0Z0Zpop
chess, when playing with roughly one minute per move
from a particular fixed first move. Note that these are not
6
QZ0Z0m0Z
purely deterministic, and each of the given lines is merely 5
Z0Z0a0Z0
Po0OpO0Z
one of several highly promising and likely options. Here we 4
give the first 20 moves in each of the main lines, regardless
of the position. 3
Z0Z0Z0ZP
2
0O0Z0ZPZ
1
S0Z0ARJ0
a b c d e f g h
in Torpedo chess is: in Torpedo chess is:
1. e4 (book) c5 2. Nf3 d6 3. d4 Nf6 4. Nc3 cxd4 5. Nxd4 a6 1. c4 (book) c5 2. e3 e6 3. d4 d5 4. Nc3 Nc6 5. Nf3 Nf6
6. g3 h6 7. Bg2 e5 8. Nde2 Be7 9. Be3 Be6 10. Nd5 Nbd7 6. a3 h6 7. dxc5 Bxc5 8. cxd5 exd5 9. b4 Bd6 10. Bb2 O-O
11. c4 Rc8 12. b3 Ng4 13. O-O Nxe3 14. Nxe3 h4 15. Nf5 11. Be2 a5 12. b5 Ne7 13. O-O Re8 14. Rc1 Be6 15. Bd3
Kf8 16. Qd2 Nf6 17. Nc3 g6 18. Nxe7 Qxe7 19. Rad1 Rc6 Ng6 16. Ne2 a4 17. Rc2 Qe7 18. Qa1 Nf8 19. Nfd4 N8d7
20. Rc1 Kg7 20. Ng3 Ng4
41
8
rZ0ZrZkZ 8
rZbl0s0j
7
ZpZnlpo0 7
Z0Z0a0op
6
0Z0abZ0o 6
0opZBZ0Z
5
ZPZpZ0Z0 5
Z0O0mpO0
4
pZ0M0ZnZ 4
0o0OpZ0O
3
O0ZBO0M0 3
Z0Z0O0Z0
2
0ARZ0OPO 2
PA0M0O0Z
1
L0Z0ZRJ0 1
S0ZQJ0ZR

16. Bc4 Ng4 17. d6 cxd5 18. h6 Rg8
Here we showcase several instructive games that illustrate
the type of play that frequently arises in Torpedo chess,
along with some selected extracted game positions in cases
where particular (endgame) move sequences are of interest.
8
rZbl0Zrj
7
Z0Z0a0op
6
0o0Z0Z0O
Game AZ-9: AlphaZero Torpedo vs AlphaZero Tor- 5
Z0OpZpO0
pedo The first ten moves for White and Black have been
sampled randomly from AlphaZero’s opening “book”, with
4
0oBZpZnZ
the probability proportional to the time spent calculating
each move. The remaining moves follow best play, at
3
Z0Z0O0Z0
PA0M0O0Z
1. d4 d5 2. Nf3 Nf6 3. c4 e6 4. Nc3 c6 5. e3 Nbd7 6. g3 Ne4
7. Nxe4 dxe4 8. Nd2 f5 9. c5 Be7 10. h4 O-O
1
S0ZQJ0ZR
a b c d e f g h
19. hxg7+ Rxg7 20. c7 Qd7 21. Bxd5 Qxd5 22. Nc4
8
rZbl0skZ
7
opZna0op 8
rZbZ0Z0j
6
0ZpZpZ0Z 7
Z0O0a0sp
5
Z0O0ZpZ0 6
0o0Z0Z0Z
4
0Z0OpZ0O 5
Z0ZqZpO0
3
Z0Z0O0O0 4
0oNZpZnZ
2
PO0M0O0Z 3
Z0Z0O0Z0
1
S0AQJBZR 2
PA0Z0O0Z
a b c d e f g h
1
S0ZQJ0ZR
a b c d e f g h
11. g5 b6 12. b4 a5 13. Bc4 axb4 14. Bxe6+ Kh8 15. Bb2
Ne5 22. . . Qg8 23. Ne5 Nxe5 24. Bxe5 Bxg5 25. Qh5
42
rZbZ0Zqj
8 12. Bd2 cxd4 13. exd4 Ne4 14. Be1 b5 15. h3 Rc8 16. Qe2
Ndf6
7
Z0O0Z0sp
6
0o0Z0Z0Z
5
Z0Z0ApaQ
4
0o0ZpZ0Z
8
0Zrl0skZ
3
Z0Z0O0Z0
7
ZbZ0Zpop
2
PZ0Z0O0Z
6
pZ0a0m0Z
1
S0Z0J0ZR
5
ZpZpM0Z0
a b c d e f g h
4
0Z0OnO0Z
25. . . b2 26. axb3 Rxa1+ 27. Bxa1 Be7 28. f4 exf3 29. Rg1
3
O0MBZ0ZP
Bf8 30. Qg5 2
0O0ZQZPZ
8
0ZbZ0aqj
1
S0Z0ARJ0
a b c d e f g h
7
Z0O0Z0sp
6
0o0Z0Z0Z
5
Z0Z0ZpL0 A normal-looking position arises in the middlegame (this
4
0Z0Z0Z0Z is one of AlphaZero’s main lines in this variation of chess),
but the board soon explodes in tactics.
3
ZPZ0OpZ0 17. a4 b4 18. Nxe4 dxe4 19. Bxa6 Bxa6 20. Qxa6 Bxe5
2
0Z0Z0Z0Z 21. dxe5 Qd4+ 22. Kh1 Qxb2 23. Bh4 e2
1
A0Z0J0S0
a b c d e f g h
30. . . h6 31. Qxh6+ Qh7 32. Bxg7+ Bxg7 33. Qxh7+ Kxh7
34. Kf2 Be5 35. Rd1 Bb7 36. Rc1 Bc8 37. Kxf3 Kg6 38. e4
8
0ZrZ0skZ
b4 39. Rc4 Kf6 40. Rc6+ Kg5 41. Ke3 f4+ 42. Kf2 Bd4+
43. Kg2 Be5 44. Rc5 Kf6 45. Kf3 Ke6 46. Rb5 Bd7 47. Rxb4
7
Z0Z0Zpop
Bxc7 48. Rd4 Ke7 49. Rd2 Be8 50. Rh2 Bd6 51. Rh7+
Ke6 52. Rh6+ Ke7 53. Rh2 Kf6 54. Rh8 Ke7 55. Rh2 Kf6
6
QZ0Z0m0Z
56. Rh6+ Ke7 57. Rh1 Kf6 58. Rh8 Ke7 59. Rh6 Be5 60. b4
5
Z0Z0O0Z0
Kd7 61. Kf2 Bc3 62. b6 Kc8 63. Rd6 Kb7 64. Kf3 Be5
65. Rd5 Bb8 66. Rd8 Bc6 67. Rf8 Ba4 68. Ke2 Be5 69. Rf5
4
Po0Z0O0A
Bb8 70. Rf6 Be5 71. Rf5 Bb8 72. e5 Kxb6 73. Rxf4 Bb5+
74. Kd1 Bxe5 75. Rf5 Bh2 76. Rxb5+ Kxb5 1/2–1/2
3
Z0Z0Z0ZP
2
0l0ZpZPZ
Game AZ-10: AlphaZero Torpedo vs AlphaZero Tor-
pedo The first ten moves for White and Black have been
1
S0Z0ZRZK
a b c d e f g h
1. d4 d5 2. c4 e6 3. Nf3 Nf6 4. Nc3 a6 5. e3 b6 6. Bd3 Bb7 24. Rfe1 Nd5 25. e7 Rfe8 26. Qd6 Qd2 27. a6 Ra8 28. Qc6
7. O-O Bd6 8. cxd5 exd5 9. a3 O-O 10. Ne5 c5 11. f4 Nbd7 b2
43
8
rZ0ZrZkZ 8
0Z0ZrZ0Z
7
Z0Z0Opop 7
Z0m0OkZp
6
PZQZ0Z0Z 6
0Z0Z0o0Z
5
Z0ZnZ0Z0 5
Z0Z0ZpZ0
4
0Z0Z0O0A 4
0ZqZ0Z0Z
3
Z0Z0Z0ZP 3
Z0Z0Z0SP
2
0o0lpZPZ 2
0Z0Z0ZPZ
1
S0Z0S0ZK 1
Z0ZQZ0AK
Here Black utilizes a torpedo move to give back the pawn

and protect h5 via d5.
40. . . f3 41. Qxf3 Qd5 42. Qg4 Ne6 43. Be3 Rb8 44. Qa4
A series of consecutive torpedo moves had given rise to Kxe7 45. Bc1 Kf7 46. Qc2 f5 47. Kh2 Rb5 48. Qa4 h5
this incredibly sharp position, with multiple passed pawns 49. Qa7+ Qb7 50. Qa4 Qd5 51. Ba3 f4 52. Rf3 Ra5 53. Qb4
for White and Black, and the threats are culminating, as Rc5
demonstrated by the following tactical sequence.
29. Qxe8+ Rxe8 30. a8=Q Nc7
8
0Z0Z0Z0Z
7
Z0Z0ZkZ0
6
0Z0ZnZ0Z
5
Z0sqZ0Zp
8
QZ0ZrZkZ 4
0L0Z0o0Z
7
Z0m0Opop 3
A0Z0ZRZP
6
0Z0Z0Z0Z 2
0Z0Z0ZPJ
5
Z0Z0Z0Z0 1
Z0Z0Z0Z0
4
0Z0Z0O0A a b c d e f g h
3
Z0Z0Z0ZP
0o0lpZPZ
2 54. Rf2 Qf5 55. Qb2 Rd5 56. Qb7+ Kg6 57. Qc6 Kh7
58. Bb2 Rd8 59. Qb7+ Kg8 60. Rf3 Qg6 61. Be5 Qf5
1
S0Z0S0ZK 62. Ba1 Rd3 63. Qb1 Rd5 64. Qb8+ Rd8 65. Qb2 Nd4
66. Rf2 Ne6 67. Qb3 Kh7 68. Qb7+ Kg8 69. Qa6 Kh7
a b c d e f g h
70. Qa7+ Kg6 71. Qb7 Rd1 72. Qa8 Rd8 73. Qc6 Kh7
74. Qb7+ Kg6 75. Bb2 Rd1 76. Qb8 Re1 77. Bc3 Re3
78. Qg8+ Kh6 79. Qh8+ Kg6 80. Qe8+ Kh7 81. Qd7+ Kg6
82. Qc6 Kh7 83. Qb7+ Kg8 84. Qa8+ Kh7 85. Qb7+ Kg8
86. Qc8+ Kh7 87. Qh8+ Kg6 88. Qg8+ Kh6 89. Qc8 Kh7
90. Qd7+ Kg6 91. Qe8+ Kh7 92. Bd2 Rd3 93. Qb8 Rd8
31. Qa5 bxa1=Q 32. Qxd2 Qa4 33. Rxe2 f6 34. Re3 Kf7 94. Qb7+ Kg8 95. Qc6 Qg6 96. Qc3 Rd4 97. Qa5 Kh7
35. Bf2 Qb5 36. Qd6 Qf1+ 37. Bg1 Qc4 38. f5 g6 39. Rg3 98. Kg1 Re4 99. h4 Nd4 100. Qc7+ Qg7 101. Qa5 Qf7
gxf5 40. Qd1 102. Qa6 f3 103. Qh6+ Kg8 104. Qg5+ Kh7 105. Be3 Ne2+
44
8
0Z0Z0Z0Z 8
rZ0ZkZ0s
7
Z0Z0ZqZk 7
ZblnZpZp
6
0Z0Z0Z0Z 6
po0Z0ZpZ
5
Z0Z0Z0Lp 5
Z0Z0Z0Z0
4
0Z0ZrZ0O 4
0Z0O0Z0O
3
Z0Z0ApZ0 3
O0Z0ZNZ0
2
0Z0ZnSPZ 2
0O0Z0OPZ
1
Z0Z0Z0J0 1
S0mQZRJ0
analysis diagram
And the game soon ends in a draw.
106. Kh2 Qc7+ 107. Kh1 Ng3+ 108. Kg1 Ne2+ 109. Kf1
Ng3+ 110. Kg1 Ne2+ 111. Rxe2 fxe2 112. Qxh5+ Kg8
and here, White would have played 16. d6, a torpedo move –
113. Qxe2 Rxh4 114. Qa6 Qe7 115. Qc8+ Kh7 116. Qf5+
gaining an important tempo while weakening the Black king.
Kg7 117. Qg5+ Qxg5 118. Bxg5 Re4 119. Kf2 Kg6
16. . . Qc4 17. Rxc1, followed by Re1+ once the queen has
120. Kf3 Re8 121. Bf4 Re7 122. g4 Rf7 123. Kg3 Rg7
moved. AlphaZero evaluates this position as being strongly
124. Bb8 Kg5 125. Bd6 Rb7 126. Bf4+ Kg6 127. Kf3 Ra7
in White’s favour, despite the material deficit.
128. Bb8 Rb7 129. Bf4 Ra7 130. Bb8 Rb7 131. Bf4 1/2–1/2
Going back to the game continuation,
Game AZ-11: AlphaZero Torpedo vs AlphaZero Tor- 11. . . Nxd5 12. e4 O-O 13. exd5 Bxd5 14. Bg5 Qb8 15. Re1
pedo The first ten moves for White and Black have been Re8 16. Nd2 Rxe1+ 17. Qxe1 Bf8 18. Qe3 c6
1. d4 d5 2. Nf3 Nf6 3. c4 e6 4. Nc3 a6 5. e3 b6 6. Bd3 Bb7
8
rl0Z0akZ
7. a3 g6 8. cxd5 exd5 9. h4 Nbd7 10. O-O Bd6 11. Nxd5 7
Z0ZnZpZp
6
popZ0ZpZ
8
rZ0lkZ0s 5
Z0ZbZ0A0
7
ZbonZpZp 4
0Z0O0Z0O
6
po0a0mpZ 3
O0ZBL0Z0
5
Z0ZNZ0Z0 2
0O0M0OPZ
4
0Z0O0Z0O 1
S0Z0Z0J0
3
O0ZBONZ0 a b c d e f g h
2
0O0Z0OPZ
1
S0AQZRJ0 Now we see several torpedo moves taking place. First White
a b c d e f g h
takes the opportunity to plant a pawn on h6, weakening the
Black king, then Black responds by a4 and b4, getting the
An interesting tactical motif, made possible by torpedo
queenside pawns in motion and creating counterplay on the
moves. One has to wonder, after 11. . . Nxd5 12. e4, what
other side of the board.
happens on 12. . . Nf4? The game would have followed
13. e5 Nxd3 14. exd6 Nxc1 15. dxc7 Qxc7 19. h6 a4 20. Re1 Qa7 21. Bf5 b4
45
8
rZ0Z0akZ 8
0Z0ZQakZ
7
l0ZnZpZp 7
Z0Z0ZpZp
6
0ZpZ0ZpO 6
0m0Z0Zpl
5
Z0ZbZBA0 5
Z0A0Z0Z0
4
po0O0Z0Z 4
pZ0o0ZBZ
3
O0Z0L0Z0 3
OpZ0Z0O0
2
0O0M0OPZ 2
0O0Z0OKZ
1
Z0Z0S0J0 1
Z0Z0Z0Z0
A critical moment, and a decision which shows just how

valuable the advanced pawns are in this chess variation. Nor-
mally it would make sense to save the knight, but AlphaZero
decides to keep the pawn instead, and rely on promotion
22. Bg4 Qb7 23. Bh3 Rc8 24. Qd3 Ra8 25. Qe3 Nb6 26. Rc1 threats coupled with checks on d5.
Nd7 27. g3 b3 28. Bg4 Qc7 29. Re1 Nb6 30. Ne4 Bxe4
31. Qxe4 Qd6 32. Qd3 Qd5 33. Re5 Qc4 34. Qe4 c5 35. Re8 39. . . d3 40. Bxb6 Qg5 41. Bd1 Qd5+ 42. Kh2 Qe6
Rxe8 36. Qxe8 cxd4
8
0Z0ZQakZ
7
Z0Z0ZpZp
6
0A0ZqZpZ
8
0Z0ZQakZ 5
Z0Z0Z0Z0
7
Z0Z0ZpZp 4
pZ0Z0Z0Z
6
0m0Z0ZpO 3
OpZpZ0O0
5
Z0Z0Z0A0 2
0O0Z0O0J
4
pZqo0ZBZ 1
Z0ZBZ0Z0
OpZ0Z0O0
a b c d e f g h
3
2
0O0Z0O0Z Being a piece down, Black offers an exchange of queens,
1
Z0Z0Z0J0 an unusual sight, but tactically justified – Black is also
threatening to capture on a3, and that threat is hard to meet.
a b c d e f g h
White can’t passively ignore the capture and defend the b2
pawn with the bishop, because Black could capture on b2,
offering the piece for the second time – and then follow up
by an immediate a3, knowing that bxa3 would allow for
b1=Q. In addition, Black could retreat the bishop instead of
capturing on b2, to make room for a2 bxa3 and again b1=Q.
The position is getting sharp again, with Black having So, it’s again a torpedo move that makes a difference and
gained a passed pawn, and White making threats around the justifies the tactical sequence.
Black king.
43. Qb5 Qf6 44. Kg2 h5 45. Be3 Qxb2 46. Qxa4 Qxa3
37. Be7 Qc1+ 38. Kg2 Qxh6 39. Bc5 47. Qxb3 Qxb3 48. Bxb3
46
8
0Z0Z0akZ 8
rmblrZkZ
7
Z0Z0ZpZ0 7
opo0Zpop
6
0Z0Z0ZpZ 6
0Z0opm0Z
5
Z0Z0Z0Zp 5
Z0Z0O0Z0
4
0Z0Z0Z0Z 4
0ZPO0Z0Z
3
ZBZpA0O0 3
A0O0ZNZ0
2
0Z0Z0OKZ 2
PZ0Z0OPO
1
Z0Z0Z0Z0 1
S0ZQJBZR
Already we see the first torpedo move, keeping the initiative.

White is a piece up for two pawns, and has the bishop pair.
8. . . dxe5 9. Nxe5 Nbd7 10. Bd3 c5 11. O-O Qa5 12. Bb2
Yet, Black is just in time to use a torpedo move to shut the
Nxe5 13. dxe5 Nd7 14. Re1 f5 15. Qh5 Re7
White king out and exchange a pair of pawns on the h-file
(by another torpedo move).
48. . . g4 49. Bd2 Kg7 50. Kf1 f5 51. Ke1 Be7 52. Bc4 h3
8
rZbZ0ZkZ
7
opZns0op
6
0Z0ZpZ0Z
8
0Z0Z0Z0Z 5
l0o0OpZQ
7
Z0Z0a0j0 4
0ZPZ0Z0Z
6
0Z0Z0Z0Z 3
Z0OBZ0Z0
5
Z0Z0ZpZ0 2
PA0Z0OPO
4
0ZBZ0ZpZ 1
S0Z0S0J0
3
Z0ZpZ0Op a b c d e f g h
2
0Z0A0O0Z 16. Qh4 Re8 17. Qg3 Nf8 18. Bc1 Kh8 19. Be2 Bd7 20. Rd1
1
Z0Z0J0Z0 Bc6 21. Bh5 h6 22. Bxe8 Rxe8 23. Rd6 f3
a b c d e f g h
8
0Z0Zrm0j
53. gxh4 Bxh4 54. Kf1 Bg5 55. Bc3+ Bf6 56. Bd2 Bg5 7
opZ0Z0o0
0ZbSpZ0o
57. Bc3+ Bf6 58. Bxf6 Kxf6 59. Bxd3 f4 60. Be4 g3 61. Bg2
6
gxf2 62. Ba8 Kf5 63. Kxf2 Kg4 64. Bb7 Kh5 65. Ba6 f3
66. Kxf3 1/2–1/2 5
l0o0O0Z0
4
0ZPZ0Z0Z
pedo Playing from a predefined Nimzo-Indian opening
3
Z0O0ZpL0
position (the first 3 moves for each side). The remaining
2
PZ0Z0OPO
moves follow best play, at roughly one minute per move.
1. d4 (book) Nf6 (book) 2. c4 (book) e6 (book) 3. Nc3 (book)
1
S0A0Z0J0
a b c d e f g h
Bb4 (book) 4. e3 Bxc3 5. bxc3 d6 6. Nf3 O-O 7. Ba3 Re8
8. e5 Here we see an effect of another torpedo move, after the
47
0Z0Z0Z0Z
exchange sacrifice earlier, taking over the initiative and 8
creating a dangerous pawn.
24. Bd2 fxg2 25. f4 Qc7 26. Be3 Qf7 27. Bxc5 Qf5 28. Rxc6
7
Z0Z0A0ok
bxc6 29. Bxa7 Ng6 30. Be3 Ra8 31. a4 Qd3 6
0ZpZPZ0Z
5
Z0Z0Z0Z0
8
rZ0Z0Z0j
4
0ZPZ0o0Z
7
Z0Z0Z0o0
3
Z0O0S0o0
6
0ZpZpZno
2
rZ0Z0ZpZ
5
Z0Z0O0Z0
1
Z0Z0Z0J0
a b c d e f g h
4
PZPZ0O0Z
3
Z0OqA0L0 and Black manages to force a draw, as the pawns are just
too threatening.
2
0Z0Z0ZpO 41. Rf3 Ra1+ 42. Kxg2 Ra2+ 43. Kg1 Ra1+ 44. Kg2 Ra2+
1
S0Z0Z0J0 45. Kh1 Ra1+ 46. Kg2 1/2–1/2
a b c d e f g h

pedo The game starts from a predefined Ruy Lopez open-
32. Re1 Kh7 33. a5 Rxa5 34. Bb6 Qxg3 35. hxg3 Ra2 36. f5
ing position (the first 5 plies). The remaining moves follow
exf5
best play, at roughly one minute per move.
1. e4 (book) e5 (book) 2. Nf3 (book) Nc6 (book) 3. Bb5
(book) a6 4. Bxc6 dxc6 5. O-O f6 6. d4 exd4 7. Nxd4 Bd6
8
0Z0Z0Z0Z 8. Be3 c5 9. Ne2 Ne7 10. Nbc3 b6 11. Qd3 Be6 12. Rad1
Be5 13. Nd5 O-O 14. Bf4 c6 15. Nxe7+ Qxe7 16. Bxe5 fxe5
7
Z0Z0Z0ok 17. a3
6
0ApZ0Zno
5
Z0Z0OpZ0 8
rZ0Z0skZ
4
0ZPZ0Z0Z 7
Z0Z0l0op
3
Z0O0Z0O0 6
popZbZ0Z
2
rZ0Z0ZpZ 5
Z0o0o0Z0
1
Z0Z0S0J0 4
0Z0ZPZ0Z
a b c d e f g h 3
O0ZQZ0Z0
2
0OPZNOPO
The following move shows the power of advanced pawns –
37. e6!, in order to create a threat of 38. e8=Q, so Black has
1
Z0ZRZRJ0
a b c d e f g h
to block with the knight. If instead 37. e7, Black responds
by first giving the knight for the pawn – 37. . . Nxe7, and
then after 38. Rxe7 follows it up with 38. . . h4!, similar to Here comes the first torpedo move (b6-b4), gaining space
the game continuation. on the queenside.
37. e6 Ne7 38. Bc5 h4 39. Bxe7 hxg3 40. Re3 f4 17. . . b4 18. a4 h6 19. Qe3 Rad8 20. f3 a5 21. f5
48
0Z0s0skZ
8 29. . . hxg5 30. b3 Bd5
7
Z0Z0l0o0
6
0ZpZbZ0o
5
o0o0oPZ0 8
0Z0Z0ZkZ
4
Po0ZPZ0Z 7
Z0Z0Z0o0
3
Z0Z0L0Z0 6
0Zps0Z0Z
2
0OPZNZPO 5
o0ZboPo0
1
Z0ZRZRJ0 4
Po0oPZ0Z
a b c d e f g h
Here we see an effect of another torpedo move, f3-f5, ad-

