A DEMAND MODEL WITH DEPARTURE TIME CHOICE

FOR WITHIN-DAY DYNAMIC TRAFFIC ASSIGNMENT

Giuseppe Bellei, Guido Gentile, Lorenzo Meschini, Natale Papola

Università degli studi di Roma “La Sapienza”
E-mail: guido.gentile@uniroma1.it

ABSTRACT

A within-day dynamic demand model is formulated, embodying, in addition to the classic

generation, distribution and modal split stages, an actual demand model taking into account
departure time choice. The work focuses on this last stage, represented through an extension
of the discrete choice framework to a continuous choice set. The dynamic multimodal supply
and equilibrium model based on implicit path enumeration, which have been developed in
previous work are outlined here, to define within-day dynamic elastic demand stochastic
multimodal equilibrium as a fixed point problem on users flows and transit line frequencies.
A MSA algorithm capable, in the case of Logit route choice models, of supplying equilibrium
flows and frequencies on real dimension networks, is presented, as well as the specific
procedures implementing the departure time choice and actual demand models. Finally, the
results obtained on a test network are presented and conclusions are drawn.

keywords: within-day dynamic traffic assignment, elastic demand, departure time choice

1 INTRODUCTION

The dynamic analysis of transportation networks gathered increasing attention through the
past years. The static models, traditionally used in this field, are in fact unable to represent
relevant phenomena, such as demand variations over time and temporary over saturation of
network elements. On the other hand, dynamic models are more complex than static ones: on
the supply side, they require to ensure temporal consistency, besides spatial consistency,
among system variables (for instance, see Cascetta, 2001); on the demand side, choice of the
time when to travel has to be modelled explicitly in order to achieve a correct representation
of users’ travel behaviour (Mahmassani and Chang, 1986; Van Vuren et al., 1998;
Mahmassani and Liu, 1999).
Different approaches to the latter problem can be found in the literature. One of them is to
regard departure time choice as a discrete choice among temporal intervals and use static
assignment to characterize the utility of each interval, as in Daly et al. (1990); models based
on this approach yield a rough representation of travel demand during day time as a sequence
of static equilibriums.

1
An alternative approach, relying on a more realistic representation of dynamic travel times,
can be found in De Palma et al. (1983) and Ben-Akiva et al. (1986) where a within day
dynamic stochastic equilibrium model and a doubly dynamic stochastic model, respectively,
are presented. Travel, departure time and path choices are addressed through a mixed
discrete/continuous nested Logit model and travel times are determined by means of a
deterministic queuing model; however, both models are applied only to single origin
destination pair idealized networks.
Another approach is based on the hypothesis that departure time and path choices are made
jointly, so that, for each origin-destination pair, a discrete choice set is defined whose finite
number of alternatives is equal to the number of departure intervals multiplied by the number
of paths (Arnott et al., 1990; Cascetta et al., 1992); models based on this approach require
explicit path enumeration and the introduction of a diachronic graph, as in Van der Zijpp and
Lindveld (2000), so they can be hardly applied to congested urban networks.
An extensive analysis of departure time choice for shopping trips is presented in Bhat (1998)
and Bhat and Steed (2002). In the first paper, a discrete choice model is proposed, able to
represent correlation among adjacent departure time periods. In the second paper, a
continuous-time model of departure time is proposed, accommodating time-varying
coefficients. Both papers however doesn’t investigate path choice, since travel times and cost
are assumed exogenously.
In a recent paper by Bellei, Gentile, Papola (2003) the within-day road Dynamic Traffic
Assignment (DTA) was regarded as a stochastic dynamic user equilibrium. A new fixed point
formulation of the problem in terms of time-continuous real valued temporal profiles of arc
flows and arc performances (travel times and generalized costs) was presented, where the
concept of network loading map, yielding arc flows for given demand flows consistently with
certain arc performances, was extended to the dynamic case, thus avoiding the introduction of
both the dynamic network loading as a sub-problem, and the explicit path enumeration. OD
flows temporal profiles were taken as given. An implicit path enumeration algorithm was thus
proposed for the Logit case.
Based on the above modelling framework, two more papers, namely Gentile, Meschini,
Papola (2002) and Gentile, Meschini, Papola (2003), focused on a new dynamic transit supply
model relying on a frequency based approach; the introduction of a diachronic graph is thus
avoided and implicit path enumeration is allowed. This model is able to represent both intra
and inter modal congestion effects (i.e. interaction among cars and bus flows). In order to
define a multimodal within-day DTA, a dynamic mode choice model was introduced in these
papers, while demand flows were assumed to be rigid with respect to any other choice
dimension.
In this paper a mixed discrete/continuous nested Logit dynamic demand model with five
choice levels is presented, where, besides the usual to travel or not to travel (generation),
destination, mode and path choices, the departure time choice is introduced. The model is
conceived to extend previous work to the most general case of elastic demand multimodal

2
within-day DTA. With reference to departure time choice, the proposed demand model adopts
a continuous approach, thus not requiring to enumerate explicitly the desired departure time
intervals. The resulting within-day DTA model is then capable of representing both supply
and demand dynamic phenomena concerning congested multimodal urban networks, and
leads to a fixed point formulation that can be solved by an efficient implicit path MSA
algorithm applicable to real networks.

2 THE CHOICE AND DEMAND MODELS

In modelling travel demand we follow the behavioural approach based on random utility
theory, where it is assumed that each user is a rational decision-maker who, when making his
travel choice: a) considers a positive, finite number of mutually exclusive travel alternatives
constituting his choice set J; b) associates to each travel alternative j of his choice set a
perceived utility, not known with certainty, and thus regarded by the analyst as a random
variable Uj; and c) selects the travel alternative that maximises his utility. With these
hypotheses the probability of alternative j is formally expressed as:

Pj = Prob[∩ k∈J εk ≥ Vj -Vk +εj ] , (1)

where Vj and εj are respectively the systematic utility and the random residual of the generic
alternative j. The expected value of the maximum perceived utility is called satisfaction:

W = E[max j∈J {Vj +εj }] (2)

