Homeostasis and Rein Control: From Daisyworld to Active Perception

Inman Harvey
Evolutionary and Adaptive Systems Group
Centre for Computational Neuroscience and Robotics
Centre for the Study of Evolution
COGS/Informatics, School of Science and Technology, University of Sussex, BN1 9QH, UK
inmanh@cogs.susx.ac.uk

Abstract regulating a global variable. Individual acts of behaviour

Homeostasis refers to the ability of organisms to maintain typically have a cost to the individual; yet the net effect
vital properties, such as body temperature, within a zone of from just one individual on global temperature is
viability, or of comfort, and the Gaia Hypothesis proposes insignificant. So selection at the individual level will favour
that the Earth with its biota acts as a homeostatic whole. The those profligates who do not care to do their bit towards
Daisyworld model was proposed as one possible mechanism global homeostasis. These arguments appeared initially to
for providing this homeostatic regulation. Here a new and give sound evolutionary reasons why the global regulation
much simplified version of this model is presented, proposed in the Gaia Hypothesis could not be maintained,
demonstrating that the combination of any ‘Hat function’ or indeed even have arisen in the first place.
with any feedback, positive or negative, can lead to
homeostasis through ‘Rein Control’. This principle is so
general that it can be extended to other domains such as Daisyworld
active perception, here demonstrated in a simulated robot. The Gaia Hypothesis was originally restricted to the claim
(based on observation) that such global regulation existed,
without any theory or mechanism to explain how it might
Introduction happen; it initially met with much skepticism. The ‘parable
Living organisms have many physiological variables that of Daisyworld’ was then proposed as a possible mechanism
must be maintained within upper and lower bounds for (Lovelock 1983, Watson and Lovelock 1983).
continued survival. Typically there are regulatory Whereas the surface temperature of a lifeless planet
mechanisms that maintain these variables within these would have changed dramatically with the increase in
bounds even in the face of substantial environmental luminosity of the sun over geological timescales,
perturbations. These homeostatic systems can take many observation indicates that on our Earth it has remained
forms. For mammals such as humans that need to maintain remarkably constant around temperatures suitable for life.
their body temperature within fairly tight limits, the variety Daisyworld is a deliberately simplified model of an
of mechanisms includes physiological, reflex behaviour imaginary planet with just two species of daisies, black and
such as shivering and sweating, and more considered white, that demonstrates how this could happen. The
behaviour such as moving towards or away from heat. growth rate of the daisies depends only on their local
Such regulatory mechanisms are important for the temperature, but in turn the daisies modify this because of
survival of the organism, and often appear to be complex, differences in the way they absorb radiation; black daisies
subtle, and so crafted as to provide near-optimal conditions have low albedo (or reflectivity) and heat up easily,
for the organism; their origins and maintenance are usually whereas white daisies with higher albedo tend to reflect the
attributed to the power of selection over many generations sun’s radiation back.
of Darwinian evolution. So controversy was inevitable
1.2
when Lovelock (1972) proposed the Gaia Hypothesis that
the Earth with its biota acts as a homeostatic geo- 1

physiological system that regulates global properties, e.g. 0.8

temperature at the Earth’s surface, within a range that
Growth rate

0.6
provides viable living conditions for the biota; for a review response
0.4
see (Lenton 1998). No immediate explanation was
available for the origin or cause of any such regulation, as 0.2
clearly the Earth as a whole had not evolved through 0
successive generations of selection within a population of 0 10 20 30 40 50
-0.2
Earths. Further, as species within the biota, and individuals Temperature
within each species, were competing with each other, it was
difficult to see how selection at the individual level could Figure 1.Hat function: daisy response to temperature
favour behaviour that led to global cooperation in
 The dependency of growth rates on temperature is to stay within their sensitive regions; in other words,
assumed to be a Hat-shaped function as in figure 1. The phototaxis despite the random nature of the feedback.
Daisyworld model demonstrates how these feedbacks via +ve
the environment, both positive for black and negative for
white daisies, result in regulation of the planetary DB TB
temperature. Watson and Lovelock (1983) demonstrate that
External
with their particular parameter values and equations, the Forcing
resulting close-coupled system regulates the temperature to Hat-functions Leakage
within the viability zone (here 50 to 400C) over a far wider
range of solar luminosities than would have been the case
in the absence of any daisies. D TW
W