3
ZPZ0Z0Z0
vancing towards the Black king.
2
0ZPL0J0O
21. . . Bc4 22. Rde1 Qd6 23. Rd1 Qxd1 24. Rxd1 Rxd1+
25. Kf2 Rd4 26. g3 Rfd8
1
Z0Z0Z0Z0
a b c d e f g h
8
0Z0s0ZkZ
7
Z0Z0Z0o0
6
0ZpZ0Z0o The Black bishop can’t be taken, due to a torpedo threat
e3+!
5
o0o0oPZ0 31. Qxg5 Bxe4 32. f7+
4
PobsPZ0Z
3
Z0Z0L0O0
2
0OPZNJ0O
1
Z0Z0Z0Z0 8
0Z0Z0ZkZ
a b c d e f g h 7
Z0Z0ZPo0
27. Nxd4 cxd4 28. Qd2 Rd6 29. g5 6
0Zps0Z0Z
5
o0Z0o0L0
8
0Z0Z0ZkZ 4
Po0obZ0Z
7
Z0Z0Z0o0 3
ZPZ0Z0Z0
6
0Zps0Z0o 2
0ZPZ0J0O
5
o0Z0oPO0 1
Z0Z0Z0Z0
4
PoboPZ0Z a b c d e f g h
3
Z0Z0Z0Z0
2
0OPL0J0O
1
Z0Z0Z0Z0 And yet another torpedo strike, in order to capture on e5.
a b c d e f g h
32. . . Kxf7 33. Qxe5 Rf6+ 34. Ke1 Bxc2 35. Qxd4 Bxb3
White uses a torpedo move to generate play on the kingside. 36. Qd7+ Kg8 37. Qd8+ Kh7 38. Qd3+ g6 39. Qxb3 c5
49
8
0Z0Z0Z0Z 8
0Z0Z0Z0Z
7
Z0Z0Z0Zk 7
Z0Z0Z0ok
6
0Z0Z0spZ 6
QZ0ZPZ0Z
5
o0o0Z0Z0 5
Z0Z0ZpZ0
4
Po0Z0Z0Z 4
pZ0ZpZqZ
3
ZQZ0Z0Z0 3
Z0Z0a0O0
2
0Z0Z0Z0O 2
PZ0Z0O0Z
1
Z0Z0J0Z0 1
A0Z0Z0J0
White ends up with the queen against the rook and two White is one torpedo move away from queening, but has to
pawns, but this ends up being a draw, as the pawns are first try to safeguard the king.
simply too fast and need to remain blocked. Normally the 34. Be5 Qd1+ 35. Qf1 Bxf2+
queen on b3 would prevent the c5 pawn from moving, but a
c5-c3 torpedo move shows that this is no longer the case!
40. Kd1 c3 41. Qc4 Rf5 42. Kc2 Rf2+ 43. Kb1 Rb2+ 44. Kc1
8
0Z0Z0Z0Z
Rd2 45. Kb1 Kh6 46. h3 Rd1+ 47. Kc2 Rd2+ 48. Kb1 Rd1+
49. Kc2 Rh1 50. Qf4+ Kh7 51. Qc7+ Kh6 52. Qb8 Rf1
7
Z0Z0Z0ok
53. Qh8+ Kg5 54. Qd8+ Kh6 55. h4 Rf2+ 56. Kb1 Rf1+ 6
0Z0ZPZ0Z
Z0Z0ApZ0
57. Kc2 Rf2+ 58. Kb1 Rf1+ 59. Kc2 1/2–1/2 5
Game AZ-14: AlphaZero Torpedo vs AlphaZero Tor- 4

pZ0ZpZ0Z
Z0Z0Z0O0
pedo The position below, with Black to move, is taken 3
from a game that was played with roughly one minute per
move: 2
PZ0Z0a0Z
8
0Z0Z0Z0Z
1
Z0ZqZQJ0
a b c d e f g h
7
Z0Z0Z0ok Black is in time, due to the torpedo threats involving the
6
QZ0Z0Z0o e-pawn.
5
Z0Z0OpZ0 36. Kxf2 e3+ 37. Kxe3 Qxf1
4
pZ0ZpZqZ
3
Z0Z0O0O0
8
0Z0Z0Z0Z
2
PZ0Z0O0O
7
Z0Z0Z0ok
1
A0a0Z0J0
6
0Z0ZPZ0Z
a b c d e f g h
5
Z0Z0ApZ0
A dynamic position from an endgame reached in one of the
4
pZ0Z0Z0Z
AlphaZero games. White has an advanced passed pawn, 3
Z0Z0J0O0
PZ0Z0Z0Z
which is quite threatening – and Black tries to respond by
2
creating threats around the White king. To achieve that,
Black starts with a torpedo move: 1
Z0Z0ZqZ0
31. . . h4 32. e6 hxg3 33. hxg3 Bxe3 a b c d e f g h
50
0s0Z0ZkZ
Black captures White’s queen, but White creates a new one, 8
with a torpedo move.
38. e8=Q Qe1+ 39. Kd3 Qb1+ 40. Kc3 Qa1+ 41. Kb4 Qxa2
7
Z0oqZ0op
6
0Z0s0Z0Z
5
ZpZPZpZ0
8
0Z0ZQZ0Z
4
0Z0Z0Z0Z
7
Z0Z0Z0ok
3
Z0ZQO0O0
6
0Z0Z0Z0Z
2
0Z0Z0O0O
5
Z0Z0ApZ0
1
Z0SRZ0J0
a b c d e f g h
4
pJ0Z0Z0Z A position from one of the AlphaZero games, illustrating
3
Z0Z0Z0O0 the utilization of pawns in a heavy piece endgame. The
qZ0Z0Z0Z
2 b-pawn is fast, and it gets pushed down the board via a
torpedo move.
1
Z0Z0Z0Z0 26. . . h5 27. h4 b3 28. Qc4 Rb7 29. Rb1 Qb5 30. Qd4 c6
a b c d e f g h 31. dxc6
8
0Z0Z0ZkZ
An interesting endgame arises, where White is up a piece,
7
ZrZ0Z0o0
given that Black had to give away its bishop in the tactics
earlier, and Black will soon only have a single pawn in
6
0ZPs0Z0Z
return. Yet, after a long struggle, AlphaZero manages to
5
ZqZ0ZpZp
defend as Black and achieve a draw. 4
0Z0L0Z0O
ZpZ0O0O0
42. Qe7 Qb3+ 43. Ka5 Qg8 44. Kxa4 Kg6 45. Bf4 Qc4+ 3
46. Ka5 Qd5+ 47. Kb4 Qd4+ 48. Kb3 Qd3+ 49. Kb2 Qd4+
50. Kc2 Qd5 51. Qe3 Qc4+ 52. Kd2 Qb4+ 53. Kd3 Qb3+
54. Kd4 Qb4+ 55. Kd5 Qb5+ 56. Kd6 Qa6+ 57. Kd7 Qb5+
2
0Z0Z0O0Z
58. Ke7 Qb7+ 59. Kd8 Qd5+ 60. Kc7 Qc4+ 61. Kb6 Qb4+
62. Kc6 Qa4+ 63. Kb7 Qd7+ 64. Kb6 Kh5 65. Qf3+ Kg6
1
ZRZRZ0J0
a b c d e f g h
66. Kc5 Qa7+ 67. Kb4 Qb6+ 68. Kc3 Qa5+ 69. Kc2 Qa4+
70. Kd2 Qa5+ 71. Ke2 Qb5+ 72. Kf2 Qb2+ 73. Kg1 Qb1+ Unlike in Classical chess, this capture is possible, even
74. Kg2 Qb2+ 75. Kh3 Qc2 76. Bd6 Qc1 77. Bf4 Qc2 though it seemingly hangs the queen. If Black were to cap-
78. Qe3 Qc6 79. Qe1 Kf6 80. Kh4 Kg6 81. Kh3 Kf6 82. Kh2 ture it with the rook, the c-pawn would queen with check in
Qc2+ 83. Bd2 Qd3 84. Bf4 Qd5 85. Qe2 Qd4 86. Bd2 Kf7 a single move! The threat of c8=Q forces Black to recapture
87. Bg5 Qd5 88. Bc1 Qc6 89. Bf4 Kf6 90. Qd3 Qe6 91. Bd2 the pawn instead.
Kf7 92. Kh3 Qf6 93. Qd5+ Ke7 94. g5 fxg4+ 95. Kxg4
31. . . Rxc6 32. e5 fxe4 33. Qxe4 Rb8 34. Rb2 Qc4 35. Qe5
Qe6+ 96. Qxe6 Kxe6 1/2–1/2
and the game soon ended in a draw. 1/2–1/2

pedo The first ten moves for White and Black were sam-
pled randomly from AlphaZero’s opening “book”, with the
probability proportional to the time spent calculating each
move. The remaining moves follow best play, at roughly
pedo The position below, with Black to move, is taken
one minute per move.
from a game that was played with roughly one minute per
move: 1. d4 Nf6 2. c4 e6 3. Nc3 d5 4. Nf3 a6 5. e3 b6 6. g3 dxc4
51
0ZkZ0s0s
7. e5 Nd5 8. Bxc4 Be7 9. O-O Bb7 10. Re1 h6 11. a3 b5 8
12. Bb3 Nxc3 13. bxc3 a4
7
Z0l0ZpZ0
6
0ZpOpZ0o
5
ZpObZ0o0
8
rm0lkZ0s 4
pZnO0L0Z
7
Zbo0apo0 3
O0A0Z0OP
6
0Z0ZpZ0o 2
0ZBZ0O0Z
5
ZpZ0O0Z0 1
S0Z0S0J0
4
pZ0O0Z0Z a b c d e f g h
3
OBO0ZNO0
2
0Z0Z0O0O
1
S0AQS0J0
a b c d e f g h
26. Qf6 Qd8 27. Qxd8+ Rxd8 28. Bd3 Rdg8 29. f3 h5
30. Be4 Re8 31. Bd3 Rh6 32. Rf1 f5 33. Bxc4 Bxc4 34. Rf2
Bd5 35. Kh2 Rg6 36. Rg1 Reg8 37. Bd2 R6g7 38. Bb4 Kd7
39. f4 gxf4 40. Rxf4 Kc8 41. Be1 b3
In the early stage of the game, we see White using a torpedo
e3-e5 move to expand in the center and Black responding
by an a6-a4 torpedo move to gain space on the queenside.
14. Bc2 Bd5 15. Qe2 c6 16. Nd2 Qa5 17. Ne4 Nd7 18. Bd2
Qc7 19. Qg4 g6 20. h3 O-O-O 21. Qf4 Nb6 22. c5
8
0ZkZ0ZrZ
7
Z0Z0Z0s0
8
0Zks0Z0s
6
0ZpOpZ0Z
7
Z0l0apZ0
5
Z0ObZpZp
6
0mpZpZpo
4
pZ0O0S0Z
5
ZpObO0Z0
3
OpZ0Z0OP
4
pZ0ONL0Z
2
0Z0Z0Z0J
3
O0Z0Z0OP
1
Z0Z0A0S0
a b c d e f g h
2
0ZBA0O0Z
1
S0Z0S0J0
a b c d e f g h
Black uses two consecutive torpedo moves (b5-b3, a4-a2)

on the queenside to create a dangerous passed pawn on a2.
White moves forward with a c3-c5 torpedo move.
42. axb4 a2 43. Bc3 Ra7 44. R4f1 Kd8 45. Ra1 Kd7 46. Bb2
22. . . Nc4 23. Bc3 Rdf8 24. Nd6+ Bxd6 25. exd6 g5 Ra4 47. Bc3 Ra3 48. Rac1 Be4 49. h4 Rg4 50. Bd2 f3
52
8
0Z0Z0Z0Z 8
0Z0Z0Z0s
7
Z0ZkZ0Z0 7
Z0ZkZ0ZP
6
0ZpOpZ0Z 6
0ZpOpZ0Z
5
Z0O0Z0Zp 5
Z0O0Z0Z0
4
0O0ObZrO 4
0ObO0Z0Z
3
s0Z0ZpO0 3
Z0A0ZpZ0
2
pZ0A0Z0J 2
pZ0Z0JpZ
1
Z0S0Z0S0 1
Z0Z0Z0ZR
Black uses another torpedo move (f5-f3) to advance further The torpedo move g4-g2 forces the White rook away from
on the kingside and create another passed pawn. the h-file.
51. Rf1 Rg8 52. Ra1 Rga8 53. Rf2 Rb3 54. Kh3 Rb1 55. Bc3 69. Re1 Rxh7 70. b6
Bd5 56. g4 Rb3 57. Be1 hxg4+ 58. Kg3 Rb1 59. Bc3 Rb3
60. Bd2 Rb1 61. Bc3 Rb3 62. Bd2 Rb2 63. h6
8
0Z0Z0Z0Z
8
rZ0Z0Z0Z
7
Z0ZkZ0Zr
7
Z0ZkZ0Z0
6
0OpOpZ0Z
6
0ZpOpZ0O
5
Z0O0Z0Z0
5
Z0ObZ0Z0
4
0ZbO0Z0Z
4
0O0O0ZpZ
3
Z0A0ZpZ0
3
Z0Z0ZpJ0
2
pZ0Z0JpZ
2
ps0A0S0Z
1
Z0Z0S0Z0
a b c d e f g h
1
S0Z0Z0Z0
a b c d e f g h
White needs to generate immediate counterplay, and does

so via b4-b6, another torpedo move. White then uses a b6-
White advances the h-pawn with an h4-h6 torpedo move, b8=Q torpedo move to promote to a queen in the next move,
seeking counterplay. demonstrating how fast the pawns are in this variation of
chess.
63. . . Rg8 64. Raf1 Bc4 65. h7 Rf8 66. Rh1 Rh8 67. Bc3
Rxf2 68. Kxf2 g2 70. . . Rh1 71. b8=Q Rf1+
53
8
0L0Z0Z0Z 8
rZ0Z0ZkZ
7
Z0ZkZ0Z0 7
o0Z0m0Zb
6
0ZpOpZ0Z 6
0Z0ZPo0O
5
Z0O0Z0Z0 5
Z0ZpZ0Z0
4
0ZbO0Z0Z 4
QOpZ0o0Z
3
Z0A0ZpZ0 3
Z0Z0ZPl0
2
pZ0Z0JpZ 2
0Z0Z0ZNZ
1
Z0Z0SrZ0 1
Z0Z0ZRJR
32... Qg6 33. Rh4 Kh8 34. Rg4 Qe8 35. Qa1 Qf8 36. Qc3
72. Rxf1 gxf1=Q+ 73. Kg3 and the game eventually ended Bf5 37. Rxf4 a6 38. Re1 d3 39. Rxc4 Bxe6 40. Rd4 Bf5 41.
in a draw due to mutual threats and ensuing checks. 1/2–1/2 Qc7 Ng6 42. Kf2 Qxh6
Game AZ-17: AlphaZero Torpedo No-castling vs Alp-

8
rZ0Z0Z0j
haZero Torpedo No-castling This game was an experi-
ment combining the No-castling chess with Torpedo chess,
7
Z0L0Z0Z0
resulting in a highly tactical position. The first ten moves for
6
pZ0Z0onl
White and Black were sampled randomly from AlphaZero’s
opening “book”, with the probability proportional to the
5
Z0Z0ZbZ0
time spent calculating each move. The remaining moves
4
0O0S0Z0Z
3
Z0ZpZPZ0
8
rZ0Z0j0Z
2
0Z0Z0JNZ
7
o0Z0m0Zb
1
Z0Z0S0Z0
a b c d e f g h
6
0l0ZPo0O 43. Qc1 Qxc1 44. Rxc1 Ne7 45. Ne3 Bg6 46. Ra1 Nc6 47.
5
Z0ZpZ0o0 Rh4+ Kg7 48. b5 Nb8 49. Rc4 Bf7 50. Rc7 f4 51. Nd1 a4
52. Nc2 a2 53. Nxd3 Kf6 54. Rc8 Ra3 55. Nxf4 Nd7 56.
4
rOpO0A0Z Ne2 Ne5 57. b7 Rxf3+ 58. Kg2 Rb3 59. b8=Q Rxb8 60.
Z0Z0ZPZ0
3 Rxb8 and White went on to win the game easily. 1-0
2
0Z0L0JNZ B.6. Semi-torpedo
1
Z0Z0Z0SR In Semi-torpedo chess, we consider a partial extension to the
rules of pawn movement, where the pawns are allowed to
a b c d e f g h
move by two squares from the 2nd/3rd and 6th/7th rank for
White and Black respectively. This is a restricted version of
Here White executes a stunning ’double attack’: another variant we have considered (Torpedo chess) where
the option is extended to cover the entire board. Yet, even
27. Qc2!! Kg8
this partial extension adds lots of dynamic options and here
Black can’t afford to capture the Queen, due to the powerful we independently evaluate its impact on the arising play.
attack following 27... Bxc2 28. h8=Q+. White also had to
assess the consequences of 27... gxf4 B.6.1. M OTIVATION
28. Qxa4 Qxd4+ 29. Kxg3 gxf4+ 30. Kh2 Qf2 31. Rf1 As with Torpedo chess, the motivation in extending the pos-
Qg3+ 32. Kg1 sibilities for rapid pawn movement lies in adding dynamic,
54
attacking options to the middlegame. Yet, given that it is

only a partial extension, adding an extra rank for each side
from which the pawns can move by two squares, its impact
8
rZblka0s
on endgame patterns is much more limited. 7
ZpZ0Zpop
B.6.2. A SSESSMENT
6
pZnZ0m0Z
The assessment of the Semi-torpedo chess variant, as pro-
5
Z0Z0o0A0
vided by Vladimir Kramnik: 4
0Z0oPZ0Z
3
M0M0Z0Z0
2
POPZ0OPO
1
S0ZQJBZR
a b c d e f g h
Here, the game could continue 9. exd5 Bxa3