As usual, we assume that it is possible to divide the choice process into a hierarchic sequence
of decisions; at each level the user has a specific choice set, dependent on the choices made at
upper levels. The utility associated to each alternative available at a given level is the sum of a
specific term and of the satisfaction that takes into account the alternatives available at the
lower levels. The structure of the demand model parallels the structure of the choice model in
that users’ flows given as input, or having made a choice at an upper level, are split at the
lower level accordingly to the corresponding choice model.
The multimodal network, where users departing during the time horizon [0, Θ] are
represented, is defined as a graph G(N, A) where N is the node set and A is the arc set. The
generic node is denoted as x, while origin and destinations of trips, belonging to C ⊆ N, are
denoted as o and d, respectively; the generic arc is denoted as a, TL(a) and HD(a) being,
respectively, its initial and final nodes. Modal subgraphs Gm(Nm, Am) are implicitly defined
associating to each arc a set Ma ⊆ M of transport modes. Each travel alternative is
biunivocally associated with an acyclic path k of graph G from o to d on mode m, belonging
to an efficient path set Kmod, defined shortly later.
The proposed demand model is a Nested Logit with five choice levels, specifically:
generation, distribution, modal split, departure time choice, and path choice. It thus follows,

3
apart from the departure time choice, the Oppenheim (1995) approach to travel demand
modelling. Such an approach is extended to a dynamic equilibrium framework as follows:
i) the demand model takes as input the time rate N o(σ΄) ≥ 0, σ΄∈[0, Θ], of potential users
having σ΄ as their desired departure time, for each origin o;
ii) the travel behaviour, with regard to the choices of travelling, of destination d and of
mode m, is represented by models having, for each desired departure time, the same
functional form and parameters;
iii) the departure time choice is represented by a model supplying actual departure time
probabilities in the interval B(σ΄) = [max{σ΄ -ADV, 0} , min{σ΄ + DEL, Θ}], in the
form of a probability density function of actual departure time τ΄ for each o, d, m
and σ΄, where ADV and DEL are respectively the maximum delay and advance time
admitted by users;
iv) the travel behaviour, with regard to path choice, is represented by mode specific models
having, for each o, d, and τ΄, the same functional form and parameters.
It is worth noting that the departure time choice model is defined over an infinite number of
alternatives belonging to a continuous choice set in ℜ 1, while the overall structure of the
choice model is that of a discrete choice model. We will deal with this issue in section 2.3.
Assumptions ii) and iii), are consistent with model specification, but could be easily relaxed if
needed, for example allowing time varying parameters to better fit experimental data as
suggested in Bhat (2002). On the contrary, dealing with path choice model, we need
hypothesis iv) in order to adopt an implicit path enumeration approach where path
probabilities must be consistent with the arc conditional probabilities of leaving a node x =
TL(a) through arc a. This is because on a same arc, at a given instant, we can have users
coming from different origins and departing at different instants.
The choice model is of the type depicted in Figure 1. The hierarchical order proposed here
keeps a widely accepted structure and it is consistent, with reference to the position of
departure time choice, with previous experimental analysis both in De Palma et al. (1983) and
in Bhat (1998).

[Figure 1 here]

The following part of this section is organized as follows. In subsection 2.1 further definitions
and notations needed to place the demand model in a dynamic framework will be introduced.
Then the choice models related to each level will be examined in ascending order, since lower
level satisfactions are needed when an upper level model is defined. Subsection 2.2 and 2.3
will be specifically devoted respectively to implicit path choice and departure time choice,
while we will deal with mode, destination and travelling choice together in subsection 2.4.
The formal definition of the corresponding demand models, determining demand temporal
profiles, will follow in the inverse order, since upper level profiles are input to lower level
models. As for choice models, in subsection 2.5 we will deal with desired demand profiles

4
together, defined for each desired departure time and determined by generation, distribution
and modal split models; sections 2.6 and 2.7 will be specifically devoted to actual demand
model and to network flow propagation model, supplying actual demand profiles, defined for
each actual departure time, at mode - od pair level and at arc level, respectively. The arc flow
obtained are consistent with the dynamic loading to the network of implicit path flows, taking
the arc performances as given.
We will denote by g(σ) or g(τ) the generic temporal profile referred to desired or actual
departure time, respectively; then, g(σ΄) or g(τ΄) denote the value of the temporal profile at
given instants. When compact form is utilized, vectors of temporal profiles, having the
appropriate dimension, are identified by the use of bold symbols.

2.1 Definitions and notations

As already said, it is assumed that, when travelling from node x to destination d on mode m,
users consider only the subset Kmxd of efficient paths from x to d, which are defined on the
basis of node topological order (Nguyen, Pallottino and Inaudi, 1996). With reference to a
given destination d and mode m, we thus assume that the node topological order TOmd(x) is
monotone non-decreasing with some measure, independent of congestion and time, of the
distance on graph G from x to d. An arc is efficient if its tail has a higher topological order
than its head, while a path is efficient if all its arcs are efficient. The sets of the efficient arcs
exiting and entering a given node x are referred to, respectively, as the efficient forward star
FSE(x)md = {a∈A: TL(a) = x, TOmd(x) > TOmd(HD(a))} and the efficient backward star
BSE(x)md = {a∈A: HD(a) = x, TOmd(TL(a)) > TOmd(x)}.

At any time τ΄, each mode m is characterized by mode - specific arc entering flows and
performances:
fam(τ΄) mode m users flows entering at τ΄ into arc a,
ca (τ΄)
m
mode m generalized costs on arc a, for users entering at τ΄.
tam(τ΄) mode m exit time from arc a, when entering at τ΄.
If function tam(τ΄) is monotone increasing, its inverse, which also results to be monotone
increasing, can be defined as:
-1
tam (τ΄) entering time into arc a when exiting at τ΄.
The generalized arc cost for users entering at time τ is thus simply assumed to be:
cam(τ΄) = η⋅(tam (τ΄) -τ΄) +mcam(τ΄), (3)
where mca is the temporal profile of the monetary cost, while η is the Value of Time.
m

These variables, as all the arc and node variables defined in the following, are defined on a
time horizon [0, Θ + ∆Θ], ∆Θ being the time needed to complete all trips started at Θ.
Path performances are defined as a function of time consistently with arc performances by
utilizing the following arc set notation:
Akxdm set of the arcs constituting path k∈Kmxd;

5
Akxdma set of the arcs constituting the sub-path of path k∈Kmxd between node x and the tail
of arc a∈Akxdm;
together with the following path notation:
Ckxdm(τ΄) generalized cost of path k∈Kmxd for users leaving node x at time τ΄;
Tkxdm(τ΄) time when users following path k∈Kmxd and leaving node x at time τ΄ reach
destination d;
Tkxdma (τ΄) time when users following path k∈Kmxd and leaving node x at time τ΄ enter arc
a∈Akxdm;