Comprehensibility of Daisyworld
-ve
Daisyworld is intentionally a simplified model; temperature
is taken as just one example of an essential variable that
Figure 2. Interactions in the cut-down model: a black
can be regulated, and the lessons from the Daisyworld
daisybed above, separate white one below. Both receive
parable are meant to have far wider scope. One point made
external forcing from the sun, and the only interaction
early on is that the precise form of the Hat function is
between them is by ‘leakage’ or heat conductance.
unimportant, provided it has the general peaked character
around an optimum temperature, with the brim of the hat
representing here a zero growth rate outside the viability Cut-down Daisyworld
range. Nevertheless, the use of the Stefan-Boltzmann law to
calculate absorbed and emitted radiation, and use of the The simplifications are twofold: firstly, the model is
equations governing the comparative and indeed idealized into a simpler form with fewer interactions;
competitive growth rates of the different daisies, means that secondly, the Hat function and the putative underlying laws
analysis of this system is not a trivial problem. of radiation and heat conductance are simplified into linear
I have therefore adopted the strategy of radically or piecewise linear form.
simplifying the usual Daisyworld model, to see how much The black and white daisies can be modeled as growing
can be left out whilst still retaining the homeostatic on separate daisybeds, in other words not competing for
regulation. In doing so, firstly it becomes clearer how space. The interactions and feedbacks are then limited to
crucial is the difference in local temperatures between those shown in figure 2. The Hat function can be replaced
black and white daisies, something often obscured by the by a piecewise linear function of similar general form,
conventional graphs shown; secondly it becomes much which I call a ‘Witch’s Hat’ function. TB and TW are the
easier to visualize the very simple underlying feedback average temperatures of each daisybed, DB and DW are the
interactions; thirdly it becomes plainer just how much can proportional coverage of each daisybed by black and white
be generalized from this one example to other domains. daisies respectively, as determined by the Witch’s Hat
function. The temperature T of each bed (taken as uniform
Organisation of Paper within the bed) is determined by a combination of factors:
external forcing by the sun, feedback (positive or negative)
In the following sections I shall start by describing the cut-
proportional to D (the coverage of daisies in that bed), and
down version of Daisyworld. Visualisations of the
a ‘leakage’ factor whereby some (parameterized)
conditions for steady state will be shown. Then results
proportion of the heat flows from the hotter to the cooler
obtained through computer simulations integrating the
daisybed. Unlike the original model, there is no direct
equations to a steady state will be given. The conditions
interaction between DB and DW.
and parameter values will be manipulated to see just how
far they can change whilst maintaining robustness.
I shall draw some very general conclusions, and to A Single Daisybed
demonstrate their generality apply them to a very different Initially we can simplify still further by just looking at the
domain of active perception. Here a simulated robot is behaviour of a single daisybed; consider one half only of
supplied with oriented light-sensors that display a similar figure 2. T is the temperature of a bed with albedo , S the
α

Hat function response to a light source. Feedback directly temperature of the Sun, and deep space is at zero
coupled to this response will change the orientation of the temperature. In a simplified, distorted version of physics,
light-sensor in (a random choice of) either positive or heat flow into the bed from the Sun is (1 - )(S - T), and
α