10. bxa3 Nd4 11. Bd3 Qa5, and the position is
assessed as equal by AlphaZero. This variation
illustrates nicely how the torpedo moves provide
not only an additional attacking option for White,
“
Compared to Classical chess, the pawns that
but also additional equalizing options for Black,
have been played to the 3rd/6th rank become
depending on the position.
much more useful, which manifests in several
ways. First, prophylactic pawn moves to h3/h6 Semi-torpedo chess seems to be more decisive
and a3/a6 now allow for a subsequent torpedo than Classical chess, and less decisive than Tor-
push. Having played h3 for example, it is now pos- pedo chess. It is an interesting variation, to be
sible to play the pawn to h5 in a single move. This potentially considered by those who like the gen-
also means, if the goal was to push the pawn to h5 eral middlegame flavor of Torpedo chess, but are
in two moves, that there are two ways of achiev-
ing it – either via h4 and h5 or via h3 and h5 –
and doing the latter does not expose a weakness
on the g4 square and can thus be advantageous.
Secondly, fianchetto setups now allow for addi-
unwilling to abandon existing endgame theory.
B.6.3. M AIN L INES

Here we discuss “main lines” of AlphaZero under Semi-
torpedo chess, when playing with roughly one minute per
”
tional dynamic options. The g3 pawn can now be
pushed to g5 in a single move, to attack a knight move from a particular fixed first move. Note that these
on f6 – and vice versa. Thirdly, openings where are not purely deterministic, and each of the given lines is
one of the central pawns is on the 3rd/6th rank merely one of several highly promising and likely options.
change – consider the Meran for example – the Here we give the first 20 moves in each of the main lines.
e3 pawn can now go to e5 in a single move.
Theory might change in other openings as well,
like for instance the Ruy Lopez with a7-a6, given
that there would be some lines where the tor-
pedo option of playing a6-a4 might force White
to adopt a slightly different setup. AlphaZero also
likes playing g6 early for Black, with a threat of
g4 in some lines, aimed against a knight on f3 if
White starts expanding in the center. As another Main line after e4 The main line of AlphaZero after 1. e4
example, consider a pretty standard opening se- in Semi-torpedo chess is:
quence in the Sicilian defence: 1. e4 c5 2. Nf3
1. e4 (book) c5 2. c3 Nf6 3. e5 Nd5 4. Bc4 e6 5. Nf3 Be7
Nc6 3. d4 cxd4 4. Nxd4 Nf6 5. Nc3 e5 6. Ndb5 d6
6. d4 d6 7. O-O O-O 8. Re1 Nc6 9. exd6 Qxd6 10. dxc5
– it turns out that here 7. Bg5 no longer keeps the
Qxc5 11. Nbd2 b6 12. b4 Qd6 13. Qc2 Bb7 14. a3 Nf6
advantage, because of 7. . . a6 8. Na3 followed up
15. Ne4 Qc7 16. Bd3 h6 17. c5 bxc5 18. Nxc5 Bxc5 19. bxc5
by a torpedo move 8. . . d4:
Na5 20. Ne5 Rac8
55
8
0ZrZ0skZ 8
rZ0ZrZ0Z
7
obl0Zpo0 7
opZ0Z0ak
6
0Z0Zpm0o 6
0ZnZ0lpo
5
m0O0M0Z0 5
Z0ZbZ0Z0
4
0Z0Z0Z0Z 4
0Z0O0Z0Z
3
O0ZBZ0Z0 3
Z0Z0A0O0
2
0ZQZ0OPO 2
PO0L0OBZ
1
S0A0S0J0 1
S0Z0ZRJ0
and after 21. Bb2 White would have compensation for the B.6.4. I NSTRUCTIVE GAMES
pawn. There are also tactical resources in this position, for
instance White could consider a more forcing line of play – Game AZ-18: AlphaZero Semi-torpedo vs AlphaZero
21. Bxh6!? gxh6 22. Qd2 Kg7 23. Re3 Rh8 24. Rg3+ Kf8 Semi-torpedo The first ten moves for White and Black
25. Rae1 h4 26. Rg7! Kxg7 27. Qg5+ Kf8 28. Qxf6 Rg8 have been sampled randomly from AlphaZero’s opening
29. Ng6+ Rxg6 30. Bxg6 – potentially leading to a draw by “book”, with the probability proportional to the time spent
perpetual check. calculating each move. The remaining moves follow best
Main line after d4 The main line of AlphaZero after 1. d4 1. d4 Nf6 2. c4 e6 3. e3 d5 4. cxd5 exd5 5. Nc3 Bd6 6. Nb5
in Semi-torpedo chess is: c6 7. Nxd6+ Qxd6 8. Bd3 Ne4 9. f3 Qb4+ 10. Bd2 Nxd2
1. d4 (book) Nf6 2. c4 e6 3. e3 d5 4. cxd5 exd5 5. Nc3 Bd6 11. Qxd2 Qd6 12. Ne2 O-O 13. O-O Nd7 14. g4 Nf6 15. Kg2
6. Bd3 O-O 7. Nge2 a6 8. O-O Re8 9. b3 Nc6 10. Ng3 Bg4 Bd7 16. Ng3 Kh8 17. Rae1 Rae8 18. Bb1 Ng8 19. h3 Ne7
11. f3 Bc8 12. a3 Ne7 13. Bb2 h6 14. Qd2 c6 15. e5 dxe4 20. f5
16. Ncxe4 Ned5 17. Nxd6 Qxd6 18. Rae1 Qd8 19. Rxe8+
Nxe8 20. Re1 Bd7
8
rZ0lnZkZ
7
ZpZbZpo0
8
0Z0Zrs0j
6
pZpZ0Z0o
7
opZbmpop
5
Z0ZnZ0Z0
6
0Zpl0Z0Z
4
0Z0O0Z0Z
5
Z0ZpZPZ0
3
OPZBZPM0
4
0Z0O0ZPZ
2
0A0L0ZPO
3
Z0Z0O0MP
1
Z0Z0S0J0
2
PO0L0ZKZ
a b c d e f g h 1
ZBZ0SRZ0
a b c d e f g h
in Semi-torpedo chess is:
1. c4 (book) c5 2. g3 g6 3. Bg2 Bg7 4. e3 e6 5. d4 cxd4
6. exd4 Ne7 7. Nc3 O-O 8. Nge2 d5 9. cxd5 Nxd5 10. h4
Here we see the first torpedo move of the game, f3-f5, claim-
Bd7 11. Nxd5 exd5 12. Be3 Re8 13. Nc3 Nc6 14. O-O Be6
ing space before Black has the chance to play f5.
15. h5 h6 16. hxg6 fxg6 17. Qd2 Kh7 18. Ne2 Qf6 19. Nf4
Bf7 20. Nxd5 Bxd5 20. . . f6 21. a3 b6 22. Nh5 Rb8 23. Qf2 b4
56
8
0s0Z0s0j 8
0ZrZ0s0j
7
o0Zbm0op 7
o0Zbm0op
6
0Zpl0o0Z 6
0Z0l0o0Z
5
Z0ZpZPZN 5
O0ZpOPZ0
4
0o0O0ZPZ 4
0ZpO0MPZ
3
O0Z0O0ZP 3
ZpZ0ZQZP
2
0O0Z0LKZ 2
0O0ZRZKZ
1
ZBZ0SRZ0 1
ZBZRZ0Z0
Black utilizes a torpedo move of its own, b6-b4, to initiate

White expands in the center with another torpedo move,
counterplay on the queenside.
e3-e5.
24. Qf4 Nc8 25. a4 b3 26. Qf3 Nb6 27. Nf4 Rbe8 28. Re2
32. . . dxe4 33. Bxe4 Rfe8 34. Ne6 Nc6 35. Bxc6 Bxc6
c4
36. d5 Ba8 37. Re3 Bb7 38. h5
And c6-c4 comes as another torpedo move, speeding up the
queenside expansion. White chooses not to take en passant,
but to play a5 instead in reply.
29. a5 Nc8
8
0ZrZrZ0j
7
obZ0Z0op
8
0ZnZrs0j
6
0Z0lNo0Z
7
o0ZbZ0op
5
O0ZPZPZP
6
0Z0l0o0Z
4
0ZpZ0ZPZ
5
O0ZpZPZ0
3
ZpZ0SQZ0
4
0ZpO0MPZ
2
0O0Z0ZKZ
3
ZpZ0OQZP
1
Z0ZRZ0Z0
a b c d e f g h
2
0O0ZRZKZ
1
ZBZ0ZRZ0
a b c d e f g h
Here comes another torpedo advance, h3-h5, creating threats

on the kingside.
38. . . h6 39. Kh3 Qd7 40. Rc3 Re7 41. Qf1 Qb5 42. d6 Rd7
30. Rfe1 Ne7 31. Rd1 Rc8 32. e5 43. Qf4 Ba6
57
0ZrZ0Z0j
8 Game AZ-19: AlphaZero Semi-torpedo vs AlphaZero
Semi-torpedo The position below, with Black to move, is
7
o0ZrZ0o0 taken from a game that was played with roughly one minute
per move:
6
bZ0ONo0o
5
OqZ0ZPZP
4
0ZpZ0LPZ
3
ZpS0Z0ZK 8
0ZrlrZkZ
2
0O0Z0Z0Z 7
ZbZnZ0a0
1
Z0ZRZ0Z0 6
0o0o0Zpo
a b c d e f g h
5
o0ZPoPZn
44. Qd2 Qe5 45. Rg3 Rc6 46. Nf8 4
PZ0Z0Z0Z
8
0Z0Z0M0j
3
Z0MBA0OP
7
o0ZrZ0o0
2
0OQM0O0Z
6
bZrO0o0o
1
Z0ZRJ0ZR
a b c d e f g h
5
O0Z0lPZP
4
0ZpZ0ZPZ
3
ZpZ0Z0SK
0O0L0Z0Z
18. . . Bxd5 19. Nde4 Bxe4 20. Qb3+ Kh8 21. Nxe4 d4
2
1
Z0ZRZ0Z0
a b c d e f g h
0ZrlrZ0j
46. . . Rcxd6 47. Qb4 Qb5 48. Qxd6 Rxd6 49. Rxd6 Qb8
8
50. Ng6+ Kh7 51. Rxa6 Qb7
7
Z0ZnZ0a0
8
0Z0Z0Z0Z 6
0o0Z0Zpo
7
oqZ0Z0ok 5
o0Z0oPZn
6
RZ0Z0oNo 4
PZ0oNZ0Z
5
O0Z0ZPZP 3
ZQZBA0OP
4
0ZpZ0ZPZ 2
0O0Z0O0Z
3
ZpZ0Z0SK 1
Z0ZRJ0ZR
2
0O0Z0Z0Z a b c d e f g h
1
Z0Z0Z0Z0
a b c d e f g h
52. Re3 Qh1+ 53. Kg3 Qg1+ 54. Kf3 Qf1+ 55. Kg3 Qg1+
Here, a torpedo move (d6-d4) unleashes a tactical sequence.
56. Kf3 Qf1+ 57. Ke4 Qg2+ 58. Kf4 Qf2+ 59. Ke4 Qg2+
60. Kf4 Qf2+ 61. Ke4 Qg2+ 1/2–1/2 22. Nd6 Rf8 23. Be2 Rc7 24. fxg6 Nc5
58
8
0Z0l0s0j 8
0Z0l0Z0j
7
Z0s0Z0a0 7
Z0Z0Z0a0
6
0o0M0ZPo 6
0Z0Z0ZPo
5
o0m0o0Zn 5
Z0Z0Z0ZB
4
PZ0o0Z0Z 4
0Z0ZpZ0Z
3
ZQZ0A0OP 3
Z0ZrZ0OP
2
0O0ZBO0Z 2
0O0Z0Z0Z
1
Z0ZRJ0ZR 1
Z0L0J0S0
25. Qxb6 Nxa4 26. Nf7+

37. Rg2 Qa5+ 38. Kf1 Qf5+ 39. Qf4 Qxf4+ 40. gxf4 Rxh3
41. Bd1 Rh4 42. Kf2 Rxf4+ 43. Ke3 Rf1 44. Bg4 Rf6 with
8
0Z0l0s0j a draw soon to follow. 1/2–1/2
7
Z0s0ZNa0
6
0L0Z0ZPo Game AZ-20: AlphaZero Semi-torpedo vs AlphaZero
5
o0Z0o0Zn Semi-torpedo The first ten moves for White and Black
4
nZ0o0Z0Z “book”, with the probability proportional to the time spent
Z0Z0A0OP
3
2
0O0ZBO0Z 1. d4 Nf6 2. Nf3 d5 3. c4 e6 4. a3 dxc4 5. e3 c6 6. Bxc4 b5
Z0ZRJ0ZR
1 7. Bd3 Bb7 8. Nc3 a6 9. e5
a b c d e f g h
26. . . R8xf7 27. Qxa5 Rfd7 28. Qxa4 dxe3

8
rm0lka0s
7
ZbZ0Zpop
8
0Z0l0Z0j 6
pZpZpm0Z
7
Z0srZ0a0 5
ZpZ0O0Z0
6
0Z0Z0ZPo 4
0Z0O0Z0Z
5
Z0Z0o0Zn 3
O0MBZNZ0
4
QZ0Z0Z0Z 2
0O0Z0OPO
3
Z0Z0o0OP 1
S0AQJ0ZR
2
0O0ZBO0Z a b c d e f g h
1
Z0ZRJ0ZR Here we see another typical central torpedo move (e3-e5),
a b c d e f g h
claiming space.
29. Bxh5 Rxd1+ 30. Qxd1 exf2+ 31. Kxf2 Rd7 32. Qc1 9. . . Nd5 10. Be4 Be7 11. h3 Nxc3 12. bxc3 Nd7 13. O-O
Rd2+ 33. Ke1 Rd3 34. Kf2 Rd2+ 35. Ke1 Rd3 36. Rg1 e4 Rb8 14. Qe2 c4
59
8
0s0lkZ0s 8
rZ0l0skZ
7
ZbZnapop 7
ZpZbZpZ0
6
pZ0ZpZ0Z 6
pZ0apZ0o
5
ZpZ0O0Z0 5
Z0ZpMno0
4
0ZpOBZ0Z 4
0Z0OnZ0Z
3
O0O0ZNZP 3
Z0OBA0ZN
2
0Z0ZQOPZ 2
PO0Z0OPO
1
S0A0ZRJ0 1
S0ZQS0J0
16. f3 Bxe5 17. dxe5 Nxe3 18. Rxe3 f5
Black uses a torpedo move as a counter (c6-c4), expanding

on the queenside.
8
rZ0l0skZ
15. Bxb7 Rxb7 16. Qe4 Rc7 17. Qg4 g6 18. a5 7
ZpZbZ0Z0
6
pZ0ZpZ0o
5
Z0ZpOpo0
8
0Z0lkZ0s 4
0Z0ZnZ0Z
7
Z0snapZp 3
Z0OBSPZN
6
pZ0ZpZpZ 2
PO0Z0ZPO
5
OpZ0O0Z0 1
S0ZQZ0J0
4
0ZpO0ZQZ a b c d e f g h
3
Z0O0ZNZP 19. exf6 Qb6 20. Qc1 Nxf6 21. Nf2 e4
2
0Z0Z0OPZ
1
S0A0ZRJ0
8
rZ0Z0skZ
a b c d e f g h 7
ZpZbZ0Z0
6
pl0Z0m0o
Another torpedo move follows (a3-a5), giving rise to a the- 5
Z0ZpZ0o0
matic pawn structure. 4
0Z0ZpZ0Z
Z0OBSPZ0
18. . . h5 19. Qg3 Nb8 20. d5 Qxd5 21. Bg5 Qd8 22. Rad1
3
Rd7 23. Bxe7 Qxe7 24. Ng5 O-O 25. Ne4 Rxd1 26. Rxd1
Rd8 27. Rd6 Rxd6 28. exd6 Qd8 29. Qe5 Nd7 30. Qd4 Qh4
and the game eventually ended in a draw. 1/2–1/2
2
PO0Z0MPO
1
S0L0Z0J0
a b c d e f g h
Here we see a torpedo move e6-e4 being used in a tactical

Game AZ-21: AlphaZero Semi-torpedo vs AlphaZero
sequence in center of the board.
Semi-torpedo The position below, with White to move,
is taken from a game that was played with roughly one 22. fxe4 Rae8 23. e5 Ng4 24. Nxg4 Bxg4 25. Kh1 Rf2
minute per move: 26. b4 Ref8
60
0Z0Z0skZ
8 B.7. Pawn-back
In the Pawn-back variation of chess, the pawns are allowed
7
ZpZ0Z0Z0 to move one square backwards, up to the 2nd/7th rank for
6
pl0Z0Z0o White and Black respectively. In addition, if the pawn moves
back to its starting rank, it is allowed to move by two squares
5
Z0ZpO0o0 again on its next move. In this particular implementation,
the two-square pawn move is always allowed from the 2nd
4
0O0Z0ZbZ or the 7th rank, regardless of whether the pawn has moved
3
Z0OBS0Z0 before. A different implementation of this variation of chess
might consider disallowing this, though it is unlikely to
2
PZ0Z0sPO make a big difference. Because the pawns are allowed to
move backwards and pawn moves are now reversible in this
1
S0L0Z0ZK implementation of chess, the 50 move rule is modified so
a b c d e f g h that 50 moves without captures lead to a draw, regardless of
whether any pawn moves were made in the meantime.
27. Qe1 Be6 28. Re2 R2f4 29. a3 Kg7 30. h3 Qd8 31. Re3
h5 32. Rd1 g4 33. Rd2 h4 34. hxg4 Qg5 B.7.1. M OTIVATION
In Classical chess, pawns that move forwards leave weak-
8
0Z0Z0s0Z nesses behind. Some of these remain long-term weaknesses,
resulting in squares that can be easily occupied by the op-
7
ZpZ0Z0j0 ponent’s pieces. If the pawns could move backwards, they
could come back to help fight for those squares and therefore
6
pZ0ZbZ0Z reduce the number of weaknesses in a position. Allowing
5
Z0ZpO0l0 the pawns to move backwards would therefore make it easier
to push them forward, as the effect would not be irreversible.
4
0O0Z0sPo This might make advancing in a position easier, but equally,
it could provide defensive options for the weaker side, such
3
O0OBS0Z0 as retreating from a less favourable situation and covering a
2
0Z0S0ZPZ weaknesses in front of the king.
1
Z0Z0L0ZK B.7.2. A SSESSMENT
a b c d e f g h The assessment of the Pawn-back chess variant, as provided
by Vladimir Kramnik:
35. Rh3 Rxg4 36. Qe3 d4 37. cxd4 R8f4 38. Rf3 Bd5
“
There are quite a few educational motifs in this
variation of chess. The backward pawn moves
8
0Z0Z0Z0Z can be used to open the diagonals for the bishops,
7
ZpZ0Z0j0 or make squares available for the knights. The
bishops can therefore become more powerful, as
6
pZ0Z0Z0Z they are easier to activate. The pawns can be
pushed in the center more aggressively than in
5
Z0ZbO0l0 classical chess, as they can always be pulled back.
4
0O0O0sro Exposing the king is not as big of an issue, as the
pawns can always move back to protect. Weak
3
O0ZBLRZ0 squares are much less important for positional
assessment in this variation, given that they can
2
0Z0S0ZPZ almost always be protected via moving the pawns
1
Z0Z0Z0ZK back.
It was interesting to see AlphaZero’s strong pref-
a b c d e f g h
erence for playing the French defence under
these rules, the point being that the light-squared
39. Rxf4 Qxf4 40. Qxf4 Rxf4 41. Kg1 Rxd4 and the game
bishop is no longer bad, as it can be developed
soon ended in a draw. 1/2–1/2
61
via c8-b7 followed by a timely d5-d6 back-move. from a particular fixed first move. Note that these are not
purely deterministic, and each of the given lines is merely
Other openings change as well. After the standard one of several highly promising and likely options. Here we
1. e4 e5 2. Nf3 Nc6, there comes a surprise: 3. c4! give the first 20 moves in each of the main lines, regardless
of the position.
8
rZblkans
opopZpop
7 Main line after e4 The main line of AlphaZero after 1. e4
in Pawn-back chess is:
6
0ZnZ0Z0Z 1. e4 (book) e6 2. Nc3 d5 3. d4 Nf6 4. e5 Nfd7 5. f4 c5 6. Nf3
5
Z0Z0o0Z0 a6 7. Be3 b5 8. f5 Nc6 9. fxe6 fxe6 10. e4 cxd4 11. Nxd4
Nxd4 12. Qxd4 b4 13. Ne2 Nf6 14. exd5 Qxd5 15. Nf4
4
0ZPZPZ0Z Qxd4 16. Bxd4 Bd6 17. Nd3 a5 18. Be5 Ke7 19. Bxd6+
Z0Z0ZNZ0
Kxd6 20. O-O-O Ke7
3
2
PO0O0OPO 8
rZbZ0Z0s
1
SNAQJBZR 7
Z0Z0j0op
a b c d e f g h
6
0Z0Zpm0Z
It is followed by 3. . . Bc5 4. e3 (a back-move!)
Bb6 5. d4 d6
5
o0Z0Z0Z0
4
0o0Z0Z0Z
8
rZblkZns 3
Z0ZNZ0Z0
7
opo0Zpop 2
POPZ0ZPO
6
0ano0Z0Z 1
Z0JRZBZR
5
Z0Z0o0Z0 a b c d e f g h
4
0ZPO0Z0Z Main line after d4 The main line of AlphaZero after 1. d4
3
Z0Z0ONZ0 in Pawn-back chess is:
2
PO0Z0OPO 1. d4 (book) d5 2. e3 Nf6 3. Nf3 e6 4. c4 Be7 5. b3 O-O
6. cxd5 exd5 7. Bd3 Re8 8. Bb2 a5 9. O-O Bf8 10. Nc3
1
SNAQJBZR c6 11. Qc2 b6 12. Ne2 Ra7 13. Rac1 Rc7 14. Rfe1 Bb4
15. Nc3 Ba6 16. Bxa6 Nxa6 17. h4 b7 18. a3 Bf8 19. Ne2
a b c d e f g h
Rc8 20. Nf4 Nc7
Who would have guessed that we are on move 5,
0ZrlrakZ
after the game having started with e4 e5? 8
The Pawn-back version of chess allows for more
fluid and flexible pawn structures and could po- 7
Zpm0Zpop
0ZpZ0m0Z
tentially be interesting for players who like such 6
strategic manoeuvring. Given that Pawn-back
chess offers additional defensive resources, win-
ning with White seems to be slightly harder, so the
5
o0ZpZ0Z0
variant might also appeal to players who enjoy 4
0Z0O0M0O
defending and attackers looking for a challenge.
B.7.3. M AIN L INES