For each path k∈Kmxd and arc a∈Akxdm, the travel time of the sub-path between node x and
node TL(a) at time τ΄ is the sum of the travel times of its arcs b∈Akxdma , each of them referred
to the time Tkxdmb(τ΄) when users leaving x at τ΄ reach TL(b); that is:

Tkxdma(τ΄) = τ΄ + ∑b∈Akxdma[tbm(Tkxdmb(τ΄)) - Tkxdmb(τ΄)]. (4)

Assuming additive costs, the generalized cost of path k∈Kmxd at time τ΄ is the sum of the costs
of its arcs a∈Akxdm, each of them referred to the time Tkxdma(τ΄) when users leaving node x at τ
reach node TL(a); that is:

Ckxdm(τ΄) = ∑ a∈Akxdm ca(Tkxdma(τ΄)). (5)

2.2 Implicit path choice model

The implicit path choice model presented in this paper is the same introduced and described in
Bellei, Gentile, Papola (2003), and is founded on the concepts of arc conditional probability
and node satisfaction, which are respectively defined as follows:
padm(τ΄) probability that users on mode m follow their trip to destination d with arc a,
conditional on being at node TL(a) at time τ΄;
wxdm(τ΄) expected value of the maximum perceived utility among all paths to d on mode m
departing from x at time τ΄.
It is worth noting that, if demand temporal profiles are defined, as in this case, with reference
to the departure time from the origin, the implicit path choice model, as well as the network
flow propagation model defined in section 2.7, have necessarily to be formalized with respect
to destination. A formalization with respect to the origins is required, in fact, when demand
temporal profiles are defined with respect to the arrival time to the destination, which in this
case are unknown and are actually obtained from the network flow propagation model.
Nevertheless, the node satisfaction temporal profiles yield path choice satisfactions for each
o-d pair and mode m as the node satisfaction wodm(τ΄) at o, while the choice probability of the
generic path k∈Kmod from o to d on mode m at time τ΄ is equal to the product of the
conditional probabilities of its arcs a∈Akodm, each of them referred to the time Tkodma(τ) when
users leaving o at τ΄ reach TL(a); that is:

6
Pkodm(τ΄) = ∏ a∈Akodm padm(Tkodma(τ΄)) (6)
With reference to the Logit case, node satisfaction can be expressed in a recursive form and
used to calculate arc conditional probabilities:

wdmd(τ΄) = 0

wxmd(τ΄) = θR ⋅ln(∑a∈FSEmd(x) exp((-cam(τ΄) +wHD(a)md(tam(τ΄))) /θR)) ; (7)

pamd(τ΄) = exp((-cam(τ΄) +wHD(a)md(tam(τ΄)) -wTL(a)md(τ΄)) /θ R) ; (8)

which, in compact form, can be formally expressed through the functionals:

w(τ) = w(c(τ), t(τ)) ; (9)

p(τ) = p(w(τ), t(τ), c(τ)) . (10)
Moreover, it can be proven (Bellei, Gentile, Papola, 2003) that path probabilities defined by
(6) are the same as the path probabilities derived by an explicit path choice model based on
path costs (5):

 Ckmod (τ ' ) 
exp  − 
 θR 
Pk (τ ' ) =
mod
. (11)
 C mod (τ ' ) 
∑ exp  − j
 θ R 

j∈K od
m 

2.3 Departure time choice model

With reference to a desired departure time σ΄, we assume, extending, to the continuous choice
set B(σ΄) the standard assumption of the Logit model, that the random residuals associated to
the infinitesimal alternatives constituting a partition of B(σ΄) are independently and identically
distributed (i.i.d.) Gumbel variables as in De Palma et al. (1983).
Any consideration concerning correlation among alternatives is left to further investigations.
We choose on purpose a simple and well known choice probability model in order to focus on
our main objective, that is, to provide a modelling framework for the simulation of elastic
demand in the context of within-day DTA. In any case, as stated by De Palma et al. (1983),
there are experimental evidences that the logit model can serve as a reasonable model of
departure time behaviour.
Although we refer to users travelling from origin o to destination d by mode m, the
corresponding indices will be dropped in order to improve readability. Let’s then define:
p(τ΄/σ΄) probability density of leaving at time τ΄, conditional to desired departure time σ΄;
V(τ΄,σ΄) specific utility of leaving at time τ΄, when σ΄is the desired departure time;
w(τ΄) path choice satisfaction, given, for each o-d pair and mode m, by (7)
The Logit formula can be used to calculate the probability of choosing the generic departure
interval [τ΄-dτ /2, τ΄+dτ /2] as follows:

7
  V (τ ', σ ' ) + w (τ ') 
exp  
 θ DT 
p (τ '/ σ ' ) ⋅ dτ = σ ' + DEL ⋅ dτ , (12)
 V ( x, σ ' ) + w ( x ) 
∫ exp 
σ ' − ADV
θ DT
 ⋅ dx

where, clearly, the summation is replaced by an integration. On the analogy of the discrete
case, the denominator in equation (12) is directly related to the departure time satisfaction:

 σ '+ DEL  V ( x, σ ' ) + w ( x )  

W (σ ') = θ DT ⋅ ln  ∫ exp   ⋅ dx  (13)
σ '− ADV  θ DT  
Finally we obtain, by assuming that the specific utility is proportional to the departure
advance or delay:

V (τ ′,σ ′) = − max {bADV ⋅ (σ ′ − τ ′) ,bDEL ⋅ (τ ′ − σ ′)} , (14)

an explicit expression of the actual departure time τ΄ probability density function, conditional
to desired departure time σ΄ and dependent on path choice satisfaction at τ΄:

 w (τ ′ ) − max {bADV ⋅ (σ ′ − τ ′ ) , bDEL ⋅ (τ ′ − σ ′ )} 

exp  
 θ DT
p (τ '/ σ ') = σ ′+ DEL   (15)
 w ( x ) − max {bADV ⋅ (σ ′ − x ) , bDEL ⋅ ( x − σ ′ )} 
∫ exp 
 θ
 ⋅ dx

′−
σ ADV  DT 
The departure time choice model is expressed in compact form through the functionals:

pDT(τ /σ) = p(V(τ, σ), w(τ)) (16)

WDT(σ) = W(V(τ, σ), w(τ)) (17)

