Vol 16 No 6, June 2007 c 2007 Chin. Phys. Soc.

1009-1963/2007/16(06)/1608-07 Chinese Physics and IOP Publishing Ltd

Behaviours in a dynamical model of traffic

assignment with elastic demand∗
Xu Meng(M ˆ)† and Gao Zi-You(pgl)
School of Traffic and Transportation, Beijing Jiaotong University, Beijing 100044, China

(Received 15 September 2006; revised manuscript received 16 October 2006)

This paper investigates the dynamical behaviour of network traffic flow. Assume that trip rates may be influenced
by the level of service on the network and travellers are willing to take a faster route. A discrete dynamical model
for the day-to-day adjustment process of route choice is presented. The model is then applied to a simple network for
analysing the day-to-day behaviours of network flow. It finds that equilibrium is arrived if network flow consists of
travellers not very sensitive to the differences of travel cost. Oscillations and chaos of network traffic flow are also found
when travellers are sensitive to the travel cost and travel demand in a simple network.

Keywords: discrete dynamical system, network traffic flow, traffic assignment problem, chaos
PACC: 0545, 0547, 0250

1. Introduction tions began to be asked about the behaviour basis of

the equilibrium paradigm, particularly in the sense of
its usefulness for representing travellers’ response to
The notion of equilibrium in the analysis of urban exogenous information. This led to a growing interest
transportation networks stems from the dependence of in day-to-day dynamical models, whereby the evolu-
the link travel costs on the link flows. Network equilib- tion over days of travel choices and traffic congestion
rium models have a long history in the traffic flow re- is linked through a learning model based on travellers’
search literature.[1−3] Particularly, the classes of static past experiences.[9−11]
equilibrium models have been seen to be versatile,[4,5]
Many general contrasts may readily be made be-
with extensions able to represent perceptual variations
tween the day-to-day dynamics and traditional equi-
across the driver population, mixed vehicle types, and
librium approaches. For example, the estimation of
some mixed mode operations.
equilibrium has been typically achieved through the
In the 1980s, with a growing international inter- solution of some optimization, complementarity or
est in real-time information and control systems, a variational inequality problem, yet this makes the ap-
couple of fundamental issues of network equilibrium proach rather restrictive in terms of the feasible gen-
models began to be questioned:[6,7] for example, what eralizations. On the other hand, given a suitably con-
the required conditions for equilibrium are, if equi- structed equilibrium solution algorithm, the interpre-
librium is really reached, what the relationship be- tation of model outputs is typically straightforward,
tween travellers’ learning processes and network be- whereas in simulations of the day-to-day approach
haviour is, and also, their inability to deal with many the interpretation stage is often the most challenging
of the dynamic aspects of traffic systems. Initially, task.[12,13] However, perhaps the greatest contrast to
the predominant focus was on efforts to generalize the be made is the fundamental basis of each approach.
equilibrium model to represent within-day dynamical In traditional equilibrium models, the underlying hy-
models, that is to say, the dependence of route choice pothesis is the very notion of a market in equilib-
on departure time, while accounting for the proper rium: a typically isolated, ‘self consistent’ state of
temporal and spatial evolution of congestion across the network which, if attained, would persist under
the network.[8] Subsequently, more fundamental ques- certain rational rules of behaviour. In the day-to-day
∗ Project supported by the National Basic Research Program of China (Project No 2006CB705500), the National Natural Science
Foundation of China (Grant No 70471088), and Science and Technology Foundation of Beijing Jiaotong University (Grant No
2006RC018)
† E-mail: xumeng@jtys.bjtu.edu.cn

http://www.iop.org/journals/cp http://cp.iphy.ac.cn
No. 6 Behaviours in a dynamical model of traffic assignment with elastic demand 1609