Here we discuss “main lines” of AlphaZero under Pawn-
back chess, when playing with roughly one minute per move
” 3
2
1
OPZ0ONZ0
0AQZ0OPZ
Z0S0S0J0
a b c d e f g h
62
0Zks0Z0Z
Main line after c4 The main line of AlphaZero after 1. c4 8
in Pawn-back chess is:
1. c4 (book) e5 2. e3 c5 3. Nc3 Nc6 4. Nf3 f5 5. d4 e4 6. Nd2
7
opZnlpo0
Nf6 7. d5 Ne5 8. Be2 g6 9. d4 Nf7 10. dxc5 Bxc5 11. a3 Bf8
12. b4 Bg7 13. Bb2 O-O 14. O-O d6 15. a4 Be6 16. Qb3 a5
6
0Zpo0mbZ
17. Rfd1 b6 18. bxa5 bxa5 19. Qa3 e5 20. c5 Qb8 5
Z0ZPZ0Z0
8
rl0Z0skZ
4
0ZPZ0Z0Z
7
Z0Z0Znap
3
Z0O0APO0
6
0Z0obmpZ
2
PZ0ZQJ0Z
5
o0O0opZ0
1
S0Z0ZBM0
a b c d e f g h
4
PZ0Z0Z0Z
3
L0M0O0Z0
2
0A0MBOPO
S0ZRZ0J0
Black is putting pressure on d5, so White uses the back-
1
move d5-d4 option to reconfigure the central pawn structure,
a b c d e f g h rather than release the tension.
B.7.4. I NSTRUCTIVE GAMES 17. d4 d5 18. c5
Game AZ-22: AlphaZero Pawn-back vs AlphaZero

Pawn-back The first ten moves for White and Black have
8
0Zks0Z0Z
1. e4 c6 2. d4 d5 3. e5 Bf5 4. h4 h5 5. c4 d6
7
opZnlpo0
6
0ZpZ0mbZ
8
rm0lkans 5
Z0OpZ0Z0
7
opZ0opo0 4
0Z0O0Z0Z
6
0Zpo0Z0Z 3
Z0O0APO0
5
Z0Z0ObZp 2
PZ0ZQJ0Z
4
0ZPO0Z0O 1
S0Z0ZBM0
3
Z0Z0Z0Z0 a b c d e f g h
2
PO0Z0OPZ
1
SNAQJBMR
a b c d e f g h
Black and White repeat back-moves a couple of times. Each
Here we see d5-d6 as the first back-move of the game, time that Black challenges the c5 pawn via a d5-d6 back-
challenging White’s (over)extended center – an option that move, White responds by c5-c4, refusing to exchange on
would not have been available in classical chess. that square.
6. exd6 exd6 7. d5 Be7 8. Nc3 Bxh4 9. Be3 Qe7 10. g3 Bf6 18. . . d6 19. c4 d5 20. Rc1 Rh8 21. c5 d6 22. c4 d5 23. Bf4
11. Rxh5 Rxh5 12. Qxh5 Bg6 13. Qe2 Bxc3+ 14. bxc3 Nd7 Qxe2+ 24. Bxe2 dxc4 25. Bxc4 b5 26. Bf1 Nd5 27. Bd2
15. f3 O-O-O 16. Kf2 Ngf6 N7b6
63
8
0ZkZ0Z0s 8
0ZkZ0Z0Z
7
o0Z0Zpo0 7
o0Z0Zpo0
6
0mpZ0ZbZ 6
0m0Z0ZbZ
5
ZpZnZ0Z0 5
ZpZnZ0Z0
4
0Z0O0Z0Z 4
0ZpZ0O0Z
3
Z0O0ZPO0 3
Z0ZPZ0O0
2
PZ0A0J0Z 2
PZPA0J0s
1
Z0S0ZBM0 1
Z0S0ZBM0
At this point it should come as no surprise how White should

respond to the rook invasion using a back-move g3-g2!
31. g2 c3 32. Be1 Bf5 33. Nf3 Rh6 34. g3 Rd6 35. Bg2 a6
Here we see an example of how back-moves can help cover 36. a3 Na4 37. Ng5 f6 38. Ne4 Rd7 39. Kf3 Kc7 40. Bf2
weak squares. Black is threatening to invade on the light
0Z0Z0Z0Z
squares on the queenside at an opportune moment, but White 8
utilizes a back-move d4-d3 and protects c4. This, however,
enables Black to go forward and Black takes the opportunity
to play c6-c5.
7
Z0jrZ0o0
28. d3 c5 29. f4 c4
6
pZ0Z0o0Z
5
ZpZnZbZ0
4
nZ0ZNO0Z
3
O0oPZKO0
8
0ZkZ0Z0s
2
0ZPZ0ABZ
7
o0Z0Zpo0
1
Z0S0Z0Z0
a b c d e f g h
6
0m0Z0ZbZ
5
ZpZnZ0Z0 40. . . Bxe4 41. Kxe4 Kd6 42. Re1 Rc7 43. Kf5 Ne7+
44. Kg4 c4 45. d2
4
0ZpZ0O0Z
3
Z0OPZ0O0 8
0Z0Z0Z0Z
2
PZ0A0J0Z 7
Z0s0m0o0
1
Z0S0ZBM0 6
pZ0j0o0Z
a b c d e f g h 5
ZpZ0Z0Z0
4
nZpZ0OKZ
3
O0Z0Z0O0
White decides to keep retreating here and not give up the
2
0ZPO0ABZ
light squares with a back-move c3-c2. 1
Z0Z0S0Z0
30. c2 Rh2+ a b c d e f g h
64
Here we see both Black and White having retreated from the 14. . . exf4 15. exd5 Qb6+ 16. Kh1 Na7 17. Qe1+
interaction on the queenside, Black via a back-move c3-c4
and White by playing the d-pawn back to d2. The game
soon ended in a draw.
8
rZbZka0s
45. . . f5+ 46. Kf3 c3 47. d3 c4 48. d2 c3 49. d3 c4 50. g4
7
mpZ0Zpo0
cxd3 51. cxd3 Rc3 52. gxf5 Rxd3+ 53. Kg4 Rd2 54. Re6+
Kd7 55. Kf3 Rd3+ 56. Kg4 Rd2 57. Kf3 Rd3+ 58. Kg4 Rd2
6
0l0Z0m0o
1/2–1/2 5
oNZPZPZ0
4
PZ0Z0o0Z
Pawn-back The position below, with Black to move, is 3
ZNZ0Z0O0
0OPZ0ZBO
taken from a game that was played with roughly one minute
2
per move:
1
S0Z0LRZK
8
rZblka0s a b c d e f g h
7
ZpZ0Zpo0 17. . . Kd8 18. N5d4 Bd7 19. Nxa5 fxg3 20. Rd1 Bb4
21. Nxb7+
6
0ZnZ0m0o
5
oNZpZPZ0 8
rZ0j0Z0s
4
PZ0ZpA0Z 7
mNZbZpo0
3
ZNZ0Z0O0 6
0l0Z0m0o
2
0OPZPZBO 5
Z0ZPZPZ0
1
S0ZQZRJ0 4
Pa0M0Z0Z
a b c d e f g h 3
Z0Z0Z0o0
White is targeting c7 with the bishop and the knight, but 2
0OPZ0ZBO
Z0ZRLRZK
here Black plays a back-move, e4-e5. It initiates a long 1
forced tactical sequence, showcasing that things can indeed
get quite tactical in this variation of chess, depending on the a b c d e f g h
line of play.
Sacrificing another piece!
13. . . e5 14. e4
21. . . Qxb7 22. Qxg3 Rg8 23. Ne6+
8
rZblka0s 8
rZ0j0ZrZ
7
ZpZ0Zpo0 7
mqZbZpo0
6
0ZnZ0m0o 6
0Z0ZNm0o
5
oNZpoPZ0 5
Z0ZPZPZ0
4
PZ0ZPA0Z 4
Pa0Z0Z0Z
3
ZNZ0Z0O0 3
Z0Z0Z0L0
2
0OPZ0ZBO 2
0OPZ0ZBO
1
S0ZQZRJ0 1
Z0ZRZRZK
AlphaZero decides to sacrifice a piece for the initiative! Third consecutive piece sacrifice by White!
65
rZ0lka0s
23. . . fxe6 24. dxe6 Qc7 25. Bxa8 Qxg3 26. hxg3 Kc7 8
27. Bg2 Bc6
7
ZbZnZpop
8
0Z0Z0ZrZ
6
pZnZpZ0Z
7
m0j0Z0o0
5
ZpopO0Z0
6
0ZbZPm0o
4
0Z0O0O0Z
5
Z0Z0ZPZ0
3
O0O0ANZ0
4
Pa0Z0Z0Z
2
0O0ZNZPO
3
Z0Z0Z0O0
1
S0ZQJBZR
a b c d e f g h
2
0OPZ0ZBZ
1
Z0ZRZRZK
a b c d e f g h
This looks like a pretty normal French position, but here

It’s time to take stock – White has a rook and 4 pawns for 3 comes Black’s main equalizing resource, a back move d5-
pieces, a very unusual material imbalance. d6! Maybe that’s all that was needed to make the French an
28. Rf4 Rb8 29. Rc4 Bd6 30. Rb1 Kb6 31. Re1 Nh5 32. g4 undeniably good opening for Black?
Nf6 33. Re3 Rc8 34. Rec3 Be5 35. a5+ Ka6 36. Rxc6+ Rxc6 10. . . d6
37. Rxc6+ Nxc6 38. Bxc6 Bxb2 39. e7 Kxa5
8
0Z0Z0Z0Z
7
Z0Z0O0o0 8
rZ0lka0s
6
0ZBZ0m0o 7
ZbZnZpop
5
j0Z0ZPZ0 6
pZnopZ0Z
4
0Z0Z0ZPZ 5
Zpo0O0Z0
3
Z0Z0Z0Z0 4
0Z0O0O0Z
2
0aPZ0Z0Z 3
O0O0ANZ0
1
Z0Z0Z0ZK 2
0O0ZNZPO
a b c d e f g h
1
S0ZQJBZR
And the game soon ended in a draw. 1/2–1/2 a b c d e f g h

been sampled randomly from AlphaZero’s opening “book”, This completely changes the nature of the position, as
with the probability proportional to the time spent calculat- the center is suddenly not static and Black’s light-squared
ing each move. The remaining moves follow best play, at bishop can find good use on the a8-h1 diagonal.
11. Ng3 dxe5 12. fxe5 Qb6 13. Bf2 Rd8 14. Qb1 cxd4
1. e4 e6 2. Nc3 d5 3. d4 Nf6 4. e5 Nfd7 5. f4 c5 6. Nf3 a6 15. cxd4 b4 16. Be2 bxa3 17. bxa3 Qa5+ 18. Kf1 Rb8
7. a3 Nc6 8. Be3 b5 9. Ne2 Bb7 10. c3 19. h4 Be7 20. Kg1 O-O 21. Qd3 Rfc8 22. Kh2 Qd8
66
8
0srl0ZkZ 8
rZblkZ0s
7
ZbZnapop 7
Z0Z0m0ap
6
pZnZpZ0Z 6
pZ0ZpSpZ
5
Z0Z0O0Z0 5
ZpZpm0Z0
4
0Z0O0Z0O 4
0Z0M0Z0Z
3
O0ZQZNM0 3
Z0MBA0Z0
2
0Z0ZBAPJ 2
POPZ0ZPO
1
S0Z0Z0ZR 1
S0ZQZ0ZK
Here AlphaZero prefers a solid back-move h4-h3 to a further

expansion with h5.
23. h3 Na5 24. Rhc1 Nf8 25. Qe3
8
0srl0mkZ 15. Nxe6 Bxe6 16. Rxe6 O-O 17. Bg5 Ra7 18. Be2 Nf7
19. Bh4 g5
7
ZbZ0apop
6
pZ0ZpZ0Z
5
m0Z0O0Z0
4
0Z0O0Z0Z
3
O0Z0LNMP
8
0Z0l0skZ
2
0Z0ZBAPJ
7
s0Z0mnap
1
S0S0Z0Z0
6
pZ0ZRZ0Z
a b c d e f g h 5
ZpZpZ0o0
The a6 pawn is under pressure from the e2 bishop, and
4
0Z0Z0Z0A
simply moves back to a7. The game soon fizzles out to a
draw.
3
Z0M0Z0Z0
25. . . a7 26. a4 Ng6 27. Rab1 Rxc1 28. Qxc1 Rc8 29. Qd1
2
POPZBZPO
Ba8 30. Ba6 Rb8 31. Bf1 Rxb1 32. Qxb1 Bc6 33. Bb5 Qb8
34. Qd3 Qb7 35. Ne2 h6 36. Bg3 Be4 37. Qe3 Bb4 38. Bf2
1
S0ZQZ0ZK
a b c d e f g h
Bd5 39. Qd3 Be4 40. Qe3 Bc6 41. Qd3 Be7 42. Bg3 Be4
43. Qe3 Bb4 44. Bf2 1/2–1/2

with the probability proportional to the time spent calculat- Here we see that moves like g5, that would potentially other-
ing each move. The remaining moves follow best play, at wise be quite weakening, are perfectly playable, given that
roughly one minute per move. the g-pawn can (and soon will) move back to g6, and in the
meantime the threatening bishop is forced to move back and
1. e4 e6 2. Nc3 d5 3. d4 Nf6 4. e5 Nfd7 5. Be3 c5 6. f4 a6
unpin the Black knight on e7.
7. Nf3 b5 8. f5 Nc6 9. fxe6 fxe6 10. Bd3 g6 11. O-O cxd4
12. Nxd4 Ndxe5 13. Kh1 Ne7 14. Rf6 Bg7 20. Bf2 d4 21. Ne4 Nf5 22. Qd3 g6
67
8
0Z0l0skZ 8
rZ0lrZkZ
7
s0Z0Znap 7
ZpZ0Zpo0
6
pZ0ZRZpZ 6
0Zpa0m0o
5
ZpZ0ZnZ0 5
o0ZpZ0Z0
4
0Z0oNZ0Z 4
0m0O0Z0O
3
Z0ZQZ0Z0 3
ZPZ0ONZ0
2
POPZBAPO 2
PA0ZNOPZ
1
S0Z0Z0ZK 1
S0ZQZRJ0
After moving the pawn back to g6 with a back-move, Black Here we see the first back-move of the game, opening the
safeguards the kingside, justifying the previous g5 pawn diagonal for the White bishop – d4-d3!
push, which was helpful in achieving development. 18. d3 Nd7 19. a4 c5
23. g4 Ne3 24. Bxe3 dxe3 25. Qxe3 Re7 26. Rxe7 Qxe7
27. a4 Nd6 28. Bd3 Bxb2 29. Rb1 Qe5 30. axb5 axb5
31. Qe2 Ba3
8
rZ0lrZkZ
7
ZpZnZpo0
8
0Z0Z0skZ 6
0Z0a0Z0o
7
Z0Z0Z0Zp 5
o0opZ0Z0
6
0Z0m0ZpZ 4
Pm0Z0Z0O
5
ZpZ0l0Z0 3
ZPZPONZ0
4
0Z0ZNZPZ 2
0A0ZNOPZ
3
a0ZBZ0Z0 1
S0ZQZRJ0
0ZPZQZ0O
2 a b c d e f g h
ZRZ0Z0ZK
1 Just having played a4 on the previous move, White plays a
back-move a4-a3 to challenge the b4 knight, given that the
a b c d e f g h circumstances have changed due to Black having played c5.
As a mirror-motif to Black’s g5-g6, here White plays g4-g3 20. a3 Na6 21. d4
to improve the safety of its king.
32. g3 Nxe4 33. Qxe4 Qxe4 34. Bxe4 b4 and the game soon
8
rZ0lrZkZ
ended in a draw. 1/2–1/2 7
ZpZnZpo0
Game AZ-26: AlphaZero Pawn-back vs AlphaZero 6
nZ0a0Z0o
o0opZ0Z0
5
4
0Z0O0Z0O
OPZ0ONZ0
1. d4 Nf6 2. c4 e6 3. Nf3 a6 4. Nc3 d5 5. cxd5 exd5 6. b3 Bb4
7. Bd2 Be7 8. e3 O-O 9. Bc1 Bf5 10. Bd3 Bxd3 11. Qxd3 c6
2
0A0ZNOPZ
12. Qc2 Re8 13. O-O a5 14. h4 Na6 15. Ne2 Nb4 16. Qd1 1
S0ZQZRJ0
Bd6 17. Bb2 h6 a b c d e f g h
68
0ZrZ0akZ
White goes back to the previous plan and plays the pawn to 8
d4 again, despite having moved it back before, showcasing
the fluidity of pawn structures Black responds by moving
the c-pawn back, to avoid having an isolated pawn.
7
ZpZ0Z0Z0
21. . . c6 22. g3 Nc7 23. a4 Ne6 24. Kg2 Nf6 25. Rc1 Bf8
6
0ZpZ0o0o
26. d3 5
o0ZpZ0o0
4
PZ0OrZPO
8
rZ0lrakZ
3
ZPZ0ONZ0
7
ZpZ0Zpo0
2
0Z0Z0OKZ
6
0ZpZnm0o
1
Z0S0ZRZ0
a b c d e f g h
5
o0ZpZ0Z0
4
PZ0Z0Z0O
3
ZPZPONO0
2
0A0ZNOKZ Black is attacking White’s pawn on g4, so it just moves back
Z0SQZRZ0
1 to g3.
a b c d e f g h 43. g3 Kf7 44. Rh1 g4 45. Ne1 Bd6 46. h3 g5
White opens the Bishop’s diagonal with a back-move, again.