2.4 Mode, destination and travelling choice models

Referring to users travelling from origin o toward destination d and to a desired departure
time σ΄, let’s define:
Pmod(σ΄) choice probability of modal alternative m;
Vmod(σ΄) specific utility of modal alternative m;
Wmod(σ΄) departure time choice satisfaction, given by (13).
In the Logit case, the mode choice probabilities are:

 Vmod (σ ') + W od
m (σ ' ) 
exp  
 θM 
Pm (σ ') =
od
, (18)
 Vm ' (σ ') + W m ' (σ ') 
od od

∑ exp  
m '∈M
 θM 
in compact form: PM(σ) = P(VM(σ), WDT(σ))
and the resulting mode choice satisfaction is:

8
   V od (σ ') + W od
m (σ ' ) 

W od (σ ') = θ M ⋅ ln  ∑ exp  m  (19)
 m∈M  θM  
in compact form: WM(σ) = W(VM(σ), WDT(σ))
We assume Vmod(σ΄) = βM T⋅Xmod(σ΄), where Xmod(σ΄) is a vector of time-dependent mode m
attributes for users travelling from o to d and βM is the corresponding vector of coefficients.
Referring to users departing from origin o and to a desired departure time σ΄, let’s define:
Pd o(σ΄) choice probability of destination alternative d;
V (σ΄)
od
specific utility of destination alternative d,
while the mode choice satisfaction Wod(σ΄) is given by (19).
In the Logit case, the destination choice probabilities are::

 V od (σ ') + W od (σ ') 
exp  
 θD 
Pd (σ ') =
o
, (20)
 V (σ ' ) + W (σ ' ) 
od ' od

∑ exp  
d '∈C
 θD 
in compact form: PD(σ) = P(VD(σ), WM(σ))
and the resulting destination choice satisfaction is:

  V o (τ ') + W od (τ ')  
W o (τ ') = θ D ⋅ ln  ∑ exp  d  (21)
 d∈C  θD  
in compact form: WD(σ) = W(VD(σ), WM(σ))
We assume: Vod(σ΄) = βDT⋅Xod(σ΄), where Xod(σ΄) is a vector of time-dependent destination d
attributes for users travelling from o and βD is the corresponding vector of coefficients.
Referring to users potentially departing from origin o and to a desired departure time σ΄, let’s
define:
P o(σ΄) choice probability of travelling
V (σ΄)
o
specific utility of travelling
while the destination choice satisfaction Wo(σ΄) is given by (21).
In the Logit case, the travelling choice probabilities are:

 V o (σ ' ) + W o (σ ' ) 
exp  
 θE  ,
P (σ ' ) =
o
(22)
 V ( σ ' ) + W (σ ' ) 
o o
1 + exp  
 θE 
in compact form: PE(σ) = P(VE(σ),WD(σ)),
where utility of not travelling is set to 0 because the model is additive and the resulting
travelling choice satisfaction, which is the perceived utility of potential users is:

9
   V o ( σ ' ) + W o (σ ' )  
S o (σ ') = θ E ⋅ ln 1 + exp   , (23)
  θ E  
in compact form: S(σ) = S(VE(σ), WD(σ)).
We assume Vo(σ΄) = βET⋅Xo(σ΄), where Xo(σ΄) is a vector of time-dependent choice of
travelling attributes for potential users in o and βE is the corresponding vector of coefficients.

2.5 Desired demand models

We remember that the flow N o(σ΄) of users potentially travelling from origin o at time σ΄ is
assumed to be known. Let’s denote the desired demand flows by:
q o(σ΄) flow from origin o with desired departure time σ΄
q (σ΄)
od
flow from origin o to destination d with desired departure time σ΄
q mod(σ΄) flow from origin o to destination d on mode m with desired departure time σ΄
Thus, the generation model is expressed by:

q o(σ΄) = N o(σ΄)⋅P o(σ΄) , (24)

where P o(τ΄) has the expression (22), the distribution model by:

q od(σ΄) = q o(σ΄)⋅Pd o(σ΄) , (25)

where P od(σ΄) has the expression (20), and the modal split model by:

q mod(σ΄) = q od(σ΄)⋅Pmod(σ΄) (26)

where Pmod(σ΄) has the expression (18).
On the basis of relations (24), (25) and (26), it is possible to express the three desired demand
models together as q mod(σ΄) = N o(σ΄)⋅P o(σ΄)⋅P od(σ΄)⋅Pmod(σ΄), or, in compact form, through
the functional:

q(σ) = q(N(σ), PE(σ), PD(σ), PM(σ)) (27)

2.6 Actual demand model

The role of the actual demand model is to transform the desired demand temporal profile of
any o-d pair and mode m into the corresponding actual demand temporal profile, on the basis
of the departure time probabilities produced by the departure time choice model. As in section
2.3 the o, d and m indices will be dropped. The desired demand profile provided by (24), (25)
and (26) is thus denoted as q(σ).
To transform q(σ) into the actual demand temporal profile d(τ) the contributes coming from
departure time choices corresponding to different desired departure times have to be
considered. The number of trips started within the infinitesimal departure interval [τ΄-dτ /2,
τ΄+dτ /2] is thus given by the integral, for each σ΄: τ΄∈B(σ΄), of the flow q(σ΄) of trips with
desired departure time within the infinitesimal interval [σ΄-dσ /2, σ΄+dσ /2], multiplied by the
probability p(τ΄/σ΄) that their actual departure time is in [τ΄-dτ /2, τ΄+dτ /2]; that is:

10
 τ ′+ ADV
d (τ ′) ⋅ dτ = ∫ q (σ ) ⋅ dσ ⋅ p (τ ′ / σ ) ⋅ dτ (28)
τ ′− DEL

On the basis of (15) and (28) the actual demand profile may be expressed as:

 w(τ ′) − max{bADV ⋅ (σ −τ ′) , bDEL ⋅ (τ ′ − σ )} 

exp  
τ ′+ ADV  θ 
d (τ ′) = ∫ q (σ ) ⋅  DT  ⋅ dσ (29)
τ ′−DEL
σ +DEL  w( x ) − max{bADV ⋅ (σ − x) , bDEL ⋅ (τ − x )} 
∫ exp
 θ
 ⋅ dx

σ − ADV  DT 
As in general the profiles q(σ), and w(τ) can assume any form, the integrals in equations (15)
and (29) cannot be solved in closed form. An ad-hoc procedure, based on the assumption that
the profile q(σ) is piece-wise constant, while the profile w(τ) is piece-wise linear, is presented
in section 4.2.
The actual demand model is thus expressed in compact form through the functional:

d(τ) = d(q(σ), pDT(τ/σ)) = d(q(σ), p(V(τ, σ), w(τ))) (30)

in fact, when solving the integral in (29) σ΄ disappears, while profile q(σ) does not.