negative direction. Collectively the coupling between many out of the bed into deep space is (T – 0), i.e. T.
such individual light-sensors determines the global D is now the quantity of daisies (rather than growth rate),
orientation of the robot. The result is homeostasis in the which varies according to a Hat function of the local
sense that collectively the system acts so as to maintain, as temperature T: D=H(T). In the simplest version, where we
far as possible, the light-sensors oriented to the light so as assume there is feedback linearly proportional to D to raise
(or lower) the local temperature T, this feedback is uD: for towards A, rather than its default left limit of D. Clynes
black daisies u is a positive feedback to increase the (1969) put forward the notion of Rein Control, in
temperature, for white daisies u is negative. commenting that biological systems typically have (at least)
The rate of change of flowerbed temperature is two channels for sensing and regulating variables: one (or
dT more) in one direction from the norm, another in the other
=(1−α)(S −T) −T +u.H(T) direction. This notion has received relatively little currency,
dt
although it is taken up in recent work drawing ideas from
Equilibrium is when the rate of change is zero: Daisyworld theory and applying them in modified form to
0 = (1 − α ) S − (2 − α )T + u.H (T ) physiological control (Saunders et al. 1998). The rein
(2 − α )T − (1 − α ) S metaphor is appropriate as a rein can only pull, not push.
H (T ) =
u Hence for control in both directions we need a further
For fixed S , u , α , this is linear in T with zero value when feedback loop, as in the following extension to the
(1 − α ) S simulation; we need both reins.
T=
(2 − α )
Two Daisybeds
The line has slope ( 2 − α ) / u . The equilibrium points are
For the simulation, we assume 2 daisybeds whose bare
where this straight line crosses the Hat function.
ground is grey with albedo 0.5. One bed can support only
black daisies with a lower albedo (typical value used 0.0);
Amount of Daisies

the other can only support white daisies (typical albedo

B
C 1.0). For each bed, the average albedo depends on the
proportion of cover by black daisies (0 DB 1) or white
≤ ≤

daisies (0 DW 1). The consequent temperatures, assumed

≤ ≤

to be uniform across each bed, are TB and TW. These

temperatures are then potentially modified by heat transfer
A1 A D A2 Temperature between the beds from the hotter black one to the white.
This transfer is parameterised by a factor 0 L 1. When
≤ ≤

Figure 3. Slope meets Witch’s Hat function at A, B, and C.

L=0, no heat transfer takes place, but if L=1 then the beds
each have their temperature modified to the mid-
In figure 3 the heavy sloping line has a positive slope
temperature (TB + TW)/2; for intermediate values of L the
(black daisies for positive feedback) and crosses H(T) at A,
temperatures are scaled linearly between these two extreme
B, and C. A implies no daisies, B proves to be an unstable
cases.
equilibrium, but C is a stable non-zero equilibrium. For
different amounts of external forcing from changing
To find through computation any non-zero stable
luminosity of the sun, the sloping line shifts, in parallel
equilibrium point for fixed values of S and L , we initialise
fashion, along the temperature axis. There will be a stable
DB and DW to 0.5, and then iterate this loop:
equilibrium point C, with a positive quantity of daisies, for
any such line between the lighter sloping lines through A1 1. Calculate albedo for each bed from DB, DW .
and A2 in the figure; these indicate the limits for 2. Calculate TB and TW from S and these albedos.
intersecting the RHS of the Hat. This implies that the range 3. Adjust these temperatures by the between-bed
of viability that allows some daisies to survive extends all heat transfer, or ‘leakage’, parameterised by L .
the way from A1 to A2, rather than the limited range D to 4. Use the Hat function to calculate D'B = H (TB )
A2 available if there is no feedback. The slope of the line and D 'W = H (TW )
reflects the degree of feedback, with a vertical line
corresponding to u=0. The stronger the feedback, the larger 5. Adjust DB and DW a small proportion of the way
u is, the further away the slope is from vertical; and hence towards these new values D' B and D'W by:
the further away to the left the viability range is extended.
A1 lies at a distance u /(2 − α ) to the left of the central D ⇐ (1 − δ ) D + δD' for a small value of δ .
optimum temperature of the Witch’s Hat, if one takes the 6. Go back to 1.
maximum height of that hat to be scaled to 1.0. If the line δ should be chosen small enough to ensure that the values
has negative slope (white daisies for negative feedback) change smoothly over successive iterations of this loop, and
then the mirror image case holds, and the range of viability then the loop must be repeated sufficiently many times until
is extended out to the right instead. So regardless of the the changes in values at each iteration are vanishingly
sign of the feedback, the range of viability is extended. small. In practice it was found, for the range of parameters
used here, that δ = 0.0001 and 200000 iterations of the loop
made further changes in the variables invisible at the level
Rein Control of double precision floating point numbers.
In a thoroughly mixed metaphor, in figure 3 the line AC Results are shown in graphical form for various values of
can be thought of as a rein pulling the zone of viability the parameter L . In each case the resulting equilibria are
shown across the full range of external forcing by the sun,
as it varies from excessively cool to excessively hot. 160