approach, the underlying belief is in the behavioural 2. Model

dynamics, namely how the behaviour on day n is af-
fected by behaviour and the state of the network on
days n − 1 and earlier. It is therefore tempted to re-
Consider in a urban transportation G = (N, A),
gard the two approaches as competing philosophies.
where N is the set of nodes, A is the links set. As-
Yet the classical preliminary analyses of dynamical
sume that trip rates may be influenced by the level
approaches involve examining the ‘equilibrium’ be-
of service on the network and travellers’ willingness of
haviour of the system. While the concept of equi-
taking a faster route. For example, as congestion in-
librium of a dynamical system does not relate directly
creases, travellers may decide to use a different mode
to the traditional economic equilibrium of traffic net-
of travel (e.g. subway), shift the time of travel (out-
work models, there are many important links between
side the design period), or forgo some trips altogether.
the two approaches.[14−17]
The trip rate, qrs , between every O-D pair r − s in the
Indeed, dynamical models can be seen as a gen- network can be assumed to be a function of travel time
eralization of equilibrium models since they simulate between origin r and destination s. In other words,
the convergence of the supply-demand system towards
possibly different equilibrium states and transient
states visited due to modifications of supply and/or qrs = Drs (urs ), ∀r,s . (1)
demand.[18−20] Furthermore, under some rather mild
assumptions, equilibrium configurations of the system Where urs is the minimum travel cost between r and
can be modelled as attractors of the system, for ex- s, and Drs (·) is the demand function (for vehicular
ample, states in which the evolution of the system trips) between r and s. It can include that O-D spe-
stops. Finally, the dynamical approaches allow anal- cific parameters reflect population size, income distri-
ysis of the stability of equilibrium configurations and bution, and vehicle ownership for origin nodes, as well
provide a complete statistical description of the sys- as employment intensity or retail activity variables for
tem’s evolution.[21] destination nodes.
Theoretical day-to-day dynamical models can be
The demand function can be expected to be
classified into two groups according to whether they
monotonically decreasing (or at least non-increasing)
model the system in continuous time or discrete time.
in the O-D travel time. In other words, as urs grows,
The former is mainly differential equation systems,
qrs decreases, and vice versa. This function is also
which assumes revisions on a route, may be in contin-
bounded from above.[26] For example, the maximum
uous time.[22,23] The latter is formulated as difference
number of vehicular trips generated between a given
equations or simulation models. In reality, repeated
origin and a given destination in a certain design pe-
trips are subjected to active constraints which mean
riod maybe bounded by the population size at the
that they are not continuously adjustable in time (i.e.
origin.
over days), therefore, a more appropriate representa-
tion would seem to be a discrete time system, criti- The concept behind the traffic flow dynamics (or
cally, this allows an interpretation to be given to the the adjustment process) is that path flow along paths
rate of adjustment.[24,25] increases if path cost is less than the O-D travel cost,
In this study, we consider a general network equi- while it decreases if the path cost exceeds O-D travel
librium model with elastic demands. It is assumed cost; similarly, the O-D travel cost increases if total
that one has associated each origin–destination (O-D) path flow on the same O-D pair is less than the O-
pair in the network with a travel disutility function. D demand flow, while it decreases if the path flow
A model for the process by which a route choice is ad- exceeds the O-D demand flow.[27] This may be seen
justed day-to-day is formulated as a discrete dynam- as natural, and is similar to the projected dynami-
ical system. Then this model is applied to a simple cal system. These dynamics can be expressed by the
network for a behaviour analysis of network flow. following difference equation:
1610 Xu Meng et al Vol. 16

fp (i + 1) − fp (i) = Ψp ⌊fp (i)⌋, (2a)

 n o
 max η [urs (i) − cp (i)]λ , −fp (i) , if cp (i) − urs (i) > 0,
Ψp [fp (i)] = (2b)
 η [urs (i) − cp (i)]λ , otherwise.

urs (i + 1) − urs (i) = Φrs [urs (i)], (3a)

  " #λ 

  P  P
 max η qrs [urs (i)] −

 f p (i) , −u rs (i) , if fp (i) − qrs [urs (i)] > 0,
p∈Prs p∈Prs
  
Φrs [urs (i)] = " #λ (3b)

 P
 η qrs [urs (i)] − fp (i) , otherwise.



p∈Prs

Define the function Fp (fp ), Urs (urs ) as follows:


 max f (i) + η[u (i) − c (i)]λ , 0 , if cp (i) − urs (i) > 0,
p rs p
Fp (fp ) = fp + Ψp (fp ) = (4a)
 fp (i) + η[urs (i) − cp (i)]λ , otherwise.