2.7 Network flow propagation model

The path flows could be easily derived from the actual demand profiles d(τ), provided by (29)
for each o, d and m, and by the path choice probabilities (5). This however would lead, to
ensure consistency with travel times when they are loaded to the network, to the formulation
of Dynamic Network Loading (DNL) problem. As shown in Bellei, Gentile, Papola (2003),
an implicit path enumeration approach allows to define an equivalent fixed point DTA
formulation, not requiring DNL formulation, which is applied to the problem at hand in the
following section.
Within this approach, an arc network loading consistent with travel times, actual demand
profiles and arc conditional probabilities is obtained, trough what we call a network flow
propagation model, since it describes how arc flow profiles by mode and destination
propagate from upstream to downstream arcs:
-1 -1
famd(τ΄) = pamd(τ΄)⋅[dmTL(a)d(τ΄) + ∑ b∈BSEmd(TL(a)) [ fbmd(tbm (τ΄))⋅∂tbm (τ΄) /∂τ]] (31)
where dmTL(a)d(τ΄) = 0 ∀τ΄∈[0, Θ] if TL(a)∉C. Arc flows by mode are then obtained by simply
summing over destinations:
fam(τ΄) = ∑ d∈C famd(τ΄) (32)
The network flow propagation model is expressed in compact form through the functional:

f (τ) = ω(d(τ), p(τ), t(τ)) (33)

11
3 SUPPLY AND EQUILIBRIUM MODELS

In analogy with the static case, the within-day DTA, regarded as a dynamic user equilibrium,
can be consistently formalized through a fixed point problem expressed in terms of the
temporal profiles of arc flows and transit frequencies, by combining the arc performance
function with the Network Loading Map (NLM) and thus avoiding to introduce the DNL. To
this purpose, the NLM is here extended to the case of elastic demand with departure time
choice, as depicted in Figure 2.

[Figure 2 here]

In the following subsections, the arc performance function and the equilibrium model will be
briefly described only in a qualitative manner, providing references where this matters are
widely investigated and discussed.

3.1 Arc performance models

The concept of equivalent flow va(τ) is introduced first, in order to represent the congestion
phenomena considered. With reference to the arcs of the transit network, the equivalent flows
coincide with the user flows, while, with reference to each arc a of the road network, we
assume that the equivalent flow is given by a linear combination of flows for modes m∈Ma,
including transit vehicles. Then, equivalent flows are expressed in compact form as:

v(τ) = v( f(τ), φ(τ)) (34)

With reference to the road network, the dynamic arc performance function is specified
through the macroscopic and non-stationary link-node model presented in Bellei, Gentile,
Papola (2003), where the vehicular flow, considered as a monodimensional partially
compressible fluid, behave accordingly with the simplified kinematic wave theory and with a
triangular shaped fundamental diagram. The arc is modelled by considering two serial phases:
a running phase, modelled through a bottleneck, which simulates the link travel time
including the delay due to the over-saturated queue; a waiting phase, which represents the
average delay due to the under-saturated queue proper to an intermittent service (such a
traffic light).
With reference to the transit network, the dynamic performance function is specified through
the transit supply model presented in Gentile, Meschini, Papola (2002) and (2003). With
reference to the generic line l, the characterizing element of this model is the non separable
stop model, depicted in Figure 3.

[Figure 3 here]

12
The line arc represents both the dwelling time at the stop corresponding to its tail, and the
running time between two successive stops. The dwelling time is dependent on the flows that
are boarding, alighting, and on board the line at the corresponding stop; the running time is
dependent on the congestion level along the road arcs employed by the line.
The waiting arc represents the average waiting time for the transit vehicle at the stop, and it is
assumed to be proportional to the inverse of the line frequency.
The queue arc, modelled through a bottleneck with variable outgoing capacity, represents the
additional travel time due to an over saturation of the line, which occurs whenever boarding
and onboard flows overwhelm the line capacity.
The stop and waiting arcs, both with no travel time and cost allow expressing the non
separable dynamic stop model as a function of arc entering flow temporal profiles only.
The stop access arc allows distinguishing between different stop nodes corresponding to the
same walking node.
The alighting and walking arcs are assumed uncongested.
On the basis of the above outlined models, congestion phenomena are represented assuming
that, in general, the temporal profiles of exit times depend on the temporal profiles of both the
arc equivalent flows and the line frequencies:

t(τ) = t(v(τ), φ(τ)) (35)

The arc cost is assumed to be the sum of a term proportional to the travel time, and a term
taking into account the monetary cost. Then, the temporal profiles of arc cost are expressed in
compact form by the functional:

c(τ) = c(v(τ), φ(τ)) (36)

In order to complete the dynamic representation of the supply we need to define a line model,
which, for given transit frequencies at terminals, yields the frequency pattern along each line.
Contrary to the static case in fact, the line frequency temporal profiles are generally not
constant along the line: on one side, there is a translation of the terminal frequency in space
and time; on the other side, the variation in time of the line travel times makes vehicles spread
and jam (see for example Figure 4).

[Figure 4 here]

As explained in Gentile, Meschini, Papola (2002) and (2003), an effective representation of

this phenomenon can be achieved regarding transit frequencies as flows on the network. In
this way, in analogy with the network flow propagation, the temporal profiles of the
frequency at each stop of the generic line l is expressed as:

φa(τ΄) = φFA(l)(Tal -1(τ΄)) ⋅(∂Tal -1(τ΄) /∂τ) (37)

where: φFA(l)(τ) is the temporal profile of the line frequency at terminal (assumed given);
Tal(τ΄) express the time when a vehicle running line l and departed from terminal at time τ΄

13
reaches TL(a); Tal -1(τ) is the inverse of the temporal profile just defined, and express the
departure time from the terminal for a vehicle running line l and reaching TL(a). Expressing
profile Tal(τ) as a function of arc exit time profiles through equation (4), we have in compact
form:

φ(τ) = φ(t(τ)) (38)

3.2 Formulation of the dynamic assignment model

Relations (35), (36), and (38) determine a circular dependence between transit frequencies
and arc performances, in addition to the classical one between user flows and arc
performances. Then, the DTA is formulated as a fixed-point problem expressed in terms of
the temporal profiles of arc inflows and transit frequencies on the time horizon [0, Θ + ∆Θ].
Formally, combining (9), (10), (16), (17), (18), (19), (20), (21), (22), (27), (30), (33), (34),
(35), (36) and (38) it is:

[ f(τ), φ(τ)] = Φ[ f(τ), φ(τ)] , (39)

where the functional relation denoted by (39), whose expression is quite complex, is
explained through the flow-diagram depicted in Figure 5.