140

Full Conductance 120

T_B
140 100
T_W
80 Avge Temp
Temperature

120
D_B
100 60
T_B D_W

80 T_W 40
Avge Temp
60 20
%age Daisies

D_B
D_W
40 0
50 60 70 80 90 100 110 120 130 140 150
20

0 Figure 6. Conductance L=0.5. Coexistence of both daisies

75 80 85 90 95 100 105 110 115 120 125
between 82 and 118.5
Intermediate conductance. Figure 6 shows the more
Figure 4. Conductance or Leakage L=1. The
general picture, where although the extension of black
superimposed lines for DB and DW indicate the Witch’s Hat
daisies left, white daisies right, is not as far as in figure 6,
function, with a viability zone between 90 and 110 on the
the range of coexistence of both daisies is greater. Note that
lower scale. Vertical axis indicates temperatures (TB, TW
at all times that either type of daisy is viable, the black
and average, here all superimposed) and also percentage
daisies are hotter than their optimal temperature of 1000,
of Daisies in each bed ((DB, DW, here also superimposed).
and the white daisies are cooler than this.
Maximum heat conductance. If L=1, the temperatures of
black and white beds are identical. Since they follow the
140
same Hat function, there is always the same number of
black daisies in one bed as white daisies in the other. Hence 120
the overall average effect is that of grey, albedo 0.5. In
100 T_B
other words, there is no net feedback, and (regardless of
T_W
how many daisies there are) the temperature is the same as 80
Avge Temp
if there were none. Maximum conductance means D_B
60
minimum, or zero, homeostasis. See figure 4, where the D_W
horizontal axis indicates the sun’s output, scaled according 40
to the corresponding temperature of a lifeless planet; here
20
the temperatures TB and TW are the same as this.
250 0
70 90 110 130

200
Figure 7. The daisies are given different, narrower, Hat
T_B functions: black’s from 115-120, white’s from 80-85. There
150 T_W is coexistence of both daisies between 93 and 112.
Avge Temp

100 D_B
D_W Moving the Hat functions. Figure 7 shows that there is
still homeostasis when the viability zones (Hat functions) of
50 the daisies are shifted relative to each other. Note from the
figures in the caption, black is here shifted towards the
0 hotter end, white towards the cooler end.
50 70 90 110 130 150 170 190 210

Figure 5. L=0. Daisybeds are independent, and only in the Moving to Active Perception
range 90-110 on horizontal axis do both daisies coexist.
We have seen above how the simple yet powerful
Minimum conductance. When L=0, the two daisybeds are combination of a Hat function with a feedback loop (either
completely unconnected, and behave as if they were positive or negative) produces Rein Control, and in the
separate planets each regulating itself; the black bed context of Daisyworld, homeostasis; the Hat function
extends its viability only towards lower sun temperatures, directly relates to the idea of a zone of viability, and Rein
the white bed only towards higher. See figure 5. Control tends to regulate a system to stay within it. The cut-
down Daisyworld has reproduced the basic homeostatic rotation of the robot, are calculated in simulation with
results of the usual version; though the simpler equations similar liberties and simplifications used for modeling the
used means that it does not reproduce the phenomenon physics as were used in modeling Daisyworld, whilst
whereby the average planetary temperature actually respecting the general principles. The end result is that,
decreases slightly as solar luminosity increases. despite the random parameterization of all the tentacles,
Such a powerful principle can be extended to other this robot efficiently performs phototaxis.
0.6
domains, and here it is demonstrated with active perception
in a simple simulated 2-dimensional robot. Despite the very 0.4
different domain, the underlying principles are identical.
D 0.2
Target Orientation