26. . . Rc8 27. Nfd4 Nxd4 28. Bxd4 c5 29. Ba1 Nh5 30. Ng1
g6 31. Nf3 Ng7 32. Bxg7 Bxg7 33. d4 c6 34. Qd2 Bf8
35. h5 g5 8
0ZrZ0Z0Z
7
ZpZ0ZkZ0
8
0ZrlrakZ
6
0Zpa0o0o
7
ZpZ0ZpZ0
5
o0ZpZ0o0
6
0ZpZ0Z0o
4
PZ0OrZ0Z
5
o0ZpZ0oP
3
ZPZ0O0OP
4
PZ0O0Z0Z
2
0Z0Z0OKZ
3
ZPZ0ONO0
1
Z0S0M0ZR
a b c d e f g h
2
0Z0L0OKZ
1
Z0S0ZRZ0
a b c d e f g h
After having been challenged by a h4-h3 back-move, Black

Having just played h5, White plays h5-h4 now, to attack retreats with g4-g5 as well.
Black’s g-pawn again. They repeat once before continuing
47. Nd3 Ke8 48. h4 g4 49. h3 g5 50. h4 g6 51. Kf3 Kd7
with other plans.
52. Nf4 Rg8 53. Ne2 h5 54. Nf4 Bxf4 55. gxf4 b6 56. a3 f7
36. h4 g6 37. h5 g5 38. Qd3 Rc7 39. Qf5 Qc8 40. Qxc8 57. Rc2 Ra8 58. Rb1 Re6 59. Ke2 Rf6 60. 2f3 Rf5 61. Kf2
Rcxc8 41. h4 f6 42. g4 Re4 d6 62. a4 Re8 63. Rbc1 b7
69
8
0Z0ZrZ0Z 8
0Z0Z0Z0Z
7
ZpZkZpZ0 7
ZPZ0Z0Z0
6
0Zpo0ZpZ 6
0Z0Z0Z0Z
5
o0Z0ZrZp 5
Z0Z0Z0Z0
4
PZ0O0O0O 4
0Z0j0J0Z
3
ZPZ0OPZ0 3
ZrZ0Z0Z0
2
0ZRZ0J0Z 2
0Z0Z0Z0Z
1
Z0S0Z0Z0 1
Z0Z0Z0Z0
a b c d e f g h
a b c d e f g h
White takes aim at the c6 pawn, but Black simply plays b6-
b7, guarding it. With no clear way forward in this position, In classical chess, White would be completely lost.
and after many more pawn structure reconfigurations, the Here, White can play b7-a7 or b7-c7, changing
game unsurprisingly ended in a draw. 1/2–1/2 files. The rook can follow, but the pawn can al-
ways step aside. In this particular position, after
B.8. Pawn-sideways b7-c7, Rc3, c7-d7 – Black has no way of stopping
the pawn from queening, and instead of losing –
In the Pawn-sideways version of chess, pawns are allowed White actually wins!
an additional option of moving sideways by one square,
It almost appears as if being a pawn up might give
when available.
better chances of winning than being up a piece
for a pawn. In fact, AlphaZero often chooses to
B.8.1. M OTIVATION
play with two pawns against a piece, or a mi-
Allowing the pawns to move laterally introduces lots of new nor piece and a pawn against a rook, suggesting
tactics into chess, while keeping the pawn structures very that pawns are indeed more valuable here than in
flexible and fluid. It makes pawns much more powerful than classical chess.
before and drastically increases the complexity of the game, This variant of chess is quite different and at times
as there are many more moves to consider at each juncture – hard to understand, but could be interesting for
and no static weaknesses to exploit. players who are open to experimenting with few
B.8.2. A SSESSMENT
The assessment of the Pawn-sideways chess variant, as pro-
vided by Vladimir Kramnik:
attachments to the original game!
B.8.3. M AIN L INES

Here we discuss “main lines” of AlphaZero under Pawn-
”
“
This is the most perplexing and “alien” of all sideways chess, when playing with roughly one minute per
variants of chess that we have considered. Even move from a particular fixed first move. Note that these
after having looked at how AlphaZero plays Pawn- are not purely deterministic, and each of the given lines is
side chess, the principles of play remain somewhat merely one of several highly promising and likely options.
mysterious – it is not entirely clear what each side Here we give the first 20 moves in each of the main lines,
should aim for. The patterns are very different regardless of the position.
and this makes many moves visually appear very
strange, as they would be mistakes in Classical
chess.
Lateral pawn moves change all stages of the game.
Endgame theory changes entirely, given that the
pawns can now “run away” laterally to the edge Main line after e4 The main line of AlphaZero after 1. e4
of the board, and it is hard to block them and pin in Pawn-sideways chess is:
them down. Consider, for instance, the following
1. e4 (book) c5 2. c3 b6 3. dd4 Bb7 4. Nd2 g6 5. Bd3 Bg7
position, with White to move:
6. Ngf3 a5
70
8
rm0lkZns 8
rm0l0skZ
7
ZbZpopap 7
ZbZ0opap
6
0o0Z0ZpZ 6
0o0o0mpZ
5
o0o0Z0Z0 5
o0ZpZ0Z0
4
0Z0OPZ0Z 4
0O0OPZ0Z
3
Z0OBZNZ0 3
O0OBZNZ0
2
PO0M0OPO 2
0Z0M0OPO
1
S0AQJ0ZR 1
ZRAQZRJ0
11. Qc2 Nxe4 12. Nxe4 dxe4 13. Bxe4 Bxe4 14. Qxe4 Nd7
15. Be3 ab5
The previous move (a5) seems very unusual to a Classical
chess player’s eye. Black chooses to disregard the cen-
tre, while creating a glaring weakness on b5. Yet, there
8
rZ0l0skZ
is method to this “madness”. It seems that rushing to grab 7
Z0Znopap
0o0o0ZpZ
space early is not good in this setup, so White’s most promis-
6
ing plan according to AlphaZero is to prepare b4. Apart
from fighting against that advance, a5 prepares for playing
a5-b5! later in this line, as we will see. Yet, this whole line
5
ZpZ0Z0Z0
of play is hard to grasp as it violates the Classical chess 4
0O0OQZ0Z
O0O0ANZ0
principles.
3
7. O-O d6 8. Rb1 Nf6 9. a3 O-O 10. b4
2
0Z0Z0OPO
1
ZRZ0ZRJ0
a b c d e f g h
8
rm0l0skZ As mentioned earlier, the a5 pawn finds a new purpose – on
7
ZbZ0opap b5! The b6 pawn will soon move to c6, in the process of
0o0o0mpZ
6 reconfiguring the pawn structure.
16. ab3 Nf6 17. Qd3 bc6 18. cc4 Qb8 19. a4 b4 20. c3 Rxa4
5
o0o0Z0Z0
4
0O0OPZ0Z 8
0l0Z0skZ
3
O0OBZNZ0 7
Z0Z0opap
2
0Z0M0OPO 6
0Zpo0mpZ
1
ZRAQZRJ0 5
Z0Z0Z0Z0
a b c d e f g h
4
roPO0Z0Z
3
Z0OQANZ0
White has achieved the desired advance, to which Black
2
0Z0Z0OPO
responds with a lateral move – c5-d5! 1
ZRZ0ZRJ0
10. . . cd5 a b c d e f g h
71
rZ0Z0skZ
Main line after d4 The main line of AlphaZero after 1. d4 8
in Pawn-sideways chess is:
1. d4 (book) d5 2. e3 e6 3. cc4 dxc4 4. Bxc4 a6 5. a4 c5
7
Zpl0mpap
6. Nf3 Nc6 7. Be2 cxd4 8. exd4 g6 9. b3 Nge7 10. Bb2 Bg7
11. Na3
6
pZnZbo0Z
5
Z0ZpZ0Z0
4
PZ0O0Z0Z
8
rZblkZ0s
3
M0O0ZNZP
7
ZpZ0mpap
2
0A0Z0OPZ
6
pZnZpZpZ
1
S0ZQSBJ0
a b c d e f g h
5
Z0Z0Z0Z0 17. Nc2 Rfd8 18. Qb1 Rab8 19. hg3 b5 20. b4 Ra8
4
PZ0O0Z0Z
3
MPZ0ZNZ0 8
rZ0s0ZkZ
2
0A0ZBOPO 7
Z0l0mpap
1
S0ZQJ0ZR 6
pZnZbo0Z
a b c d e f g h 5
ZpZpZ0Z0
4
0O0O0Z0Z
Here Black has a way of opening the light-squared bishop
3
Z0O0ZNO0
while safeguarding the e5 square, by playing: 2
0ANZ0OPZ
11. . . d6 12. O-O O-O 13. c3 d5 14. Re1 Qc7 15. Bf1 Be6
16. h3
1
SQZ0SBJ0
a b c d e f g h

in Pawn-sideways chess is:
8
rZ0Z0skZ 1. c4 (book) c5 2. Nc3 Nc6 3. g3 g6 4. Bg2 e6 5. e4 a6 6. a3
7
Zpl0mpap Rb8 7. Nge2 Bg7 8. Rb1 dd6 9. bb4
6
pZnZbZpZ
5
Z0ZpZ0Z0
8
0sblkZns
4
PZ0O0Z0Z
7
ZpZ0Zpap
3
M0O0ZNZP
6
pZnopZpZ
2
0A0Z0OPZ
5
Z0o0Z0Z0
1
S0ZQSBJ0
4
0OPZPZ0Z
a b c d e f g h
3
O0M0Z0O0
2
0Z0ONOBO
In this position, Black utilizes a rather unique defensive
1
ZRAQJ0ZR
resource: a b c d e f g h
16. . . gf6 Here, Black responds with a typical lateral move.
72
9. . . c7 Here another lateral move proves useful:

17. b4 Qc4 18. Bf4 e5 19. Bg5 gf6 20. Be3 e6
8
0sblkZns
7
Z0o0Zpap
6
pZnopZpZ
8
0sbs0ZkZ
5
Z0o0Z0Z0
7
Z0o0Zpap
4
0OPZPZ0Z
6
pZnZpZ0Z
3
O0M0Z0O0
5
Z0ZPo0Z0
2
0Z0ONOBO
4
0OqZPZ0O
1
ZRAQJ0ZR
3
O0M0A0O0
a b c d e f g h 2
0Z0ZNOBZ
10. O-O Nge7 11. bxc5 Rxb1 12. cxd6
1
Z0ZQZRJ0
a b c d e f g h
8
0ZblkZ0s
7
Z0o0mpap Black moves the g6 pawn first to f6 and then to e6, reaching
this position. The continuation shown here is not forced, and
6
pZnOpZpZ in some of its games, AlphaZero opts for slightly different
lines with Black, as this seems to be a very rich opening.
5
Z0Z0Z0Z0
4
0ZPZPZ0Z B.8.4. I NSTRUCTIVE GAMES
O0M0Z0O0
3 Game AZ-27: AlphaZero Pawn-sideways vs AlphaZero
Pawn-sideways The game is played from a fixed opening
2
0Z0ONOBO position that arises after: 1. e4 e5 2. Nf3 Nc6 3. Bc4. The
remaining moves follow best play, at roughly one minute
1
ZrAQZRJ0 per move.
a b c d e f g h 1. e4 (book) e5 (book) 2. Nf3 (book) Nc6 (book) 3. Bc4
(book) d6 4. O-O Be6 5. Bb3 g5 6. dd4
White fights for the advantage by going for this kind of a
material imbalance, an exchange down.
rZ0lkans
12. . . Rb8 13. dxe7 Qxe7 14. dd4 O-O 15. h4 Rd8 16. d5 8
Qc5
7
opo0ZpZp
8
0sbs0ZkZ 6
0ZnobZ0Z
7
Z0o0Zpap 5
Z0Z0o0o0
6
pZnZpZpZ 4
0Z0OPZ0Z
5
Z0lPZ0Z0 3
ZBZ0ZNZ0
4
0ZPZPZ0O 2
POPZ0OPO
3
O0M0Z0O0 1
SNAQZRJ0
2
0Z0ZNOBZ a b c d e f g h
1
Z0AQZRJ0
a b c d e f g h 6. . . Bxb3 7. axb3 g4 8. Nxe5
73
rZ0lkans
8 White uses a lateral move (e4-d4) to create threats on the
e-file.
7
opo0ZpZp 13. . . e4 14. cc4 Bd6 15. g3 Kf8 16. Qg2 Ng4
6
0Zno0Z0Z
5
Z0Z0M0Z0
4
0Z0OPZpZ
3
ZPZ0Z0Z0
8
rZ0l0j0s
2
0OPZ0OPO
7
opo0Zpo0
1
SNAQZRJ0
6
0Z0a0ZnZ
a b c d e f g h 5
Z0ZPZ0Z0
Already, things are getting very tactical and very unortho-
4
0ZPOpZnZ
dox. 3
ZPZ0Z0O0
0O0Z0OQO
8. . . dxe5 9. d5 2
8
rZ0lkans 1
SNA0S0J0
a b c d e f g h
7
opo0ZpZp
6
0ZnZ0Z0Z
5
Z0ZPo0Z0
0Z0ZPZpZ
4 Black goes for the attack.
17. Rxe4 Nxh2 18. Nd2 f5 19. Re1 Bf4
3
ZPZ0Z0Z0
2
0OPZ0OPO
1
SNAQZRJ0
rZ0l0j0s
a b c d e f g h 8
Black leaves the knight on c6 and goes on with creating
counter-threats.
7
opo0Z0o0
9. . . hg7 10. Qxg4 Nf6 11. Qf3 Ne7 12. Re1 Ng6 13. d4
6
0Z0Z0ZnZ
5
Z0ZPZpZ0
8
rZ0lka0s 4
0ZPO0a0Z
7
opo0Zpo0 3
ZPZ0Z0O0
6
0Z0Z0mnZ 2
0O0M0OQm
5
Z0ZPo0Z0 1
S0A0S0J0
0Z0O0Z0Z
4 a b c d e f g h
3
ZPZ0ZQZ0
2
0OPZ0OPO
1
SNA0S0J0 Offering a piece on f4.
a b c d e f g h 20. gxf4 Nxf4 21. Qg3 gg5
74
rZ0l0j0s
8 30. c3 a6 31. Re1+ Kd7 32. Kg2 Ne4 33. e5 Ke6 34. d3
Rg8+ 35. Kf3 Ng5+ 36. Kg3 c6 37. Re3 Rd8 38. ed5+ Kf6
7
opo0Z0Z0 39. h4
6
0Z0Z0Z0Z 8
0Z0s0Z0Z
5
Z0ZPZpo0 7
ZpZ0Z0Z0
4
0ZPO0m0Z 6
pZpZ0j0Z
3
ZPZ0Z0L0 5
Z0ZPZpm0
2
0O0M0O0m 4
0ZPO0Z0O
1
S0A0S0J0 3
Z0ZPS0J0
a b c d e f g h
White uses a lateral pawn move to safeguard the king.

2
0O0Z0Z0Z
22. g2 Qd6 23. Nf1 Nxf1 24. Qxg5 Nh3+
1
Z0Z0Z0Z0
a b c d e f g h
8
rZ0Z0j0s Here Black decides to take on d5 rather than try to move
the knight, and White recaptures on d5 as well rather than
7
opo0Z0Z0 taking on g5!
6
0Z0l0Z0Z 39. . . cxd5 40. cxd5 Rxd5 41. e4 fxe4 42. dxe4 Nxe4
43. Rxe4 a5
5
Z0ZPZpL0
4
0ZPO0Z0Z 8
0Z0Z0Z0Z
3
ZPZ0Z0Zn 7
ZpZ0Z0Z0
2
0O0Z0ZPZ 6
0Z0Z0j0Z
1
S0A0SnJ0 5
o0ZrZ0Z0
a b c d e f g h 4
0Z0ZRZ0O
25. gxh3 Rg8 26. Re5 Rxg5+ 27. Bxg5 Ng3 28. Be7+ Qxe7
29. Rxe7 Kxe7
3
Z0Z0Z0J0
2
0O0Z0Z0Z
8
rZ0Z0Z0Z 1
Z0Z0Z0Z0
7
opo0j0Z0 a b c d e f g h
6
0Z0Z0Z0Z And now the game moves towards a draw.
5
Z0ZPZpZ0 44. Ra4 Rd2 45. a2 ab5 46. Rf4+ Ke5 47. Rb4 c5 48. Rxb7
Rxa2 49. h5 Ra6 50. Rb5 d5 51. Kg4 Ke4 52. Kg5 d4
4
0ZPO0Z0Z with a draw to follow soon. 1/2–1/2
3
ZPZ0Z0mP Game AZ-28: AlphaZero Pawn-sideways vs AlphaZero
2
0O0Z0Z0Z Pawn-sideways The game is played from a fixed opening
1
S0Z0Z0J0 position that arises after 1. c4 c5. The remaining moves
a b c d e f g h
1. c4 (book) c5 (book) 2. Nc3 Nc6 3. g3 g6 4. Bg2 e6 5. e4
Finally the dust has settled: White having two pawns for the a6 6. a3 dd6 7. Nge2 Bg7 8. Rb1 Nge7 9. O-O Rb8 10. bb4
piece. c7 11. bxc5 Rxb1 12. cxd6 Rb8 13. dxe7 Qd7 14. c5 b6
75
0sbZkZ0s
8 Not minding to give up the piece, for getting strong passed
pawns in return.
7
Z0oqOpap 26. . . Bxe2 27. Re1
6
0onZpZpZ
5
Z0O0Z0Z0
4
0Z0ZPZ0Z
8
0s0Z0skZ
3
O0M0Z0O0
7
Z0Z0lpap
2
0Z0ONOBO
6
0O0ZpZpZ
1
Z0AQZRJ0
5
Z0ZPZ0Z0
a b c d e f g h 4
0Z0O0Z0Z
To Black’s a6-b6, White responds with c5-b5, another lateral
3
ZQZ0A0O0
move. 2
0Z0ZbOBO
15. b5 Nd4 16. Nxd4 Bxd4 17. Ne2 Bg7 18. a4 Qxe7 19. dd4
O-O 20. Qb3 a6 21. Be3 axb5 22. axb5 Ba6
1
Z0Z0S0J0
a b c d e f g h
8
0s0Z0skZ
7
Z0o0lpap And yet, Black agrees and decides to return the piece in-
stead.
6
bZ0ZpZpZ 27. . . exd5 28. Rxe2 Bxd4 29. Bxd4 Qxe2 30. Bxd5
5
ZPZ0Z0Z0 White opts to have the bishop pair and a pawn for two
4
0Z0OPZ0Z exchanges, an unbalanced position.
3
ZQZ0A0O0 30. . . gf6 31. h4 hg7 32. b7 Qa6 33. Bg2 Rfe8 34. Bc5 gg6
35. Qf3 Kg7 36. a7 Rbd8 37. Be3 Rh8
2
0Z0ZNOBO
1
Z0Z0ZRJ0
a b c d e f g h 8
0Z0s0Z0s
White uses a lateral move to protect the pawn 7
O0Z0Zpj0
23. c4 c6 24. b6 c5 25. ed4 d5 26. cxd5 6
qZ0Z0opZ
8
0s0Z0skZ
5
Z0Z0Z0Z0
7
Z0Z0lpap
4
0Z0Z0Z0O
6
bO0ZpZpZ
3
Z0Z0AQO0
5
Z0ZPZ0Z0
2
0Z0Z0OBZ
4
0Z0O0Z0Z
1
Z0Z0Z0J0
a b c d e f g h
3
ZQZ0A0O0
2
0Z0ZNOBO 38. Qf4 Rd7 39. Qb8 Rdd8 40. Qc7 Qa1+ 41. Kh2 g5
1
Z0Z0ZRJ0 42. Qc4 Qe5 43. Kh3 Rc8 44. Qg4 Qe6 45. Bb7 Qxg4+
a b c d e f g h 46. Kxg4 Rc4+ 47. Kf3 gxh4 48. a8=R Rxa8 49. Bxa8 g4+
76
8
BZ0Z0Z0Z 8
rZ0Zkans
7
Z0Z0Zpj0 7
opoblpZp
6
0Z0Z0o0Z 6
0Zno0Z0Z
5
Z0Z0Z0Z0 5
ZNZPo0Z0
4
0ZrZ0ZpZ 4
0ZBZPZ0Z
3
Z0Z0AKO0 3
Z0Z0ZQZ0
2
0Z0Z0O0Z 2
POPZ0OPO
1
Z0Z0Z0Z0 1
S0A0S0J0
50. Ke2 e6 51. ff3 gxf3+ 52. Bxf3 f5 53. Kd3 Ra4 54. Bd1
Ra3+ 55. Ke2 ee5 56. f3 Kf6 57. Bc1 Ra2+ 58. Bd2 g5
59. Bb3 Ra3 60. Bd5 ef5 61. Be3 f4 62. Bd4+ Kf5 63. Be4+ 10. . . Nd4 11. Nxd4 exd4 12. Bg5
Ke6 64. e3 fg4 65. Kf2 f5 66. Bb7 e5 67. Bc8+ Kd5 68. Bxe5
Kxe5 69. Bxg4
and the game soon ended in a draw. 1/2–1/2
8
rZ0Zkans
Game AZ-29: AlphaZero Pawn-sideways vs AlphaZero
Pawn-sideways Position from an AlphaZero game played
7
opoblpZp
at roughly one minute per move, from a predefined position.
6
0Z0o0Z0Z
5
Z0ZPZ0A0
4
0ZBoPZ0Z
8
rZ0Zkans 3
Z0Z0ZQZ0
7
opo0lpZp 2
POPZ0OPO
6
0ZnobZ0Z 1
S0Z0S0J0
5
Z0ZPo0Z0 a b c d e f g h
4
0ZBZPZpZ
3
M0Z0ZNZ0
2
POPZ0OPO with a motif of a lateral (e4-f4) discovery! In the game,
1
S0AQS0J0 Black didn’t take the bishop. So, how would have the game
proceeded if Black took the bishop? Here is one possible
a b c d e f g h
continuation from AlphaZero: 12. . . Qxg5 13. f4+ Qe7
14. Rxe7+ Nxe7 15. c5 dxc5 16. Qxb7 Rc8 17. Re1 Kd8
18. Qxa7 Nc6 19. Qa4 hg7 20. c3 Rh6 21. Bb5 Rb8 22. g3
8. . . gxf3 9. Qxf3 Bd7 10. Nb5 Rd6 23. d3 f6 24. h4 e6 25. h5 f7 26. Rb1 Rb6 27. Kg2. The
continuation is assessed as better for White.
Instead of capturing the knight, White has something else
in mind. . . 12. . . f6 13. f4 de6
77
8
rZ0Zkans 8
0j0srZ0Z
7
opobl0Zp 7
Zpo0m0Z0
6
0Z0Zpo0Z 6
0Z0aPo0Z
5
Z0ZPZ0A0 5
Z0Z0Z0o0
4
0ZBo0O0Z 4
0Z0Z0O0Z
3
Z0Z0ZQZ0 3
ZPO0Z0OP
2
POPZ0OPO 2
0Z0ZRO0A
1
S0Z0S0J0 1
S0Z0Z0J0
29. fxg5 fxg5 30. h4 gxh4 31. gxh4 Rh8 32. Bxd6 cxd6
33. Ra4 Rc8 34. Rg4 Nf5 35. e7 Kc7 36. Rf4 Nh6 37. g4
Kd7 38. f3 Rhe8 39. Rh2 Rxc3 40. Rxh6 Rxe7 41. Kh2 d5
Black uses lateral moves to cover the file as well. 42. b4 e5 43. Rf8 Rb3 44. g5 e4 45. fxe4 Rxb4 46. f4 Rb3
47. Kg1 Re2 48. Rh7+ Kd6 49. Rd8+ Kc5 50. Rd1 Rg3+
14. dxe6 Bc6 15. Bd5 O-O-O 16. Bxc6 bxc6 17. Qxc6 51. Kh1 Re4 52. Rf1 Rg4 53. Rxb7 Rexf4 54. Rxf4 Rxf4
55. Kh2 Rg4 56. Rg7 Kd6 57. Kh3 Rg1 58. Kh4
8
0Z0Z0Z0Z
8
0Zks0ans
7
Z0Z0Z0S0
7
o0o0l0Zp
6
0Z0j0Z0Z
6
0ZQZPo0Z
5
Z0Z0Z0O0
5
Z0Z0Z0A0
4
0Z0Z0Z0J
4
0Z0o0O0Z
3
Z0Z0Z0Z0
3
Z0Z0Z0Z0
2
0Z0Z0Z0Z
2
POPZ0OPO
1
Z0Z0Z0s0
a b c d e f g h
1
S0Z0S0J0
a b c d e f g h
and White soon won the game. 1–0
Game AZ-30: AlphaZero Pawn-sideways vs AlphaZero