[Figure 5 here]

To be notice that, from the point of view of fixed point definition, σ is an auxiliary variable,
since it just serves to take into account the effects on the demand of the difference between
actual and desired departure times.

4 ALGORITHM

In order to implement the proposed DTA model, the period of analysis is divided into I time
intervals identified by the sequence of instants (τ 0, … , τ i, … ,τ I ). In the following we
assume to approximate the generic temporal profile x through either a piece-wise constant or a
piece-wise linear function defined by the values taken at such instants, so that for the two
cases we have respectively:

x(τ 0) = x 0 , x(τ) = x i , τ ∈(τ i-1, τ i] , i = 1, … , I (40a)

x(τ 0) = x0 , x(τ) = xi-1 + (τ -τ i-1) ⋅(x i -x i-1) /(τ i -τ i-1) , τ ∈(τ i-1, τ i] , i = 1, … , I (40b)
Specifically, the temporal profiles of the flows are assumed piece-wise constant, while the
temporal profiles of performances, satisfactions and choice probabilities are assumed to be
piece-wise linear.
With these assumptions, it is possible to devise an efficient elastic demand DTA solution
algorithm, outlined in the first section, that requires only some more trivial calculus than a

14
rigid demand DTA with regard to the inclusion of generation, distribution and modal split
models. Since taking into account the choice of departure time is somewhat more demanding,
the corresponding procedures for the calculation of choice probabilities and of actual demand
are illustrated in more detail in the second section.

4.1 Elastic Demand Dynamic Traffic Assignment

The resulting fixed point problem formalizing the elastic demand DTA is solved through the
Method of Successive Averages (MSA). It can be thus defined by the following macro steps:
0) k = 0 , f k+1 = {0 | f iniz} initialization
1) k = k +1 updates the iteration counter
2) v k = v( f k, φ k) calculates equivalent flows
3) t k = t(v k, φ k) , c k = c(v k, φ k) calculates the arc performances
4) w k = w(c k, t k) , p k = p(w k, c k, t k) calculates arc conditional probabilities
5) PDT k = P(VDT, w k) , WDT k = W(VDT, w k) calculates departure time probabilities
6) PM k = P(VM, WDT k) , WM k = W(VM, WDT k) calculates modal choice probabilities
7) PD k = P(VD, WM k) , WD k = W(VD, WM k) calculates distribution probabilities
8) PE k = P(VE, WD k) , S k = S(VE, WD k) calculates generation probabilities
9) q k = q(N, PG k, PD k, PM k) calculates the desired demand flows
10) d k = d(q k, WDT k, w k) calculates the actual demand flows
11) fNLM k = ω( p k, t k, d k) performs network flow propagation
12) φNLM k = φ(t k) performs line frequency propagation
k+1 k k k
13) f = f +1/k ⋅( fNLM – f ) updates the arc flows
14) φ k+1
= φ +1/k ⋅(φNLM – φ )
k k k
updates the line frequencies
15) if max a∈A i∈I |ya -fa | > ε and k < kmax then goto 2
ik ik
stop criterion

Steps 3), 4), 11) and 12) present algorithmic procedures quite specific for the dynamic
context, and are detailed in Bellei, Gentile, Papola (2003) and in Gentile, Meschini, Papola
(2003).

4.2 Departure time choice and actual demand

Since we approximate flow temporal profiles through piece-wise constant functions over
predefined time intervals, we are only concerned in determining the time interval in which
each user actually depart, and not the exact user departure time. Hence, we only need to
evaluate the probability P((τ i-1, τ i]/σ΄) of departing during interval (τ i-1, τ i], when σ΄ is the
desired departure time. This can be done integrating equation (15) over the interval (τ i-1, τ i]:

15
  w (τ ) − max {bADV ⋅ (σ ′ − τ ) , bDEL ⋅ (τ − σ ′)} 
exp  
τ i
 θ 
( )
P (τ i −1 ,τ i  σ ' = ∫ ′+
i −1
σ DEL

 w ( x ) − max {b ⋅ (
DT

σ ′ − x ) , b ⋅ ( x − σ ′ )}


⋅ dτ (41)

∫
τ ADV DEL
exp   ⋅ dx
 θ 
σ ADV
′−  DT 
Since profile w(τ) is expressed through relation (40b), the arguments of the integrals in
equation (41) are piece-wise linear. The points Pi j = P((τ i-1, τ i]/σ j) and W j, i = 1, …, I,
j = 1, …, I, which respectively define, coherently with the relation (40b), profiles P((τ i-1, τ
i
]/σ) and W(σ), can be calculated through the following procedure:

Function departure_time_choice
A = 0; D = 0
For j = 1 To I
1) For i = 1 To I
pi j = 0
Next i
2) Do While τ A ≤ σ j – ADV And A < I
A=A+1
Loop
3) Do While τ D < σ j + DEL And D < I
D=D+1
Loop
 wi − wi −1 
4) α =  wi −1 − τ i −1 ⋅ − σ A ⋅ bADV  θ DT
 τ −τ
i i −1

 wi − wi −1 
β =  i i −1 + bADV  θ DT
 τ −τ 

(
PA j = exp (α ) β ⋅ exp (τ A ⋅ β ) − exp (σ j − ADV ) ⋅ β 
  )
j j
W = pADV(j)
5) For i = A + 1 To j
 wi − wi −1 
α =  wi −1 − τ i −1 ⋅ − σ j ⋅ bADV  θ DT
 τ −τ
i i −1

Pi j = exp (α ) β ⋅ exp (τ i ⋅ β ) − exp (τ i −1 ⋅ β ) 
W j = W j + pi j
Next i
6) For i = j + 1 To D – 1
 wi − wi −1 
β =  i i −1 − bDEL  θ DT
 τ −τ 