Urs (urs ) = urs + Φrs (urs )

  " #λ 

  P  P


 max u rs (i) + η qrs [u rs (i)] − f p (i) , 0 if fp (i) − qrs [urs (i)] > 0,
p∈Prs p∈Prs
  
= " # λ
(4b)


 u (i) + η q [u (i)] − P f (i) ,

otherwise.
 rs
 rs rs p
p∈Prs

The dynamics can then be rewritten as the following equations:

fp (i + 1) = Fp ⌊fp (i)⌋, (5a)

urs (i + 1) = Urs [urs (i)]. (5b)

Or,

fp (i + 1) = max{0, η[urs(i) − cp (i)]λ + fp (i)}, ∀p ∈ P rs , ∀r,s , (6a)

  λ 

 

X
urs (i + 1) = max 0, η qrs (urs (i)) − fp (i) + urs (i) , ∀r,s , (6b)
 
 p∈Prs 

where fp (i) is path flow on day i and on path p, cp (i) is depressive about the difference of the travel times.
is path cost on day i and on path p, urs (i) is the cost From the above model, if path flow fp (i) equals to 0,
on day i between origin r to destination s, η and λ are which means that the path travel cost is larger than
positive parameters. qrs (urs (i)) is the trip demand of O-D cost; if O-D cost urs (i) equals to 0, which means
origin r to destination s on day i. The former param- that the path flow is larger than O-D demand.
eter, η, represents the sensitivity of the adjustment.
If the parameter, λ, is more than 1.0, travellers react Let Fp′ (fp ), Urs′
(urs ) denote the derivative of
more sensitively when the difference is large. Oth- Fp (fp ) and Urs (urs ) respectively. The derivative of
erwise, if λ is less than 1.0, the adjustment function Fp (fp ) and Urs (urs ) at equilibrium, Fp′ (fp∗ ), Urs
′
(u∗rs ),
No. 6 Behaviours in a dynamical model of traffic assignment with elastic demand 1611

can be classified into three cases for the stability anal- 3. Numerical analysis
ysis:
The numerical test was implemented on a com-
puter conducted with double precision arithmetic on
 an Intelr Pentiumr 4 CPU 1.75 GHz, 512 MB RAM,
 1, λ ≥ 1, by using the Microsoft Window XP (SP2) operating
Fp′ (fp∗ ) = ∀p ∈ P rs , (7)
 ∞, 0 < λ < 1, system. All of the coding was carried out in Mat-
 lab 6.0. We study the discrete dynamical model on a
 1, λ>1 simple network, which is shown in Fig.1. For this net-


′
Urs (u∗rs ) = 1+ ′
ηqrs (u∗rs ), λ=1 ∀r, s.(8) work, there are three nodes: 1, 2, 3; three links: a, b, c;

and a single O-D pair (1, 3). Let path p1 = (a, c) and

∞, 0<λ<1

path p2 = (b, c).