Pawn-sideways The first ten moves for White and Black
White has gained several pawns for the piece, has a danger- have been sampled randomly from AlphaZero’s opening
ous attack and a substantial advantage, according to Alp- “book”, with the probability proportional to the time spent
haZero. Yet, Black uses a lateral pawn move here to prevent calculating each move. The remaining moves follow best
immediate disaster: play, at roughly one minute per move.
17. . . ab7 18. Qa4 Kb8 19. Bh4 Qb4 20. Qb3 g7 21. Bg3 1. c4 c5 2. Nc3 Nc6 3. g3 g6 4. Bg2 e6 5. e4 dd6 6. Rb1 a6
Bd6 22. c3 Qxb3 23. axb3 dxc3 24. bxc3 Nh6 25. h3 Nf5 7. a3 Bg7 8. Nge2 Rb8 9. bb4 c7 10. O-O Nge7 11. bxc5
26. Bh2 Rhe8 27. Re2 Ne7 28. gg3 g5 Rxb1 12. cxd6 Rb8 13. dxe7 Qd7
78
8
0sbZkZ0s 8
bs0Z0skZ
7
Z0oqOpap 7
Z0oqmpap
6
pZnZpZpZ 6
0ZpZpZpZ
5
Z0Z0Z0Z0 5
Z0Z0Z0Z0
4
0ZPZPZ0Z 4
0ZPOPZ0O
3
O0M0Z0O0 3
ZPM0A0O0
2
0Z0ONOBO 2
0ZQZNOBZ
1
Z0AQZRJ0 1
Z0Z0S0J0
In this game (unlike in the main lines section before), Black 20. Bh3 Qc8 21. Rd1 Qa6 22. Bg2 Rfd8 23. Rb1 cd6 24. d5
decides to recapture on e7 with the knight instead. exd5 25. Nxd5 Nxd5 26. exd5
14. Re1 Bb7 15. b3 Nxe7 16. dd4 O-O 17. Be3
8
bs0s0ZkZ
8
0s0Z0skZ 7
Z0o0Zpap
7
Zboqmpap 6
qZ0o0ZpZ
6
pZ0ZpZpZ 5
Z0ZPZ0Z0
5
Z0Z0Z0Z0 4
0ZPZ0Z0O
4
0ZPOPZ0Z 3
ZPZ0A0O0
3
ZPM0A0O0 2
0ZQZNOBZ
2
0Z0ZNOBO 1
ZRZ0Z0J0
a b c d e f g h
1
Z0ZQS0J0
a b c d e f g h
The d5 pawn is locking out the a8 bishop, so Black chal-

lenges the center with a lateral move, only to decide to push
Here Black plays a lateral move (a6-b6) to improve its pawn forward on the next move. This perhaps reveals a fluidity of
structure: plans as well as structures.
17. . . b6 18. h4 Ba8 19. Qc2 26. . . e6 27. Nf4 e5 28. Ne2 Bf8 29. Nc3 c6
But the pawn marches on, although not forward, opening The center is challenged again, this time from the other side,
the line for the rook with: but White has a lateral response to keep things locked:
19. . . bc6 30. dc5
79
8
bs0s0akZ 8
bs0s0akZ
7
Z0Z0ZpZp 7
Z0l0Zpo0
6
qZpZ0ZpZ 6
0Z0o0o0Z
5
Z0O0o0Z0 5
Z0ZPoPZ0
4
0ZPZ0Z0O 4
NZPZ0Z0Z
3
ZPM0A0O0 3
ZPZ0A0O0
2
0ZQZ0OBZ 2
0SQZ0OBJ
1
ZRZ0Z0J0 1
Z0Z0Z0Z0
Black and White keep reconfiguring the central pawns.

45. . . c6 46. dc5 d6 47. cxd6 Rxd6 48. c5 Rdd8 49. b4 Bxg2
50. Kxg2 Qc6+ 51. f3 d5 52. Bd4 e5 53. Bf2 d5 54. Bd4
e5 55. Bf2 d5 56. Qb3 Qb5 57. Nc3 Qc4 58. bb5 Rdc8
And Black responds with a lateral move as well, bringing 59. Nxd5 Qxb3 60. Rxb3 Bxc5 61. b6 Bd6
the h-pawn towards the center.
30. . . hg7 31. Na4 Qa5 32. hg4 gf6 33. g5 fg6 34. f5 gf6
8
0srZ0ZkZ
7
Z0Z0Zpo0
6
0O0a0o0Z
5
Z0ZNZPZ0
4
0Z0Z0Z0Z
8
bs0s0akZ 3
ZRZ0ZPO0
7
Z0Z0Zpo0 2
0Z0Z0AKZ
6
0ZpZ0o0Z 1
Z0Z0Z0Z0
5
l0O0oPZ0 a b c d e f g h
4
NZPZ0Z0Z An interesting endgame arises.
3
ZPZ0A0O0 62. Nc3 Bc5 63. Nd5 Bd6 64. Rb2 e6 65. e5
2
0ZQZ0OBZ
1
ZRZ0Z0J0
8
0srZ0ZkZ
a b c d e f g h
7
Z0Z0Zpo0
6
0O0apZ0Z
5
Z0ZNO0Z0
4
0Z0Z0Z0Z
After a sequence of lateral moves, the situation has settled 3
Z0Z0ZPO0
0S0Z0AKZ
on the kingside.
2
35. Rb2 d5 36. Bf4 e5 37. Bd2 Qc7 38. Be3 Qa5 39. Ra2
Qb4 40. Rb2 Qa5 41. Kh2 d5 42. Bf4 e5 43. Bd2 Qc7 1
Z0Z0Z0Z0
44. Be3 d6 45. d5 a b c d e f g h
80
Both sides using lateral move to create threats. 72. . . Rb8 73. a7 Ra8 74. Nd3 exd5 75. Nb4 e5 76. b7 Rb8
77. a7 Ra8 78. Na6 Bd6 79. Bc5
65. . . Bf8 66. Nf4 Bc5 67. b7 Rc7 68. Rc2 Bb6 69. Rxc7
Bxc7
8
rZ0Z0ZkZ
8
0s0Z0ZkZ 7
O0Z0ZpZ0
7
ZPa0Zpo0 6
NZ0a0Z0Z
6
0Z0ZpZ0Z 5
Z0A0o0o0
5
Z0Z0O0Z0 4
0Z0Z0Z0Z
4
0Z0Z0M0Z 3
Z0Z0ZPO0
3
Z0Z0ZPO0 2
0Z0Z0ZKZ
2
0Z0Z0AKZ 1
Z0Z0Z0Z0
1
Z0Z0Z0Z0 a b c d e f g h
a b c d e f g h
79. . . Bxc5 80. b7
But the pawn can switch files!
70. a7
8
rZ0Z0ZkZ
7
ZPZ0ZpZ0
8
0s0Z0ZkZ 6
NZ0Z0Z0Z
7
O0a0Zpo0 5
Z0a0o0o0
6
0Z0ZpZ0Z 4
0Z0Z0Z0Z
5
Z0Z0O0Z0 3
Z0Z0ZPO0
4
0Z0Z0M0Z 2
0Z0Z0ZKZ
3
Z0Z0ZPO0 1
Z0Z0Z0Z0
2
0Z0Z0AKZ a b c d e f g h
1
Z0Z0Z0Z0 80. . . Rd8 81. Nxc5 f6 82. Ne6 Rb8 83. c7 Ra8 84. Nd8
a b c d e f g h Rc8 85. Ne6 Kf7 86. d7
0ZrZ0Z0Z
70. . . Ra8 71. d5 g5 72. b7
8
8
rZ0Z0ZkZ 7
Z0ZPZkZ0
7
ZPa0ZpZ0 6
0Z0ZNo0Z
6
0Z0ZpZ0Z 5
Z0Z0o0o0
5
Z0ZPZ0o0 4
0Z0Z0Z0Z
4
0Z0Z0M0Z 3
Z0Z0ZPO0
3
Z0Z0ZPO0 2
0Z0Z0ZKZ
2
0Z0Z0AKZ 1
Z0Z0Z0Z0
Z0Z0Z0Z0
1 a b c d e f g h
a b c d e f g h 86. . . Rb8 87. d8=Q Rxd8 88. Nxd8+ Ke7 89. Nc6+ Kd6
81
90. Nd8 Ke7 91. Nb7 ff5 92. e3 g4 93. Kf2 e4 94. Ke2 Ke6 Here we see a new kind of tactic, made possible by a lateral
95. Nd8+ Ke7 96. Nc6+ Kf6 97. Nd4 pawn move!
17. . . Nxf2 18. Kxf2 e7
8
0Z0Z0Z0Z
7
Z0Z0Z0Z0 8
rZbl0skZ
6
0Z0Z0j0Z 7
ZpZ0o0ap
5
Z0Z0ZpZ0 6
0Z0MpmpZ
4
0Z0MpZpZ 5
ZNZpA0Z0
3
Z0Z0O0O0 4
0ZPZ0Z0Z
2
0Z0ZKZ0Z 3
Z0Z0Z0ZP
1
Z0Z0Z0Z0 2
PO0Z0JPZ
S0ZQSBZ0
a b c d e f g h 1
However, this position is a draw! a b c d e f g h
97. . . g5 98. Kf1 Ke5 99. Ne2 f5 100. Kg1 Kd5 101. Kf2 19. Kg1 exd6 20. Nxd6 Nh5 21. Bxg7 Nxg7 22. Nxc8 Qxc8
Ke5 102. Kf1 Kd5 103. Kf2 Ke5 104. Kf1 Kd5 105. Kg2 23. cxd5 Qc5+ 24. Kh1 exd5 25. Qb3 b6
Kc4 106. Nd4 e5 107. Nc6 Kd5 108. Ne7 Ke6 109. Nc8 f5
110. Na7 Ke5 111. Nc6+ Kd5 112. Nd4 e5 113. Nf5 Ke6
114. Ng7+ Kf7 115. Nf5 Ke6 116. Nh6 f5 117. Kf1 Kf6 8
rZ0Z0skZ
Z0Z0Z0mp
118. Ke1 Kg6 119. Ng8 Kf7 120. Nh6+ Kg6 121. Nxg4 7
fxg4 122. Kd2 Kf5 123. Kc3 ef4 124. exf4 h4 125. e4+
Kxe4 126. gxh4 Kf4 127. Kc2 Kg4 128. Kc1 Kxh4 1/2–1/2 6
0o0Z0ZpZ
Game AZ-31: AlphaZero Pawn-sideways vs AlphaZero 5
Z0lpZ0Z0
0Z0Z0Z0Z
Pawn-sideways The first ten moves for White and Black 4
3
ZQZ0Z0ZP
play, at roughly one minute per move. 2
PO0Z0ZPZ
1. c4 c5 2. e3 e6 3. dd4 cxd4 4. exd4 g6 5. Nc3 Bg7 6. Nb5
bc7 7. Bf4 Na6 8. Nf3 Nf6 9. h3 d5 10. Bd3 O-O 11. O-O
1
S0Z0SBZK
a b c d e f g h
ab7 12. Re1 c6 13. Nd6 cc5 14. Be5 cxd4 15. Nxd4 Nc5
16. Bf1 Nce4 17. N4b5 The dust has settled, and the game soon ended in a draw.
26. g4 Qd6 27. Bg2 Rad8 28. Rac1 Ne6 29. Qxd5 Qxd5
8
rZbl0skZ 30. Bxd5 Rxd5 31. Rxe6 Rf2 32. Rxb6 Rdd2 33. g5 hg7
7
ZpZ0Zpap 34. a4 Rh2+ 35. Kg1 Rdg2+ 36. Kf1 Rf2+ 37. Kg1 Rfg2+
38. Kf1 Rf2+ 39. Ke1 Rfg2 40. Rb8+ Kh7 41. Kf1 Rf2+
6
0Z0MpmpZ 42. Kg1 Rfg2+ 43. Kf1 Rf2+ 44. Kg1 Rfg2+ 45. Kf1 1/2–
1/2
5
ZNZpA0Z0
4
0ZPZnZ0Z Game AZ-32: AlphaZero Pawn-sideways vs AlphaZero
Pawn-sideways The first ten moves for White and Black
3
Z0Z0Z0ZP have been sampled randomly from AlphaZero’s opening
PO0Z0OPZ
2
1
S0ZQSBJ0 play, at roughly one minute per move.
a b c d e f g h 1. c4 c5 2. Nc3 g6 3. e3 e6 4. dd4 bc7 5. dxc5 Bxc5 6. g4
82
rmblkZns
8 ing, now finds its place on f4, where it shuts out the activity
on the b8-h2 diagonal.
7
o0opZpZp 12. . . Nf6 13. a3 b7 14. Rb1 f7 15. b4 O-O 16. Bb2 Rd8
6
0Z0ZpZpZ 17. Rc1 Bf8 18. Qb3 Bd7 19. dc4 dxc4 20. Qxc4 Be8 21. g3
Bg7 22. Qb3 Qb6 23. Nd4 Nbd7 24. aa4 Bf8 25. Ba3 ee5
5
Z0a0Z0Z0 26. fxe5 Nxe5 27. Rfd1 Neg4 28. gf3 f4
4
0ZPZ0ZPZ
3
Z0M0O0Z0
2
PO0Z0O0O 8
rZ0sbakZ
1
S0AQJBMR 7
ZpZ0ZpZ0
0lpZ0mpZ
a b c d e f g h 6
Now that is an unusual sight, the early advance of the g-
pawn.
5
Z0Z0Z0Z0
6. . . hg7 7. Bg2 c6 8. Nf3 d5 9. O-O Qc7 10. d4
4
PO0M0onZ
3
AQM0OPZ0
8
rmbZkZns 2
0Z0Z0OBZ
7
o0l0Zpo0 1
Z0SRZ0J0
6
0ZpZpZpZ a b c d e f g h
5
Z0apZ0Z0
4
0Z0O0ZPZ The game gets quite tactical here.
3
Z0M0ONZ0 29. a5 Qc7 30. exf4 Qxf4 31. fxg4 Rxd4 32. Rxd4 Qxd4
2
PO0Z0OBO 33. g5 Ng4 34. Ne4 Qe5 35. Bb2 Qh2+ 36. Kf1 Bd7 37. f3
S0AQZRJ0
Qf4 38. Re1 Re8 39. Qc4 Nh2+
1
a b c d e f g h
White plays c4-d4, a lateral move, to reinforce the center.

10. . . Bd6 11. h3 f5 12. f4 8
0Z0ZrakZ
7
ZpZbZpZ0
8
rmbZkZns 6
0ZpZ0ZpZ
7
o0l0Z0o0 5
O0Z0Z0O0
6
0ZpapZpZ 4
0OQZNl0Z
5
Z0ZpZpZ0 3
Z0Z0ZPZ0
4
0Z0O0O0Z 2
0A0Z0ZBm
3
Z0M0ONZP 1
Z0Z0SKZ0
2
PO0Z0OBZ a b c d e f g h
1
S0AQZRJ0
a b c d e f g h
The g-pawn, advanced earlier in what seemed to be weaken- 40. Kg1 Nxf3+ 41. Bxf3 Qxf3 42. Nf6+
83
8
0Z0ZrakZ 8
0L0Z0akZ
7
ZpZbZpZ0 7
ZpZro0Z0
6
0ZpZ0MpZ 6
0ZpZ0ZpZ
5
O0Z0Z0O0 5
O0Z0ZbO0
4
0OQZ0Z0Z 4
PZ0Z0Z0Z
3
Z0Z0ZqZ0 3
Z0Z0Z0Z0
2
0A0Z0Z0Z 2
0A0Z0Z0Z
1
Z0Z0S0J0 1
Z0J0Z0Z0
Black needs to give away its queen to stop the attack. This resource is what Black was keeping in reserve, as
a potential way of responding to the threats on the a3-f8
42. . . Qxf6 43. Bxf6 Rxe1+ 44. Kf2 Rd1 45. Qf4 Bf5
diagonal while the f8 bishop was pinned.
46. Qb8 Rd7
57. Qh2 Bg7 58. b4 Bxb2+ 59. Kxb2 f7
8
0L0Z0akZ
7
ZpZrZpZ0
8
0Z0Z0ZkZ
6
0ZpZ0ApZ
7
ZpZrZpZ0
5
O0Z0ZbO0
6
0ZpZ0ZpZ
4
0O0Z0Z0Z
5
O0Z0ZbO0
3
Z0Z0Z0Z0
4
0O0Z0Z0Z
2
0Z0Z0J0Z
3
Z0Z0Z0Z0
1
Z0Z0Z0Z0
2
0J0Z0Z0L
a b c d e f g h 1
Z0Z0Z0Z0
a b c d e f g h
Is this a fortress? As we will see, the question is slightly

more complicated by the fact that the pawn structure isn’t The pawn has served its purpose on e7 and moves back.
fixed, and things will eventually open up.
60. c4 Be6 61. Qb8+ Kh7 62. Kb3 Kg7 63. Kc3 f6 64. gxf6+
47. Be5 Rd2+ 48. Ke1 Rd7 49. Bc3 Bd3 50. a4 Bc2 51. Qc8 Kxf6 65. Qf8+ Bf7 66. Kb4 g5 67. Qh6+ Bg6 68. Qh8+
Re7 52. Kd2 Bf5 53. Qb8 Rd7+ 54. Kc1 Rd3 55. Bf6 Rd7 Kf7 69. Qh3 Re7 70. Kc5 f5 71. Kd6 Re6+ 72. Kd7 Re7+
56. Bb2 e7 73. Kd8 Re8+ 74. Kc7 Re7+ 75. Kb6 e5
84
8
0Z0Z0Z0Z 8
0Z0Z0Z0Z
7
ZpZ0skZ0 7
Z0Z0Z0Z0
6
0JpZ0ZbZ 6
0Z0J0Z0Z
5
O0Z0o0Z0 5
Z0o0Zrj0
4
0ZPZ0Z0Z 4
0Z0Z0ZpZ
3
Z0Z0Z0ZQ 3
Z0Z0Z0Z0
2
0Z0Z0Z0Z 2
0Z0Z0Z0Z
1
Z0Z0Z0Z0 1
Z0Z0L0Z0
140. . . c4 141. Qe7+ Kf4 142. Qe2 g3 143. Ke6 Kg5