16
 Pi j = exp (α ) β ⋅ exp (τ i ⋅ β ) − exp (τ i −1 ⋅ β ) 
W j = W j + pi j
Next i
 wi − wi −1 
7) α =  wi −1 − τ i −1 ⋅ + σ D ⋅ bDEL  θ DT
 τ −τ
i i −1


( )
PD j = exp (α ) β ⋅ exp (σ j + DEL ) ⋅ β − exp (τ D −1 ⋅ β ) 
 
j j i
W = W + pD

8) For i = ADV(j) + 1 To DEL(j) – 1

pi j = pi j / W j
Next i
W j = θDT ⋅ln(W j)
Next j

For each desired departure time σ j, step 1) resets the choice probabilities; steps 2) and 3) find
τ A andτ D, which are, respectively, the first predefined instants after σ j–ADV and σ j+DEL;
steps 4) through 7) evaluate the numerator of equation (41) for each interval (τ i-1, τ i], i = A,
…, D; steps 8) finally calculate the probability for each interval (τ i-1, τ i], i = A, …, D.
The number D i of users actually departed during the generic interval (τ i-1,τ i] is then given by
the integral, for each instant σ΄ such that σ΄∈[τ i-1–DEL, τ i+ADV], of the desired demand flow
q(σ΄) multiplied by the related probability P((τ i-1, τ i]/σ΄):
τ i + ADV
D = i

i −1
∫ ( )
q (σ ) ⋅ P (τ i −1 ,τ i  σ ⋅ dσ (42)
τ − DEL

Then, coherently with the relation (40a), profile d(τ) is simply given by:

d(τ΄) = d i = D i /(τ i – τ i-1) , τ΄∈(τ i-1, τ i] , i = 1, …, I (43)

Since profile P((τ i-1, τ i]/σ) is expressed through relations (40b), the points d i can be
calculated through the following procedure:

Function actual_demand
A = 0; D = 0
For i = 1 To I
1) Do While σ D ≤ τ j – DEL And D < I
D=D+1
Loop
2) Do While σ A < τ j + ADV And A < I
A=A+1

17
 Loop
3) (
d i = q D ⋅ 0,5 ⋅ Pi D ⋅ τ D − (τ i − DEL ) ) (τ i
− τ i −1 )
4) For j = D + 1 To A – 1
d i = d i + q j ⋅ 0,5 ⋅ ( Pi j + Pi j −1 ) ⋅ (σ j − σ j −1
) (τ i
− τ i −1 )
Next j
5) (
d i = d i + q A ⋅ 0,5 ⋅ PjA−1 ⋅ (τ i + ADV ) − τ A−1 ) (τ i
− τ i −1 )
Next i

For each actual departure interval (τ i-1, τ i], steps 1) and 2) find σ D and σ A, which are
respectively the first periodization instant after τ i-1–DEL and τ i+ADV; steps 3) to 5) evaluate
equation (42) for each interval (σ j-1, σ j], j = D, …, A.

5 NUMERICAL APPLICATION

The network of Sioux Falls, consisting of 76 directed arcs and 24 centroids, has been
considered for a numerical application. In order to investigate the effectiveness of the
proposed departure time choice model and algorithm, we compare the results of two DTA on
this network, performed considering both rigid and elastic departure time choice.
The period of analysis, 6 hours long, was subdivided in 36 time intervals 10 minutes long; the
known daily demand has been distributed consistently with an arbitrary temporal profile
simulating a morning peak; suitable values were assigned arbitrarily to all the input
parameters previously discussed throughout the paper.
Results obtained appear to be plausible. Figure 6, referring to one of the most congested arc of
the network, shows flow spreading from congested time intervals toward uncongested ones;
local increase of flow in congested time intervals, although counterintuitive, does not clash
with the effectiveness of the model, since travel times (and thus travel costs) are non-
increasing both locally (Figure 7), and globally (Figure 9), when the elasticity with respect to
the departure time is considered. These results confirm that the departure time elasticity yields
a more satisfactory flow pattern on the network, thus allowing higher loads on the most
attractive arcs ( Figure 8). Finally, Figure 10 shows a comparison in terms of the calculation
time and number of iterations needed to obtain equilibrium, with and without departure time
choice, on a PC with a 1.9 Ghz CPU. The same convergence criterion, defined above, was
adopted in both cases:

∑ ∑ f i
a , NLM − f ai, EQ ⋅ (τ i − τ i −1 ) ⋅ cai , EQ
a∈ A i∈I
≤ ε = 0.01
∑ ∑ a∈ A i∈I
f ai, EQ ⋅ (τ i − τ i −1 ) ⋅ cai , EQ

As expected, the case with departure time choice is more resource demanding, although
calculation times remains quite acceptable for real applications.

18
[Figure 6 here]

[Figure 7 here]

[ Figure 8 here]

[Figure 9 here]

[Figure 10 here]

6 CONCLUSIONS

The purpose of this paper was to provide a modelling framework for the simulation of elastic
demand in the context of within-day dynamic traffic assignment. To this end, a multimodal
within-day dynamic traffic assignment model is presented with a Nested Logit demand model
considering combined travel, destination, mode, departure time, and route choices. With
specific reference to departure time choice, a continuous version of the logit model is adopted,
so that enumerating explicitly the desired departure time intervals is not needed.
The resulting dynamic traffic assignment model, which allows us to represent demand and
supply dynamic phenomena concerning multimodal urban networks under congested
conditions, is formulated through a fixed-point problem, which can be solved through an
efficient implicit path MSA algorithm capable of dealing with real networks.
Both, the model and the devised algorithm, can be easily extended to the multi-class case, so
overcoming the restrictions due to the hypotheses we assume of constancy over time of the
functional form as well as of the parameters of the choice models.
Any question concerning existence and uniqueness of solutions has not been tackled in this
paper. Yet, we can say that the devised algorithm, applied to a test network, converges to
plausible solutions.