From Eqs.(7) and (8), when 0 < λ < 1, both |Fp′ (fp∗ )|
′
and |Urs (u∗rs )| are greater than 1 and the equilibrium
is not stable. When λ > 1, the equilibrium is locally
stable. When λ = 1, in the case that η is sufficiently
large, it is clear that the equilibrium is not (locally)
′
stable because |Urs (u∗rs )| > 1, at this time, the step Fig.1. An elastic demand example.
size, urs (i+1)−urs (i), is very large. In the case that η
is sufficiently small but not 0, the equilibrium is locally Assume that the user link cost functions are
′
stable because η > 0, λ > 0 and qrs (u∗rs ) < 0, in this
ca (f ) = 5fa + 2fb + 15,
case, the equilibrium might also be globally stable be-
cb (f ) = 7fb + fa + 15,
cause the step size, urs (i + 1) − urs(i), is so small that
urs approaches the equilibrium gradually. Further- cc (f ) = 3fc + fa + fb + 12,
more, when 0 < λ < 1, the equilibrium is not stable
and the disutility (or inverse demand) function is given
and the flow does not converge even if η is small. This
by
could imply that the network flow does not converge
to the equilibrium and is not stable when travellers u13 (d13 ) = −2d13 + 114.
react more sensitively to small cost differences than
to large. And the dynamical system of the network is

fp1 (i + 1) = max 0, η[u13 (i) − cp1 (i)]λ + fp1 (i) ,



fp2 (i + 1) = max 0, η[u13 (i) − cp2 (i)]λ + fp2 (i) ,


  λ 

 

X
u13 (i + 1) = max 0, η q13 (u13 (i)) − fp (i) + u13 (i) .
 
 p∈P13 

From the discrete dynamic model (6a),(6b), un- tive to initial conditions, when the initial condition is
der some parameters, we find complicate behaviours changed slightly, the O-D trip flow changes clearly.
happening in this test network. Oscillations and some Figures 2–4 demonstrate the different characters
apparently irregular behaviours were found in certain of two path flows under different parameters. The
intervals of different parameters. Under the irregular total iteration number is 1000 and the bifurcation pa-
behaviours, it is also found that the model is sensi- rameter η ∈ [0, 1.2] with step size 0.003, for better
1612 Xu Meng et al Vol. 16

show the characters of traffic flow, the figures visu- (5,4) with total O-D cost urs = 96. Figure 3 shows
alized the values of iteration number from 1 to 150. that the path flows have oscillation with η = 0.3001,
Figure 2 shows the two path flows of the test network, the path flow behaviour is very complex in Fig.4 with
with the same initial conditions for η = 0.1051. It η = 0.9001.
can be found that the two path flows arrive equilibria

Fig.2. Equilibrium flow on two paths with η = 0.1051.

Fig.3. Oscillation flow on two paths with η = 0.3001.

Fig.4. Chaos flow on two paths with η = 0.9001.

There are different choices for the initial O-D flow the parameter, given the initial condition at the first
at every iteration. For convenience of comparison, step. From Figs.5 and 6, by starting from different ini-
we choose two ways: one uses the same initial con- tial conditions, different path flows may be detected if
ditions for all steps of the parameter, the other uses there is more than one attractor for the same value of
the final states of the system at the previous step of the parameter. Figures 7 and 8 show the O-D travel
No. 6 Behaviours in a dynamical model of traffic assignment with elastic demand 1613

cost behaviour about η under different initial condi- state and oscillating between path flow less than O-
tions. From the figures, it is found that when the D flow and path flow larger than O-D flow; when
sensitivity parameter is in the interval [0, 0.15], that η ∈ [0.82, 0.94], more complex path flows appear, the
is, travellers are not so sensitive to the travel cost, path flow cannot be estimated; when η ∈ [0.94, 1.2],
the network flow equilibrium state can be arrived at; the path flow will change alternatively under different
when η ∈ [0.15, 0.82], the path flow begin to oscillate, states.
which causes the path flow arriving at non-equilibrium

Fig.5. Bifurcation diagram of path flow about η, initial values are the same for all values of η.