144. Qxc4 Rf6+ 145. Ke5 Rf5+ 146. Kd6 Rf6+ 147. Ke5
76. Qf3+ Kg7 77. a6 bxa6 78. Kxc6 e4 79. Qe3 Rf7 80. c5 Rf5+ 148. Ke6 Rf6+ 149. Kd7 Rf4 150. Qe2 Kh4 151. Kd6
f4 81. Qf3 Bf5 82. Kd6 Rf6+ 83. Ke5 Bd7 84. Qg2+ Kf7 Kh3 152. Ke5 Rf2 153. Qh5+ Kg2 154. Ke4 Kg1 155. Ke3
85. Qh1 Kg7 86. Qb7 Rf7 87. Qg2+ Kh7 88. Qf3 Kg7 g2 156. Qh4 Rf8 157. Ke2 Rf1 158. Qg3 Kh1 159. Qh3+
89. Kd6 Kh6 90. Qb3 Kg7 91. Qc3+ Kh7 92. Qd3+ Kg7 Kg1 160. Qh4 h2 161. Qd4+ Kh1 162. Qh4 Kg1 163. Qg5+
93. Qd4+ Kh7 94. Qd5 Kg7 95. Qg2+ Kh7 96. Qh2+ Kg8 g2 164. Qh6 Rf2+ 165. Ke3 Rf1 166. Ke2 Rf2+ 167. Ke3
97. Qg1+ Kh7 98. Qb1+ Kg7 99. Qb2+ Kg8 100. Qb8+ Kg7 Rf1 168. Qh3
101. Qb2+ Kg8 102. Qg2+ Kh7 103. Qc2+ Kg8 104. Qg6+
0Z0Z0Z0Z
Kf8 105. Qh5 Kg7 106. Qe5+ Kg8 107. Qg5+ Kh7 108. Qd5
8
Kg7 109. Qe5+ Kg8 110. Qg5+ Kh7 111. Qh5+ Kg7
112. Qf3 Kh6 113. c6 7
Z0Z0Z0Z0
6
0Z0Z0Z0Z
8
0Z0Z0Z0Z
5
Z0Z0Z0Z0
7
Z0ZbZrZ0
4
0Z0Z0Z0Z
6
pZPJ0Z0j
3
Z0Z0J0ZQ
5
Z0Z0Z0Z0
2
0Z0Z0ZpZ
4
0Z0Z0o0Z
1
Z0Z0Zrj0
a b c d e f g h
3
Z0Z0ZQZ0 And the game ended in a draw in a couple of moves.
2
0Z0Z0Z0Z 1/2–1/2
1
Z0Z0Z0Z0 B.9. Self-capture
a b c d e f g h
In Self-capture chess, we have considered extending the
rules of chess to allow players to capture their own pieces.
113. . . Bxc6 114. Kxc6 Kg5 115. Kd6 Rf5 116. Ke6 Rf6+
B.9.1. M OTIVATION
117. Ke5 Rf5+ 118. Ke4 Rf7 119. Kd4 Rd7+ 120. Kc4 b6
121. Qg2+ g4 122. Qf1 Rd6 123. Qc1+ f4 124. Qg1+ g4 The ability to capture one’s own pieces could help break
125. Qe3+ Kf5 126. Qf2+ Kg5 127. Qe3+ Kf5 128. Qg3 “deadlocks” and offer additional ways of infiltrating the
Rf6 129. Qh4 c6 130. Kd4 d6 131. Kd5 c6+ 132. Kc5 Rg6 opponent’s position, as well as quickly open files for the
133. Qg3 Rf6 134. Qh4 Rg6 135. Qg3 Rf6 136. Kb6 Kg5 attack. Self-captures provide additional defensive resources
137. Kc7 Rf3 138. Qe5+ Rf5 139. Qe1 c5 140. Kd6 as well, given that the King that is under attack can consider
85
escaping by self-capturing its own adjacent pieces. additional options and winning motifs in the
endgames.
B.9.2. A SSESSMENT
The assessment of the Self-capture chess variant, as pro- Taking one’s own piece represents another way of
vided by Vladimir Kramnik: sacrificing in chess, and material sacrifices make
chess games more spectacular and enjoyable both
“
I like this variation a lot, I would even go as far
as to say that to me this is simply an improved for public and for the players. Most of the times
version of regular chess. this is used as an attacking idea, to gain initiative
and compromise the opponent’s king.
Self-captures make a minor influence on the open-
ing stage of a chess game, though we have seen
examples of lines that become possible under this For example, consider the Dragon Sicilian, as an
rule change that were not possible before. For ex- example of a sharp opening. After 1. e4 c5 2. Nf3
ample, consider the following line 1. e4 e5 2. Nf3 d6 3. d4 cxd4 4. Nxd4 Nf6 5. Nc3 g6 6. Be3 Bg7
Nc6 3. Bb5 a6 4. Ba4 Nf6 5. 0-0 Nxe4 6. d4 exd4 7. f3 0-0 8. Qd2 Nc6 9. 0-0-0 d5 something like
7. Re1 f5 8. Nxd4 Qh4 9. g3 in the Ruy Lopez. 10. g4 e5 11. Nxc6 bxc6 is possible, at which point
there is already Qxh2, a self-capture, opening the
file against the enemy king. Of course, Black can
8
rZbZka0s (and probably should) play differently.
7
ZpopZ0op
6
pZnZ0Z0Z
5
Z0Z0ZpZ0 8
rZbl0skZ
4
BZ0MnZ0l 7
o0Z0Zpap
3
Z0Z0Z0O0 6
0ZpZ0mpZ
2
POPZ0O0O 5
Z0Zpo0Z0
1
SNAQS0J0 4
0Z0ZPZPZ
a b c d e f g h
3
Z0M0APZ0
POPZ0Z0L
While not the main line, it is possible to play in 2
Self-capture chess and AlphaZero assesses it as
equal. In classical chess, however, this position
is much better for White. The key difference is
1
Z0JRZBZR
a b c d e f g h
that in self-capture chess Black can respond to
g3 by taking its own pawn on h7 with the queen,
gaining a tempo on the open file. In fact, White
can gain the usual opening advantage earlier in
the variation, by playing 8. Ng5 d5 9. f3 Bd6 The possibilities for self-captures in this example
10. fxe4 dxe4, which AlphaZero assesses as giving don’t end, as after 12. . . d4, White could even
the 60% expected score for White after about a consider a self-capture 13. Nxe4, sacrificing an-
minute’s thought, which is usually possible to de- other pawn. This is not the best continuation
fend with precise play. In fact, there are multiple though, and AlphaZero evaluates that as being
improvements for both sides in the original line, equal. It is just an illustration of the ideas which
but discussing these is beyond the scope of this become available, and which need to be taken
example. It is worth noting that AlphaZero prefers into account in tactical calculations.
to utilise the setup of the Berlin Defence, similar
to its style of play in classical chess. In terms of endgames, self-captures affect a wide
Regardless of its relatively minor effect on the spectrum of otherwise drawish endgame positions
openings, self-captures add aesthetically beau- winning for the stronger side. Consider the fol-
tiful motifs in the middlegames and provide lowing examples:
86
merely one of several highly promising and likely options.

8
0a0Z0Z0Z Here we give the first 20 moves in each of the main lines,
regardless of the position.
7
ZPZ0Z0Z0
6
0Z0Z0Z0Z Main line after e4 The main line of AlphaZero after 1. e4
in Self-capture chess is:
5
Z0Z0ZBZ0 1. e4 (book) e5 2. Nf3 Nc6 3. Bb5 Nf6 4. O-O Nxe4 5. Re1
4
0Z0j0Z0Z Nd6 6. Nxe5 Be7 7. Bf1 Nxe5 8. Rxe5 O-O 9. Nc3 Ne8
10. Nd5 Bd6 11. Re1 c6 12. Ne3 Be7 13. c4 Nc7 14. d4 d5
3
Z0Z0Z0Z0 15. cxd5 Bb4 16. Bd2 Bxd2 17. Qxd2 Nxd5 18. Nxd5 Qxd5
0Z0Z0ZKZ
2 19. Re5 Qd6 20. Bc4 Bd7
1
Z0Z0Z0Z0 8
rZ0Z0skZ
a b c d e f g h
7
opZbZpop
In this position, under Classical rules, the game
would be an easy draw for Black. In Self-capture
6
0Zpl0Z0Z
chess, however, this is a trivial win for White, who
can play Bc8 and then capture the bishop with the
5
Z0Z0S0Z0
b7 pawn, promoting to a queen! 4
0ZBO0Z0Z
3
Z0Z0Z0Z0
8
0Z0Z0Z0Z 2
PO0L0OPO
7
Z0Z0akZ0 1
S0Z0Z0J0
6
0Z0oRo0Z a b c d e f g h
5
o0oPZPo0 Main line after d4 The main line of AlphaZero after 1. d4
4
PoPZ0ZPo in Self-capture chess is:
3
ZPZ0ZKZP 1. d4 (book) d5 2. c4 e6 3. Nc3 Nf6 4. cxd5 exd5 5. Bg5 c6
6. Qc2 Nbd7 7. e3 Be7 8. Nf3 Nh5 9. Bxe7 Qxe7 10. Be2
2
0Z0Z0Z0Z O-O 11. O-O Ndf6 12. Ne5 g6 13. Qa4 Be6 14. b4 a6
15. Qb3 Ng7 16. Na4 Ne4 17. Qb2 Qg5 18. Nf3 Qe7 19. Ne5
1
Z0Z0Z0Z0 Qg5 20. Nf3 Qe7
a b c d e f g h
This endgame, which represents a fortress in clas-

8
rZ0Z0skZ
sical chess, becomes a trivial win in self-capture
chess, due to the possibilities for the White king
7
ZpZ0lpmp
to infiltrate the Black position either via e4 and a 6
pZpZbZpZ
Z0ZpZ0Z0
self-capture on d5 or via e2, d3 and a self-capture 5
on c4.
To conclude, I would highly recommend this vari- 4
NO0OnZ0Z
Z0Z0ONZ0
ation for chess lovers who value beauty in the 3
game on top of everything else.
B.9.3. M AIN L INES

Here we discuss “main lines” of AlphaZero under Self-
capture chess, when playing with roughly one minute per
” 2
1
PL0ZBOPO
S0Z0ZRJ0
a b c d e f g h
move from a particular fixed first move. Note that these Main line after c4 The main line of AlphaZero after 1. c4
are not purely deterministic, and each of the given lines is in Self-capture chess is:
87
1. c4 (book) e5 2. g3 d5 3. cxd5 Nf6 4. Bg2 Nxd5 5. Nc3 Nb6 And here we see the first self-capture of the game, creating
6. b3 Nc6 7. Bb2 f6 8. Rc1 Bf5 9. Bxc6+ bxc6 10. Nf3 Qd7 threats down the h-file:
11. O-O Be7 12. d3 a5 13. Ne4 O-O 14. Qc2 a4 15. Qxc6
37. . . Rxh6 38. Qf3 Qh1+
Qxc6 16. Rxc6 Nd5 17. Nc3 Nxc3 18. Rxc3 axb3 19. axb3
Rfb8 20. Rxc7 Bd8
8
rs0a0ZkZ 8
rZ0Z0ZkZ
7
Z0S0Z0op 7
ZpZ0ZpZ0
6
0Z0Z0o0Z 6
pZpm0Z0s
5
Z0Z0obZ0 5
Z0ZpZ0Z0
4
0Z0Z0Z0Z 4
PO0O0ZPZ
3
ZPZPZNO0 3
Z0ZBOQZ0
2
0A0ZPO0O 2
0Z0Z0OPZ
1
Z0Z0ZRJ0 1
S0S0Z0Jq

Game AZ-33: AlphaZero Self-capture vs AlphaZero The end? Not really. In self-capture chess the king can
Self-capture The first ten moves for White and Black escape by capturing its way through its own army, and
have been sampled randomly from AlphaZero’s opening hence here it just takes on f2 and gets out of check.
“book”, with the probability proportional to the time spent 39. Kxf2 Qh4+ 40. Ke2 Re8 41. Rh1 Qxh1 42. Rxh1 Rxh1
calculating each move. The remaining moves follow best 43. Qf4 Ne4 44. Bxe4 Rxe4 45. Qb8+ Kg7 46. Qxb7 Rh6
play, at roughly one minute per move. 47. Qxa6 Rxg4 48. Kf1 Rh1+ 49. Kf2 Rg6 50. b5 Rh2 51. b6
1. d4 Nf6 2. c4 e6 3. Nc3 d5 4. cxd5 exd5 5. Bg5 c6 6. Qc2 Rgxg2+ 52. Kf3 Rf2+
Nbd7 7. Nf3 h6 8. Bh4 Be7 9. e3 O-O 10. Bd3 Re8 11. O-O
Ne4 12. Bxe4 Bxh4 13. Bh7+ Kh8 14. Nxh4 Qxh4 15. Bd3
Qe7 16. a3 Nf6 17. b4 Bd7 18. h3 Kg8 19. Rfb1 Rec8
20. Qd1 Be6 21. Ne2 a6 22. Nf4 Ne8 23. a4 Nd6 24. Qb3 8
0Z0Z0Z0Z
Z0Z0Zpj0
Qd7 25. Be2 Bf5 26. Rc1 Qd8 27. Qb2 Be4 28. Rc5 Bf5 7
29. Rc3 Ra7 30. Rcc1 Raa8 31. Rc5 Qh4 32. Bf1 Re8
33. Rcc1 g5 34. Nd3 Bxd3 35. Bxd3 g4 36. hxg4 Re6
37. Qe2
6
QOpZ0Z0Z
5
Z0ZpZ0Z0
8
rZ0Z0ZkZ 4
PZ0O0Z0Z
7
ZpZ0ZpZ0 3
Z0Z0OKZ0
6
pZpmrZ0o 2
0Z0Z0s0s
5
Z0ZpZ0Z0 1
Z0Z0Z0Z0
a b c d e f g h
4
PO0O0ZPl
3
Z0ZBO0Z0
0Z0ZQOPZ
Unlike in classical chess, White can still play on here, and
2
AlphaZero does, by advancing the king forward with a self-
1
S0S0Z0J0 capture!
a b c d e f g h 53. Kxe3 Rb2 54. a5 Rb3+
88
0Z0Z0Z0Z
8 12. Rd1 Qe7 13. h4 Nxe5 14. Bxe5 a6 15. a4 Rg8 16. hxg5
hxg5
7
Z0Z0Zpj0
6
QOpZ0Z0Z
5
O0ZpZ0Z0
4
0Z0O0Z0Z
8
rZ0ZkZrZ
3
ZrZ0J0Z0
7
ZbZ0lpa0
2
0Z0Z0Z0s
6
pZpZpm0Z
1
Z0Z0Z0Z0
5
ZpZ0A0o0
a b c d e f g h 4
PZpOPZ0Z
And, as if one pawn was not enough, White self-captures 3
Z0M0Z0Z0
0OQZBOPZ
another one by taking on d4. 2
55. Kxd4 Ra2 56. Ke5 Rb5 57. b7 Raxa5 58. Qxa5 Rxa5
59. b8=Q Ra2
1
Z0ZRJ0ZR
a b c d e f g h
8
0L0Z0Z0Z
7
Z0Z0Zpj0
6
0ZpZ0Z0Z 17. Qc1 O-O-O 18. Qxg5 Nd5 19. Qxe7 Nxe7 20. g3 Bxe5
5
Z0ZpJ0Z0 21. dxe5 Rxd1+ 22. Kxd1 Rd8+ 23. Kc1 b4
4
0Z0Z0Z0Z
3
Z0Z0Z0Z0
2
rZ0Z0Z0Z 8
0Zks0Z0Z
1
Z0Z0Z0Z0 7
ZbZ0mpZ0
a b c d e f g h
6
pZpZpZ0Z
White manages to get a queen, but in the end, Black’s de-
fensive resources prove sufficient and the game eventually
5
Z0Z0O0Z0
ends in a draw. 4
PopZPZ0Z
60. Kd6 Re2 61. Kxc6 Re6+ 62. Kd7 Rg6 63. Qa8 Re6
64. Qxd5 Kg8 65. Qa8+ Kg7
3
Z0M0Z0O0
With draw soon to follow.
2
0O0ZBO0Z
1/2–1/2
1
Z0J0Z0ZR
a b c d e f g h
Game AZ-34: AlphaZero Self-capture vs AlphaZero
Self-capture The first ten moves for White and Black
calculating each move. The remaining moves follow best Here we come to the first self-capture of the game, White
play, at roughly one minute per move. decides to give up the a4 pawn in order to get the knight to
an active square.
1. d4 d5 2. c4 e6 3. Nc3 Nf6 4. Nf3 c6 5. Bg5 h6 6. Bh4 dxc4
7. e4 g5 8. Bg3 b5 9. Be2 Bb7 10. Ne5 Nbd7 11. Qc2 Bg7 24. Nxa4
89
0Zks0Z0Z
8 And the game eventually ended in a draw. 1/2–1/2
7
ZbZ0mpZ0 Self-capture The first ten moves for White and Black have
6
pZpZpZ0Z been sampled randomly from AlphaZero’s opening “book”,
5
Z0Z0O0Z0 ing each move. The remaining moves follow best play, at
4
NopZPZ0Z 1. d4 e6 2. Nf3 Nf6 3. c4 d5 4. Bg5 dxc4 5. Nc3 a6 6. e4 b5
3
Z0Z0Z0O0 7. e5 h6 8. Bh4 g5 9. Nxg5 hxg5 10. Bxg5 Nbd7
2
0O0ZBO0Z
1
Z0J0Z0ZR 8
rZblka0s
a b c d e f g h
7
Z0onZpZ0
And Black responds in turn with a self-capture of its own,
on c6!
6
pZ0Zpm0Z
24. . . Nxc6
5
ZpZ0O0A0
4
0ZpO0Z0Z
8
0Zks0Z0Z 3
Z0M0Z0Z0
7
ZbZ0ZpZ0 2
PO0Z0OPO
6
pZnZpZ0Z 1
S0ZQJBZR
5
Z0Z0O0Z0 a b c d e f g h
4
NopZPZ0Z
Z0Z0Z0O0
3 In this highly tactical position, self-captures provide addi-
tional resources, as AlphaZero quickly demonstrates, by
2
0O0ZBO0Z a self-capture on g2, developing the bishop on the long
diagonal at the price of a pawn.
1
Z0J0Z0ZR 11. Bxg2
a b c d e f g h
25. Nb6+ Kc7 26. Nxc4 Nd4 27. Bd3 Nf3 28. Bc2 Rd4
29. Nd6 Nxe5 30. Nxb7 Kxb7 31. f4 Nd3+ 32. Bxd3 Rxd3
33. Rh7 Rxg3 34. Rxf7+ Kc6 35. Rf6 Kd7 36. Rf7+ Kc6
8
rZblka0s
37. Rf6 Kd7 38. f5 exf5 39. exf5 Rf3
7
Z0onZpZ0
6
pZ0Zpm0Z
8
0Z0Z0Z0Z 5
ZpZ0O0A0
7
Z0ZkZ0Z0 4
0ZpO0Z0Z
6
pZ0Z0S0Z 3
Z0M0Z0Z0
5
Z0Z0ZPZ0 2
PO0Z0OBO
4
0o0Z0Z0Z 1
S0ZQJ0ZR
3
Z0Z0ZrZ0 a b c d e f g h
2
0O0Z0Z0Z
1
Z0J0Z0Z0 Yet, Black responds in turn by a self-capture on a6:
a b c d e f g h 11. . . Rxa6
90
0Zblka0s
8 31. Kc2 Bxc5 32. Bxc5 Ra6 33. a3 Bxd5 34. Rxd5 Rxh4
35. Rxf5
7
Z0onZpZ0
6
rZ0Zpm0Z
8
0Z0ZkZ0Z
5
ZpZ0O0A0
7
Z0Z0ZpZ0
4
0ZpO0Z0Z
6
rZ0Z0Z0Z
3
Z0M0Z0Z0
5
Z0A0ZRZ0
2
PO0Z0OBO
4
0Z0Z0Z0s
1
S0ZQJ0ZR
3
O0O0Z0Z0
a b c d e f g h 2
0ZKZ0O0Z
12. exf6 Rg8 13. h4 Nxf6 14. Nxb5 Be7 15. Qc2 Nd5
1
Z0Z0Z0Z0
16. Qh7 a b c d e f g h
and the game eventually ended in a draw. 1/2–1/2