Radians
Robot Orientation
0
1 10000

-0.2

B A -0.4

C -0.6
Figure 9. Time runs horizontally across this graph from the
left, and the heavy line indicates the orientation of the nose
in radians. The lighter line indicates a light source passing
Figure 8.Circular agent can only rotate about its centre; in wide, sinusoidal fashion across the front of the robot.
orientation is indicated by nose C, currently facing East.
One tentacle is shown, with sensor at end A. As indicated in figure 9, the robot will immediately pick
up on a passing light source, and track it so accurately that
One sensory tentacle is shown in the plan view of figure thereafter the plots coincide. Further testing shows that this
8. For fans of Doctor Who, think in terms of a Dalek. The behaviour is exceptionally robust to changes in the allowed
tentacle rotates around the centre of the robot, and has a ranges for the randomly chosen parameters. The maximum
sensory angle of acceptance as indicated at the end A. If a angle of acceptance can be allowed to vary over 3 orders of
light source passes across this receptive field, the sensor magnitude, out as far as 3 radians (or nearly 1800) each
response is given by a Witch’s Hat function, with maximum side. The upper limit on torque parameters and spring
response when the tentacle points directly at the source. constants can be allowed to vary over more than 2 orders of
This response produces a torque D on the tentacle A, in this magnitude; phototaxis is still reliable.
case shown as left or counter-clockwise. D is counteracted 1.2
by a restraining spring B attached to the nose C. Two
parameters, specific to this tentacle, modulate the torque 1

response D and the spring modulus B.

0.8
Now consider 100 such tentacles, each with different
Sensor value

randomly chosen angles of acceptance, different directions 0.6 Data Points

and torque parameters for D, different spring constants for
B. As a light source passes in front of them, each tentacle 0.4
will respond independently. But collectively they will be
held together via the springs attached to the nose, some 0.2