Fig.6. Bifurcation diagram of path flow about η, initial values are the final states of the previous step of η.
1614 Xu Meng et al Vol. 16

4. Conclusions
This paper assumes that travellers are willing to
switch to a faster route, and a model of the day-to-day
route adjustment process with elastic demand is for-
mulated as a discrete dynamical system. The model
is applied to a simple network, in which one origin-
destination pair is connected by different paths. In
order to examine the properties of the day-to-day dy-
namics of network flow, analysis is carried out to in-
vestigate the status of network flow.
The results of the study can be summarized as fol-
Fig.7. Bifurcation diagram of O-D travel cost about η,
lows: the condition for stability of network flow is that
initial values are the same for all values of η.
the traveller is not very sensitive to differences of cost.
At this time, the network equilibrium appears under
the elastic demand. If travellers become sensitive to
the path cost difference, the flow does not converge to
equilibrium and begins to oscillate or has chaotic be-
haviours. When network flow becomes chaotic under
certain conditions of travellers reaction in the test net-
work, network flows cannot be observed with infinite
accuracy. So, if this is the case long-term forecasts
of flows cannot be made due to sensitivity to initial
conditions, which is the nature of chaos.
As a future work, this model and its analysis
should be expanded to a more general network. The
Fig.8. Bifurcation diagram of O-D travel cost about η, question of whether or not there exists network flow
initial values are the final states of the previous step of η. in the real world for which chaotic behaviour occurs
should be examined.

References [14] Horowitz J L 1984 Trans. Res. Part B 18 13

[15] Xu W, Jin YF, Li W and Ma SJ 2005 Chin. Phys. 14 1077
[1] Bell G H and Lida Y 1997 Transportation Network Anal- [16] Yang W S, Wang B H, He P, Wang W N, Quan H J and
ysis (New York: Wiley) Xie Y B 2003 Chin. Phys. 12 931
[2] Nagurney A and Zhang D 1996 Projected Dynamical [17] Zhang X and Jarrett D F 1998 Chaos 8 503
Systems and Variational Inequalities with Applications [18] Jin Y F, Xu W, Ma S J and Li W 2005 Acta Phys. Sin.
(Boston: Kluwer) 54 3480 (In Chinese)
[3] Ortúzar J de D and Willumsen L G 2001 Modelling Trans- [19] Li K P and Gao Z Y 2005 Chin. Phys. 14 930
port (third edition) (New York: Wiley) [20] Luo X S, Wang B H, Jiang F and Gao Y 2001 Chin. Phys.
[4] Patriksson M 1994 The Traffic Assignment Problem— 10 17
Models and Methods (Utrecht: VSP) [21] Dendrinos D S 1994 Chaos, Solitons & Fractals 4 605
[5] Sheffi Y 1984 Urban Transportation Networks: Equilib- [22] Friesz T L, Bernstein D, Mehta N J, Tobin R L and Gan-
rium Analysis with Mathematical Programming Methods jalizadeh S 1994 Oper. Res. 42 1120
(Englewood Cliffs, NJ: Prentice-Hall) [23] Zhang D and Nagurney A 1997 J. Opti. Theo. Appl. 93
[6] Nagurney A and Zhang D 1997 Trans. Sci. 31 147 417
[7] Nair A S, Liu J, Rilett L and Gupta S 2001 IEEE Intel. [24] Safanov L A, Tomer E, Strygin V V, Ashkenazy Y and
Trans. Sys. Conf. Proc. Oakland 681 Havlin S 2002 Euro-phys. Lett. 57 151
[8] Smith M J 1984 Trans. Sci. 18 245 [25] Van Z, Henk J, Marina S, Van G and Peter N 1999 Trans.
[9] Cascetta E 2001 Transportation Systems Engineering: Res. Rec. 1685 21
Theory and Methods (Boston: Kluwer) [26] Weidlich W 2000 Socio Dynamics: A Systematic Ap-
[10] Li K P and Gao Z Y 2004 Mod. Phys. Lett. B 18 1395 proach to Mathematical Modeling in the Social Sciences
[11] Li K P, Gao Z Y and Chen T L 2003 Chin. Phys. 12 1213 (Amsterdam: Harwood Academic)
[12] Wang L, Wang B H and Hui P M 1997 Chin. Phys. 6 29 [27] Wilson A G 1970 Entropy in Urban and Regional Model-
[13] Wang T, Gao Z Y and Zhao X M 2006 Acta Phys. Sin.
ing (London: Pion)
55 634 (in Chinese)