8
0ZblkZrZ
Z0o0apZQ
7 Game AZ-36: AlphaZero Self-capture vs AlphaZero
Self-capture The first ten moves for White and Black
6
rZ0ZpZ0Z have been sampled randomly from AlphaZero’s opening
5
ZNZnZ0A0 calculating each move. The remaining moves follow best
0ZpO0Z0O
4 play, at roughly one minute per move.
In this game, self-captures happen towards the end, but the
3
Z0Z0Z0Z0 game itself is pretty tactical and entertaining. We therefore
2
PO0Z0OBZ included the full game.
S0Z0J0ZR
1 1. Nf3 d5 2. d4 Nf6 3. c4 e6 4. Nc3 c6 5. Bg5 h6 6. Bh4
dxc4 7. e4 g5 8. Bg3 b5 9. Be2 Bb7 10. O-O Nbd7 11. Ne5
a b c d e f g h h5 12. Nxd7 Qxd7
16. . . Rf8 17. Bh6 Nf6 18. Qc2 Rg8 19. Bf3 c6 20. Nc3
Qxd4 21. Be3 Qe5 22. O-O-O Nd5
8
rZ0Zka0s
7
obZqZpZ0
8
0ZbZkZrZ 6
0ZpZpm0Z
7
Z0Z0apZ0 5
ZpZ0Z0op
6
rZpZpZ0Z 4
0ZpOPZ0Z
5
Z0Znl0Z0 3
Z0M0Z0A0
4
0ZpZ0Z0O 2
PO0ZBOPO
3
Z0M0ABZ0 1
S0ZQZRJ0
2
POQZ0O0Z a b c d e f g h
1
Z0JRZ0ZR In the game, White played the pawn to a3, but it’s interesting
a b c d e f g h to note that potential self-captures factor in the lines that
AlphaZero is calculating at this point. AlphaZero is initially
23. Kb1 Nxc3+ 24. bxc3 c5 25. Rhg1 Rh8 26. Rg4 Qf5 considering the following line: 13. Qd2 Be7 14. Qxg5 b4
27. Qxf5 exf5 28. Rxc4 Be6 29. Bd5 Rd6 30. Rxc5 Rb6+ 15. Na4 Qxc6
91
rZ0ZkZ0s
8 21. Bg3 Bh6 22. h5 Rgg8 23. b3 Rxg3
7
obZ0apZ0 8
0Zks0Z0Z
6
0ZqZpm0Z 7
obZ0lpZ0
5
Z0Z0Z0Lp 6
0ZpZpm0a
4
NopOPZ0Z 5
ZpZ0Z0ZP
3
Z0Z0Z0A0 4
0ZpOPZ0Z
2
PO0ZBOPO 3
OPM0ZQs0
1
S0Z0ZRJ0 2
0Z0ZBOpZ
a b c d e f g h
analysis diagram
1
S0ZRZ0J0
a b c d e f g h
where Black has just self-captured its c6 pawn! 16. Nc5
Nxe4 17. Qe5 with exchanges to follow. Going back to the 24. Qxg3 Rg8 25. Qh3 Bf4 26. Qh4 Bg5 27. Qh2 cxb3
game: 28. Rd3 a6 29. Rb1 c5 30. dxc5 Nd7
13. a3 Rh6 14. Qc1 h4
8
0ZkZ0ZrZ
8
rZ0Zka0Z 7
ZbZnlpZ0
7
obZqZpZ0 6
pZ0ZpZ0Z
6
0ZpZpm0s 5
ZpO0Z0aP
5
ZpZ0Z0o0 4
0Z0ZPZ0Z
4
0ZpOPZ0o 3
OpMRZ0Z0
3
O0M0Z0A0 2
0Z0ZBOpL
2
0O0ZBOPO 1
ZRZ0Z0J0
1
S0L0ZRJ0 a b c d e f g h
a b c d e f g h 31. Rg3 Rg7 32. Qxg2 f5 33. Rxb3 Nxc5 34. Rb4 Qf6
35. Bf1 fxe4
15. Be5 h3 16. Qxg5 hxg2 17. Rd1 Rg6 18. Qf4 Qe7 19. Qf3
0ZkZ0Z0Z
Bg7 20. h4 O-O-O
8
8
0Zks0Z0Z 7
ZbZ0Z0s0
7
obZ0lpa0 6
pZ0Zpl0Z
6
0ZpZpmrZ 5
Zpm0Z0aP
5
ZpZ0A0Z0 4
0S0ZpZ0Z
4
0ZpOPZ0O 3
O0M0Z0S0
3
O0M0ZQZ0 2
0Z0Z0OQZ
2
0O0ZBOpZ 1
Z0Z0ZBJ0
S0ZRZ0J0
a b c d e f g h
1
a b c d e f g h 36. Nxe4 Nxe4 37. Rxe4 Bxe4 38. Qxe4 Bf4 39. Rg6 Rxg6+
92
40. hxg6 Qg5+ 41. Bg2 Be5 42. f4 Bxf4 43. Qb7+ Kd8 have been sampled randomly from AlphaZero’s opening
44. g7 Qc5+ “book”, with the probability proportional to the time spent
8
0Z0j0Z0Z play, at roughly one minute per move.
ZQZ0Z0O0
1. d4 Nf6 2. c4 e6 3. Qc2 c5 4. dxc5 h6 5. Nf3 Bxc5 6. a3
7
O-O 7. Bf4 Qa5+ 8. Nbd2 Nc6 9. e3 Re8 10. Bg3 e5 11. Bh4
6
pZ0ZpZ0Z g5
5
Zpl0Z0Z0
4
0Z0Z0a0Z
3
O0Z0Z0Z0 8
rZbZrZkZ
2
0Z0Z0ZBZ 7
opZpZpZ0
1
Z0Z0Z0J0 6
0ZnZ0m0o
a b c d e f g h 5
l0a0o0o0
What happens next is a rather remarkable self-capture,
demonstrating that it’s not only the pawns that can justi-
4
0ZPZ0Z0A
fiably be self-captured, as the least valuable pieces. Indeed,
3
O0Z0ONZ0
White self-captures the bishop on g2, in its attempt at avoid-
ing perpetuals!
2
0OQM0OPO
45. Kxg2 Qg5+ 46. Kf1
1
S0Z0JBZR
a b c d e f g h
8
0Z0j0Z0Z
7
ZQZ0Z0O0 12. Nxg5 hxg5 13. Bxg5 Re6 14. O-O-O Bf8 15. h4 d5
6
pZ0ZpZ0Z
5
ZpZ0Z0l0
4
0Z0Z0a0Z 8
rZbZ0akZ
3
O0Z0Z0Z0 7
opZ0ZpZ0
2
0Z0Z0Z0Z 6
0ZnZrm0Z
1
Z0Z0ZKZ0 5
l0Zpo0A0
a b c d e f g h
Yet, Black responds in turn, by capturing its own bishop!

4
0ZPZ0Z0O
The game ultimately ends in a draw. 3
O0Z0O0Z0
46. . . Qxf4+ 47. Kg2 Qg4+ 48. Kf2 Qf4+ 49. Ke2 Qe5+
50. Kf3 Qf5+ 51. Ke3 Qe5+ 52. Kd3 Qd6+ 53. Ke4 Qxe6+
2
0OQM0OPZ
54. Kf4 Qf6+ 55. Ke3 Qe5+ 56. Kd3 Qd6+ 57. Ke4 Qe6+
58. Kd4 Qf6+ 59. Kd5 Qf3+ 60. Kd6 Qf6+ 61. Kc5 Qf2+
1
Z0JRZBZR
a b c d e f g h
62. Kb4 Qd2+ 63. Kb3 Qd3+ 64. Kb2 Qd2+ 65. Kb1 Qd3+
66. Kb2 Qd2+ 67. Kb3 Qd3+ 68. Kb4 Qd2+ 69. Kxa3 Qc3+
70. Ka2 Qc2+ 71. Ka1 Qc1+ 72. Ka2 Qc2+ 73. Ka1 Qc1+
74. Ka2 Qc2+ 1/2–1/2
Here we see the first self-capture move of the game, creating
threats along the h-file:
Self-capture The first ten moves for White and Black 16. Rxh4
93
8
rZbZ0akZ 8
rZbZ0ZkZ
7
opZ0ZpZ0 7
opZ0Zpa0
6
0ZnZrm0Z 6
0lnZ0ZrZ
5
l0Zpo0A0 5
Z0ZRo0Z0
4
0ZPZ0Z0S 4
0Z0Z0Z0S
3
O0Z0O0Z0 3
ONZ0A0Z0
2
0OQM0OPZ 2
0OQZ0OPZ
1
Z0JRZBZ0 1
Z0J0ZBZ0
20. . . Qc7 21. Rc5 b6 22. Rc3 Bb7 23. Bd3 Qd8 24. g3
Rd6 25. Bc4 Kf8 26. Qh7 Qf6 27. Nd2 Ne7 28. Rg4 Rad8
29. Bg5 Qxf2
It’s interesting to note that White could have also tried open-
ing the h-file a move earlier, by playing 15. Rxh2 instead of
8
0Z0s0j0Z
15. h4, but AlphaZero prefers provoking 15. . . d5 first and 7
obZ0mpaQ
having its rook on the 4th rank, where it stands more active
and controls additional squares.
6
0o0s0Z0Z
16. . . Bg7 17. Nb3 Qb6 18. cxd5 Nxd5 19. Rxd5 Rg6
5
Z0Z0o0A0
4
0ZBZ0ZRZ
3
O0S0Z0O0
2
0O0M0l0Z
8
rZbZ0ZkZ
1
Z0J0Z0Z0
a b c d e f g h
7
opZ0Zpa0 30. Bxe7+ Kxe7 31. Rxg7 Qe1+ 32. Kc2 Be4+
6
0lnZ0ZrZ
5
Z0ZRo0A0 8
0Z0s0Z0Z
4
0Z0Z0Z0S 7
o0Z0jpSQ
3
ONZ0O0Z0 6
0o0s0Z0Z
2
0OQZ0OPZ 5
Z0Z0o0Z0
1
Z0J0ZBZ0 4
0ZBZbZ0Z
a b c d e f g h 3
O0S0Z0O0
2
0OKM0Z0Z
1
Z0Z0l0Z0
a b c d e f g h
Here comes another self-capture:
33. Nxe4 Qd1+ is what is played and made possible by a
20. Bxe3 self-capture, avoiding mate:
94
rZ0l0skZ
34. Kxb2 8
7
o0Znapo0
8
0Z0s0Z0Z 6
0Z0Z0m0Z
7
o0Z0jpSQ 5
Z0o0Z0Zp
6
0o0s0Z0Z 4
0ZNZ0A0O
5
Z0Z0o0Z0 3
Z0Z0ZQZ0
4
0ZBZNZ0Z 2
PZ0ZNOPZ
3
O0S0Z0O0 1
Z0ZRZRJ0
2
0J0Z0Z0Z a b c d e f g h
1
Z0ZqZ0Z0
a b c d e f g h
In this position, with Black to play, in classical chess Black
would struggle to find a good plan and activity. Yet, here in
Here Black responds by a self-capture on b6: self-capture chess, Black plays the obvious idea – sacrificing
34. . . Rxb6+ the a7 pawn to open the a-file for its rook and initiate active
play!
19. . . Rxa7 20. Nc3 Qa8 21. Qg3 Rfd8
8
0Z0s0Z0Z
7
o0Z0jpSQ
6
0s0Z0Z0Z 8
qZ0s0ZkZ
5
Z0Z0o0Z0 7
s0Znapo0
4
0ZBZNZ0Z 6
0Z0Z0m0Z
3
O0S0Z0O0 5
Z0o0Z0Zp
2
0J0Z0Z0Z 4
0ZNZ0A0O
1
Z0ZqZ0Z0 3
Z0M0Z0L0
a b c d e f g h
2
PZ0Z0OPZ
The game soon ends in a draw.
1
Z0ZRZRJ0
a b c d e f g h
35. Rb3 Rxb3+ 36. Bxb3 Qe2+ 37. Kb1 Qf1+ 38. Ka2 Qe2+
39. Ka1 Qe1+ 40. Ka2 Qe2+ 41. Kb1 Qe1+ 42. Kc2 Qe2+
43. Kc3 Qe3+ 44. Kb2 Qe2+ 45. Kxa3 Qa6+ 46. Ba4 Qd3+
47. Ka2 Qe2+ 48. Ka1 Qe1+ 49. Kb2 Qe2+ 50. Ka1 Qe1+ Black soon managed to equalize and eventually draw the
51. Ka2 Qe2+ 52. Ka3 Qd3+ 53. Bb3 Qa6+ 54. Kb2 Qe2+ game. 1/2–1/2
55. Kb1 Qf1+ 56. Ka2 Qa6+ 57. Kb2 Qe2+ 58. Ka1 Qf1+
59. Ka2 Qa6+ 60. Kb1 Qf1+ 61. Kb2 Qe2+ 1/2–1/2
Game AZ-38: AlphaZero Self-capture vs AlphaZero Game AZ-39: AlphaZero Self-capture vs AlphaZero
Self-capture The following position, with Black to play, Self-capture The following position, with White to play,
arose in an AlphaZero game, played at roughly one minute arose in an AlphaZero game, played at roughly one minute
per move. per move.
95
8
0Z0Z0Z0Z 8
0Z0Z0Z0Z
7
oBo0Z0j0 7
oBo0Z0j0
6
0oPZ0ZpZ 6
0oPZbZpZ
5
Z0Z0Z0Z0 5
Z0Z0Z0Z0
4
0ZPZ0abo 4
0ZPZ0Z0S
3
Z0Z0Z0Z0 3
Z0Z0Z0a0
2
PZ0Z0ZPZ 2
PZ0Z0ZPZ
1
S0Z0Z0ZK 1
Z0Z0Z0ZK
38. . . Bxh4 39. cxb7
In the previous moves, AlphaZero had manoeuvred its light-

8
0Z0Z0Z0Z
squared bishop to b7 via a6, with a clear intention of setting
up threats to self-capture on b7 and promote the pawn on b8.
7
oPo0Z0j0
Yet, if attempted immediately, Black can respond in turn by
6
0o0ZbZpZ
playing c6, c5, or even self-capturing on c7 with the bishop.
If the bishop moves away from the b8-h2 diagonal, White
5
Z0Z0Z0Z0
can proceed with the plan. This explains why White plays
the following next:
4
0ZPZ0Z0a
34. Rc1
3
Z0Z0Z0Z0
2
PZ0Z0ZPZ
1
Z0Z0Z0ZK
a b c d e f g h
8
0Z0Z0Z0Z And White went on to eventually win the game. 1–0
7
oBo0Z0j0 Game AZ-40: AlphaZero Self-capture vs AlphaZero
6
0oPZ0ZpZ Self-capture The following position, with White to play,
arose in an AlphaZero game, played at roughly one minute
5
Z0Z0Z0Z0 per move.
4
0ZPZ0abo
3
Z0Z0Z0Z0
8
RZ0Z0Z0Z
2
PZ0Z0ZPZ
7
ZNZ0Z0Z0
1
Z0S0Z0ZK
6
PZ0Z0a0Z
a b c d e f g h 5
Z0Z0Z0j0
4
rZ0Z0Z0o
3
Z0Z0Z0Z0
The rook can now be taken on c1, but this would allow the
2
0Z0ZKZ0Z
promotion of the c-pawn via a self-capture. 1
Z0Z0Z0Z0
34. . . Be6 35. Rf1 Bd6 36. Rd1 Bf4 37. Rd4 Bg3 38. Rxh4 a b c d e f g h
96
In this position, White plays a self-capture, 50. axb7, giving and has very strong threats against the Black king. In Clas-
away the knight, for an immediate threat of promoting on sical chess, those might prove fatal, but here Black uses a
b8. This is a common pattern in endgames in this variation, self-capture as a defensive resource, as can be seen in the
where pieces can be used to help promote the passed pawns. following forcing sequence:
34. Rxh7+ Kxh7 35. Rh4+ Kxg8 – Black is forced to capture
its own rook to avoid checkmate – 36. f4 Ng6 37. Rh2 Qxa2
Self-capture The following position, with Black to play,
38. Qc1 Qa4 39. Qc4+ Qxc4 40. Bxc4+
arose in an AlphaZero game, played at roughly one minute
per move.
8
0ZrZ0ZkZ
8
0Z0Z0Z0Z 7
o0Z0Z0o0
7
ZpZ0ZpZ0 6
0o0Z0onZ
6
0Zpj0s0o 5
Z0m0Z0Z0
5
s0ZpZpZP 4
0ZBZPO0Z
4
PZ0OnJ0Z 3
Z0Z0Z0Z0
3
Z0ZBO0O0 2
0A0Z0Z0S
2
0ZRZ0O0Z 1
Z0Z0Z0ZK
1
ZRZ0Z0Z0 a b c d e f g h
a b c d e f g h
And here Black uses the second self-capture in this sequence,
In this position, AlphaZero as Black plays another self- 40. . . Kxg7, to secure the king.
capture motif: 75. . . fxe4+, self-capturing its own knight
with check, while attacking White’s bishop on d3. This Game AZ-43: AlphaZero Self-capture vs AlphaZero
highlights novel tactical opportunities where self-captures Self-capture The following position, with White to play,
can be utilised not only as dynamic material sacrifices for arose in a fast-play AlphaZero game, played at roughly one
the initiative, but rather a key part of tactical sequences second per move.
where material gets immediately recovered.

8
0Z0Z0skZ
Self-capture The following position, with White to play,
arose in a fast-play AlphaZero game, played at roughly one
7
ZQZ0Z0Z0
second per move. 6
0Z0Z0a0S
8
0ZrZ0Zrj
5
Z0Z0ZpZ0
7
o0Z0Zqop
4
0Z0OnO0Z
6
0o0Z0o0Z
3
Z0Z0Z0Z0
5
Z0m0m0Z0
2
0Z0Z0Z0O
4
0Z0ZPZRZ
1
Z0Z0ZqAK
a b c d e f g h
3
Z0Z0LPZ0 With White to play, in Classical chess this would result in a
2
PA0ZBZ0S mate in one move, on h7. Yet, in Self-capture chess Black
1
Z0Z0Z0ZK can escape by self-capturing its rook on f8, at Which point
White has to attend to its own king’s safety.
a b c d e f g h
45. Qh7+ Kxf8 46. Rxf6+ Nxf6 47. Qg6 Qf3+ 48. Qg2
At the moment, White is two pawns down for the attack Qxf4, leading to a simplified position.
97
0Z0j0s0Z
Game AZ-44: AlphaZero Self-capture vs AlphaZero 8
Self-capture The following position, with White to play,
arose in a fast-play AlphaZero game, played at roughly one
second per move.
7
opZbaQop
6
0Z0ZpZ0Z
5
Z0Z0l0M0
4
0ZBZ0Z0Z
8
0Z0Zka0s 3
O0Z0Z0Z0
7
opZbZpop 2
0Z0Z0ZPO
6
0Z0ZpZ0Z 1
Z0ZRZ0ZK
5
Z0Z0O0Z0 a b c d e f g h
4
0Z0oNZ0Z 25. . . Qxg5 26. Qxe6
3
O0Z0Z0L0
2
0lBZ0OPO
8
0Z0j0s0Z
1
Z0Z0ZRZK
7
opZba0op
a b c d e f g h 6
0Z0ZQZ0Z
5
Z0Z0Z0l0
4
0ZBZ0Z0Z
In this position, with White to move, White self-captures
a pawn to open up dynamic possibilities against the Black
3
O0Z0Z0Z0
king on the f-file. 2
0Z0Z0ZPO
20. Qxf2 d3 21. Qxf7+ Kd8 22. Bxd3 Qxe5 23. Rd1 Be7
24. Bc4
1
Z0ZRZ0ZK
a b c d e f g h
Here, Black utilizes a self-capture for defensive purposes,

giving up the e7 bishop
0Z0j0Z0s
26. . . Qxe7 27. Qh3 Rf6 28. Qg3 Kc8 29. Re1 Qd6 30. Qxg7
8
Bc6
7
opZbaQop
6
0Z0ZpZ0Z
8
0ZkZ0Z0Z
5
Z0Z0l0Z0
7
opZ0Z0Lp
4
0ZBZNZ0Z
6
0Zbl0s0Z
3
O0Z0Z0Z0
5
Z0Z0Z0Z0
2
0Z0Z0ZPO
4
0ZBZ0Z0Z
1
Z0ZRZ0ZK
3
O0Z0Z0Z0
a b c d e f g h 2
0Z0Z0ZPO
1
Z0Z0S0ZK
a b c d e f g h
24. . . Rf8 25. Ng5 with a roughly equal position.
98

Assessing Game Balance With Alphazero: Exploring Alternative Rule Sets in Chess

Uploaded by

Copyright:

Available Formats

Assessing Game Balance With Alphazero: Exploring Alternative Rule Sets in Chess

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Assessing Game Balance With Alphazero: Exploring Alternative Rule Sets in Chess

Uploaded by

Copyright:

Available Formats

Assessing Game Balance with AlphaZero:

Exploring Alternative Rule Sets in Chess

Nenad Tomašev* Ulrich Paquet* Demis Hassabis Vladimir Kramnik

Variant Primary rule change Secondary rule change

Table 1. A list of considered alterations to the rules of chess.

Classical 772 8820 409 Classical 18 979 3

Variant Training 1sec 1min 3.2.1. I NFERENCE FOR GAME ODDS

row column row column

Dutch Defence Chigorin Defence

Classical 218 743 39 Classical 197 782 21

Pawn back 123 855 22 Pawn back 231 756 13

Semi-torpedo 220 719 61 Semi-torpedo 245 716 39

Torpedo 339 595 66 Torpedo 393 569 38

Stalemate = win 282 670 48 Stalemate = win 193 780 27

No-castling 221 749 30 No-castling 359 619 22

No-castling (10) 229 752 19 No-castling (10) 254 728 18

Self-capture 224 750 26 Self-capture 220 769 11

Pawn sideways 139 835 26 Pawn sideways 720 273 7

Classical 204 778 18 Classical 63 703 234

No-castling 444 534 22 Pawn back 26 776 198

No-castling (10) 388 602 10 Semi-torpedo 66 679 255

Pawn back 261 721 18 Torpedo 99 630 271

Semi-torpedo 296 673 31 Stalemate = win 54 614 332

Torpedo 432 536 32 No-castling 118 669 213

Stalemate=win 228 735 37 No-castling (10) 29 674 297

Self-capture 262 717 21 Self-capture 64 717 219

Pawn-sideways 434 552 14 Pawn sideways 45 574 381

Move sequence entropy Let s = s1:t = [s1 , s2 , . . . st ]

which is not an option in no-castling chess. Yet, castling is

3.6. Differences between opening trees

Average number of candidate moves

6.0 7.0 No-castling (e4) (b)

Average additional candidate moves

s0t+1 max p(st+1 |st ), q(st+1 |st )

No-castling (10) Classical 7.17 that P and Q would consider individually.

Kramnik, V. Kramnik And AlphaZero: How To Re-

Lc0. Leela Chess Zero. https://lczero.org/ (ac-

We show below that

A.2. Additional figures

0.05 Classical 0.05 Classical

0.06 Classical Classical

0.175 Stalemate=win 0.16 Stalemate=win

Average number of candidate moves

(a) Classical chess (b) No-castling chess

Average number of candidate moves

(c) No-castling (10) chess (d) Pawn one square chess

Average number of candidate moves

8 Stalemate=win (b) 6.00 Torpedo (b)

(e) Stalemate=win chess (f) Torpedo chess

Average number of candidate moves

(g) Semi-torpedo chess (h) Pawn-back chess

Average number of candidate moves

(i) Pawn-sideways chess (j) Self-capture chess

Uniform random (w)

Uniform random (b)

(k) Uniform random moves under classical chess rules

B. Appendix B.1.2. A SSESSMENT

required to give a clear answer on the guiding

B.1.3. M AIN L INES