pulling in one direction, some in the other. In the absence

0
of any light, all these tentacles will be drawn together over -4 -2 0 2 4
the nose, but in the presence of any stimulation they will Target offset to sensors: Viewwidth scale 1.0
rotate apart independently in different directions, restrained
only by the springs. The resultant balance of these spring Figure 10. Samples from all the left-moving sensors whilst
forces on the nose will rotate the robot as a whole in one the robot is performing phototaxis. Each data point
direction or the other, about its centre. indicates on the vertical scale the sensor response,
The translation from Daisyworld is direct as far as the calibrated to a maximum of 1.0, and on the horizontal
underlying equations go, even though conceptually it is scale the angular offset of target source to tentacle
quite a leap. Each tentacle corresponds to a daisy species; direction. Although different tentacles have different angles
on average half will provide feedback in one direction and of acceptance, this offset is here rescaled so as make all the
half the other. The Hat function on sensor response individual Hat functions coincide on this graph.
corresponds to daisy dependence on temperature. The Analysis. Although each tentacle can move independently,
springs correspond to the leakage or conductance of heat and each response to sensory stimulus is in a random
between daisybeds, and the resulting direction of the nose direction, their collective coupling means that almost all the
corresponds to the average planetary temperature. The tentacles will stay approximately oriented towards the light
dynamics of the motion of each tentacle, and consequent source nearly all the time. While unqualified teleological
language is just as inappropriate here as it is in Daisyworld, (uniform) temperature of the black bed, and likewise that of
we can carefully say: “Although this robotic system only the white bed. By plotting these separately for didactic
functions this way as a whole, through multiple feedbacks, purposes, it is easier to appreciate the importance of the
it can seem to a casual observer as if the tentacles are trying difference between these temperatures, something obscured
to maintain their sensory stimulation; just as in Daisyworld in much of the previous literature where typically only
it might seem as if the daisies are trying to regulate the average planetary temperatures have been displayed.
temperature so as to stay within the viability zone.” The original Daisyworld model has extra layers of
Figure 10 shows samples of sensory inputs from all the complexity on top of this cut-down version, so it is of
counterclockwise or left-moving sensors during a run. The interest to see what is common to both. Homeostatic
outline of the Witch’s Hat function is clearly visible, with regulation is already apparent in the simpler version, but
almost all the data points on the right-hand slope, which is the phenomenon whereby average planetary temperature
where Rein Control is acting for regulation in that can actually decrease slightly as solar luminosity increases
direction. The data from right-moving sensors is the mirror is only seen in the more complex version.
image, giving the second of the pair of metaphorical reins. The plots shown in the various figures indicate that the
black daisies are almost always living in a hotter climate
than their optimum temperature. They are mostly on the
Discussion right slope of the Witch’s Hat; and vice versa for white
daisies. Figure 10 shows the equivalent for the perception
Many people have been mystified as to how homeostatic
example. What if daisy evolution allowed either species to
regulation is achieved in the Daisyworld model. How could
modify their metabolism, and so ‘shift their viability zone’
such regulation have arisen, since surely it requires some
(subject to underlying physical constraints) in the direction
care in setting up the feedback structure and the
of the climate they actually experience? Then the black Hat
parameters? An evolutionary origin appears unrealistic.
function would shift to the right, and the white one to the
Indeed any ongoing system that includes biota and seems to
left, as shown in figure 7.
require global collaboration seems susceptible to
The principles shown here are very simple and of wide
exploitation by evolution of sub-groups towards cheating.
applicability. They do not require an evolutionary origin or
The cut-down version of Daisyworld presented here
explanation, but may be quite compatible with evolution.
makes several useful pedagogical points. The analysis of a
single daisybed (see figure 3) shows that extension of the
Acknowledgment This work was motivated and stimulated
range of viability in one direction arises from the simple
by the participants in the series of EPSRC-funded
interaction of a feedback of any sign (the sloping heavy
workshops Daisyworld and Beyond, organised by Tim
line) and a Hat function. The feedback need not be linear,
Lenton and myself in Sussex and Edinburgh, 2001-03.
though normally it should be monotonic; the Hat function
can be anything to suit your millinery tastes, although the
Witch’s Hat seems near-ideal from a mathematical stance. References
Simple feedbacks are universal in natural systems, and Hat
functions are also widespread; the zone of viability Clynes, M. 1969. Cybernetic implications of rein control in
associated with any homeostasis automatically implies a perceptual and conceptual organization. Ann. NY Acad.
Hat function, and so does the typical response of any active Sci. 156: 629-670
sensor. So no special design process needs to be postulated
for the basic phenomenon illustrated here with a single Lenton, T. M. 1998. Gaia and natural selection. Nature
daisybed. We should observe such systems everywhere. 394: 439-447
The idea of Rein Control (Clynes 1969) deserves wider
currency. This phenomenon, like a rein, can only ‘pull’ and Lovelock, J. E. 1972. Gaia as seen through the atmosphere.
not ‘push’. So for homeostatic regulation in both directions Atmos. Environ. 6: 579-580.
we need feedbacks in both directions. In the active
perception example above, the 100 simulated sensor Lovelock, J. E. 1983. Gaia as seen through the atmosphere.
feedbacks in random directions would average out at In: P. Westbroek and E. W. d. Jong. Biomineralization and
roughly half in each direction, so any system that includes Biological Metal Accumulation. Dordrecht: D. Reidel
Hat functions and many arbitrary feedback loops is likely to Publishing Company, 15-25.
incorporate both reins of Rein Control.
In cut-down Daisyworld, the interaction between black Saunders, P. T., Koeslag, J. H., and Wessels, J. A. 1998.
and white daisybeds is limited to ‘leakage’ and its role is Integral rein control in physiology. J. Th. Biol. 194: 163-
made clear. Too much coupling means the opposing 173.
homeostatic tendencies will nullify each other, whereas
being uncoupled would imply, in effect, separate planets. Watson, A. J. and Lovelock, J. E. 1983. Biological
So some intermediate loose coupling between the different homeostasis of the global environment: the parable of
systems is essential, but no further global organisation is Daisyworld. Tellus 35B: 284 -289.
needed. In this simple version we can observe directly the