REFERENCES

Arnott R., De Palma A., Lindsey R. (1990). Departure time and route choice for the morning
commute. Transportation Research A, vol. 24, pp. 209-228

Bhat C. R. (1998). Analysis of travel mode and departure time choice for urban shopping trips.
Transportation Research B, vol. 32, pp. 361-371

Bhat C. R., Steed J. L. (2002). A continuous-time model of departure time choice for urban shopping
trips. Transportation Research B, vol. 36, pp. 207-224

19
Ben-Akiva M., De Palma A., Kanaroglou P. (1986). Dynamic model of peak period traffic congestion
with elastic arrival rates. Transportation Science, vol. 20, pp. 164-181

Cascetta E., Nuzzolo A., Biggiero L. (1992). Analysis and modelling of commuters’ departure time
and route choices in urban networks. Proceedings of the Second International CAPRI Seminar on
Urban Traffic Networks

Cascetta E. (2001). Transportation systems engineering: theory and methods. Kluwer Academic
Publisher

Bellei G., Gentile G., Papola N. (2003). A within-day dynamic traffic assignment model for urban
road networks. Accepted for publication in Transportation Research B. (Downloadable from
http://151.100.84.54/Pubblicazioni/20%20DTA%20road%20(TR2002s).pdf ).

Daly A.J., Gunn H.F., Hungerink G., Kroes E.P., Mijjer P. (1990). Peak-period proportions in large-
scale modelling. Presented to PTRC Summer Annual Meeting

De Palma A., Ben-Akiva M., Lefevre C., Litinas N. (1983). Stochastic equilibrium model of peak
period traffic congestion. Transportation Science, vol. 17, pp. 430-453

Gentile G., Meschini L., Papola N. (2003). Un modello Logit di assegnazione dinamica intraperiodale
alle reti multi modali di trasporto urbano. In: Cantarella G. E., Russo F. (Eds.), Metodi e tecnologie
dell’ingegneria dei trasporti. Seminario 2001, Franco Angeli, Milano, Italy, pp. 33-55 (Downloadable
from http://151.100.84.54/Pubblicazioni/17%20DTA%20multimodale%20(REGGIO2001).pdf )

Gentile G., Meschini L., Papola N. (2002). A within-day dynamic traffic assignment Logit model to
multimodal urban networks. In: Proceedings of the First Workshop on the Schedule-Based approach
in Dynamic Transit Modelling, Ischia, Italy, pp. 113-117 (Downloadable from
http://151.100.84.54/Pubblicazioni/12%20DTA%20transit%20logit%20(SBDTM2002).pdf )

Mahmassani H. S., Chang G. L. (1986). Experiments with departure time choice dynamics of urban
commuters. Transportation Research B, vol. 20, pp. 297-320

Mahmassani H.S., Liu Y.H. (1999). Dynamic of commuting decision behaviour under advanced
traveller information systems. Transportation Research C, vol. 7, pp. 91-107

Nguyen S., Pallottino S., Inaudi D. (1996). Postoptimizing equilibrium flows on large scale networks.
European Journal of Operational Research , vol. 91, pp. 507-516

Oppenheim N. (1995). Urban travel demand modeling. John-Wiley&Sons

Van Vuren T., Daly A., Hyman G. (1998). Modelling departure time choice. Colloquium
Vervoersplanologisch Speurwerk, Amsterdam

Van der Zijpp N.J., Lindveld C.D.R. (2000). Estimation of O-D demand for dynamic assignment with
simultaneous route and departure time choice. Presented at European Transport Conference,
University of Cambridge

20
 Generation (choice of travelling) Model

n y Destination choice Model

1 d D
Mode choice Model

1 m M
σ'
Actual departure time
choice Model
σ'-ADV σ'+DEL
τ'
Implicit path
choice Model
1 k Kmod

Figure 1 – Choice Tree for potential users from origin o with desired departure time τ'

21
 network loading map
potential demand

desired demand trip probabilities trip choices model

model

desired demand departure time

flows satisfactions

actual demand departure time departure time

model probabilities choice model

actual demand
node satisfactions
flows

arc conditional implicit path

network flow probabilities choice model
propagation model
line model

arc flows transit frequencies arc performances

arc performance
model

arc performance function

Figure 2 – Flow Diagram of Elastic Demand DTA Fixed Point Problem

22
 line node
previous arc stop arc line arc

boarding arc

waiting arc
alighting arc
queue arc
stop node
stop access arc

walking node walking arc

Figure 3 – Transit stop model

23
 Tal(τ΄) Tal(τ΄΄)

a∈l
φa(τ)
space

FA(l)
Terminal
φFA(l)(τ) time

τ΄ τ΄΄

Figure 4 – Frequency temporal profiles along the transit line

24
 network loading map
N(σ) S(σ)

PE(σ) P(VE(τ),WD(σ)) VE(σ) S(VE(σ), WD(σ))

WD(σ)
q(N(σ), PG(σ),
PD(σ) P(VD(σ),WM(σ)) VD(σ) W(VD(σ), WM(σ))
PD(σ), PM(σ))
WM(τ)

PM(σ) P(VM(σ),WDT(σ)) VM(σ) W(VM(σ), WDT(σ))

q(σ) WDT(σ)

d(q(σ), pDT(σ /τ)) pDT(σ /τ) p(VDT(σ,τ), w(τ)) VDT(σ,τ) W(VDT(σ,τ), w(τ))

d(τ) w(τ)

p(τ) p(w(τ), t(τ), c(τ)) w(c(τ), t(τ))

ω(d(τ), p(τ), t(τ))
φ( t(τ))
f (τ) φ(τ) t(τ) c(τ)

v(f (τ),φ(τ)) t(v(τ),φ(τ))

c(v(τ),φ(τ))
v(τ)

arc performance function

Figure 5 – Functional of the fixed point problem

25
 Arc 52 - Entering flow [us/h]
7000
6000
5000 No
4000 DTC

3000 DTC
2000
1000 Arc
Cap
0
600 4200 7800 11400 15000 18600

Figure 6 – Inflow temporal profile for a congested arc with and without departure time choice

26
 Arc 52 - Travel time [sec]
450
400
350
300
No
250 DTC
200
150 DTC
100
50
0
0 3600 7200 10800 14400 18000 21600

Figure 7 – Travel time temporal profile for a congested arc with and without departure time choice

27
 Arc 52 - Total flow [us]

18500
18000
17500
17000
16500
16000
15500
15000
No DTC DTC

Figure 8 – Total flow for a congested arc with and without departure time choice

28
 Network Total cost
[€*us]
205000

200000

195000

190000

185000

180000
No DTC DTC

Figure 9 – Total travel cost on the network with and without departure time choice

29
300 246 40
31
250
175 30
200
No DTC
150 20 15
100 DTC
10
50
0 0
iterations calculation time [sec] ε = 0.01

Figure 10 – Performances of the algorithm for the two DTA

30