TRISTAN V : The Fifth Triennial Symposium on Transportation Analysis 1

Rethinking the wait model at transit stops

Carolina Billi*, Guido Gentile**, Sang Nguyen***, Stefano Pallottino†

* Università di Pisa, Dipartimento di Informatica

billi@[Link]

** Università di Roma “La Sapienza”, Dipartimento di Idraulica, Trasporti e Strade

[Link]@[Link]

*** Université de Montréal, Département d'Informatique et de Recherche Opérationnelle

[Link]@[Link]

1 Introduction

In an urban transit system where the service is perceived in terms of frequency of the different
lines, the mathematical description of the route choice strategy is not trivial, because the wait at
the stop, as pointed out in many studies in the last 30 years, is a complex phenomenon which
requires a specific analysis (e.g. Chriqui and Robillard [1975], Marguier [1981], Gendreau [1984],
Spiess [1984]). In order to develop an effective model for planning the service, it is very
important to define a representation of the wait at transit stops which is at once consistent with
users’ behaviour, mathematically sound and practically usable within the more general
assignment models.
This is the reason that has induced us to rethink globally to the wait problem, by analyzing the
stochastic process of vehicle and passenger arrivals at transit stops and the relations among them,
without assuming as given facts some hypotheses that are usually adopted. Indeed, some works
written during the 80’s [Marguier, 1981; Gendreau, 1984; Marguier and Ceder, 1984] already
pointed out that some modelling choices are not easily justifiable, although often utilized in the
following years.

†
Our colleague Stefano Pallottino sadly passed away on April the 11th.

Le Gosier, Guadeloupe, June 13–18, 2004

2 TRISTAN V : The Fifth Triennial Symposium on Transportation Analysis

2 Modelling assumptions

In order to address the wait problem at a transit stop served by one or many lines, we have to
introduce several simplifying hypotheses relative to three distinct spheres: the relations among
vehicle arrivals of different lines and passenger arrivals, the relations among vehicle arrivals of
the same line, the users’ knowledge of the network and of its performances. In the following we
present briefly the assumptions on which we will base our mathematical framework, referring to
Gentile, Nguyen and Pallottino [2003a, 2003b] and Billi [2003] for further details about the
modelling choices made and their consequences.
With reference to the first sphere, we assume that the vehicle arrivals of different lines at the stop
are statistically independent, and that the same is true for the passenger arrivals with respect to
vehicle arrivals.
With reference to the second sphere, the analysis of the stochastic process of the vehicle arrivals
of a same line at a given stop induces to introduce some assumptions that, however, are perfectly
justifiable with respect to the physical phenomenon. In Billi [2003], the hypotheses made on the
headway probability distribution between two successive vehicles are made explicit (among these,
also the possible absence of a probability density function). The form that seems to be more
appropriate to represent the headway probability density function for a given line ℓi is the Gamma
function with positive integer parameter mi (also called Erlang function), which takes into
account the degree of service regularity [Larson and Odoni, 1981]. From this is derived the
probability density function of the waiting time denoted by fi(w):
 − m λ w mi −1 ( mi λi w )k
λ e i i ∑ , if w ≥ 0;
f i ( w) =  i k =0 k!
 0, otherwise;

where λi is the frequency of line ℓi, that is the inverse of the headway expected value.
A physical explanation of parameter mi associates the passage time of a given vehicle with the
passage times of the mi previous vehicles: after mi consecutive passages there is a loss of memory
within the arrival process. If we set mi = 1, the Erlang function reduces to the exponential
distribution (total absence of memory within the arrival process); the opposite case is found when
mi tends to plus infinity, obtaining the uniform distribution (regular vehicle passages however
long is the arrival sequence).
With reference to the third sphere, in the following we will consider users that perceive the
sequence of vehicle arrivals of each line only in terms of its frequency and level of regularity, and
that know relatively well the network topology and the transit stop locations. Moreover, we will
assume that each user waiting at a stop is able to evaluate with sufficient accuracy, for each line ℓi,
serving that stop, the expected value si of the travel time to reach its destination once boarded the

Le Gosier, Guadeloupe, June 13–18, 2004

TRISTAN V : The Fifth Triennial Symposium on Transportation Analysis 3

line, which we will utilize as a simplified measure of generalized cost.

Finally, we assume that each user is able to board any vehicle that he chooses to ride, suffering in
case the cost of discomfort as a function of the on-board passenger density [Nguyen and Pallottino,
1988].

3 Behavioral hypotheses

Each user is a rational decision maker doing his choices with the aim of minimizing his own
generalized cost. In a static model such as that we are referring to, the choice of a user is made
before the beginning of the trip and implies the iterated choice sequence of the stop to reach on
foot where to wait, of the lines to board and, for each of them, of the stop where to alight. The
classical representation of this kind of choice is a hyperpath connecting the origin of the trip to its
destination and having the diversion points in the stop nodes through the waiting hyperarc
[Nguyen and Pallottino, 1998]. Key elements of the hyperpath are the weights associated with the
branches of each waiting hyperarc, that represent the probability of boarding a vehicle of the
different lines of that specific line set, conditional on having reached that stop.
Let us consider then a single transit stop, a set L of lines serving it, and a user that during his trip is
there waiting to board one or more lines belonging to L in order to reach his destination. The
probability of boarding line ℓi ∈ L measures the chance that a vehicle of such line is the first one
arriving at the stop which is perceived as attractive by the user during his waiting process, in the
sense that it is convenient to board that vehicle instead of keep waiting.
Note that, differently from what it is usually assumed in the literature, in this paper we do to
conceive in principle the attractive set as a collection of lines that are considered always attractive
by the user during the entire waiting process. Indeed, in general we assume that, before starting
his trip, the user has already defined the lines of L he wishes to board and, for each of them, the
period of his waiting process where he considers the line to be attractive. In other words, we
assume that the attractive set of a user is defined as a function of the time t elapsed form his arrival
at the stop waiting without success; for this reason we will call it dynamic set and denote it by D(t),
defined for 0 ≤ t ≤ uD, where uD is the maximum waiting time, i.e. the first instant when with
certainty at least one vehicle has arrived at the stop while considered attractive (eventually,
uD = +∞).
In the behavioural model we take into consideration the time spent by the user waiting at the stop
without any attractive arrival; this time is an information that the user acquires during the wait and
that influences the perception of the expected remaining waiting time of each line. For example,
let us assume that line ℓi is perfectly regular ( fi(w) is the uniform distribution function) and has
headway ui = 20 min and line time si = 30 min. When the user reaches the stop and starts waiting,
the expected value of the travel time is 10 + 30 = 40 min. After 10 min waiting, the expected value
Le Gosier, Guadeloupe, June 13–18, 2004
4 TRISTAN V : The Fifth Triennial Symposium on Transportation Analysis

of the remaining waiting time is 5 min and the remaining travel time has decreased to 5 + 30 = 35
min (instead, the total travel time has increased, because to the 35 min we have to add the 10 min
already waited).

4 Properties of the dynamic set

Let D(t), 0 ≤ t ≤ uD, be the dynamic set which has been defined “a priori” by the user. A remaining
time function RTD(t), 0 ≤ t ≤ uD, is associated with D(t), expressing for each instant τ ∈ [0, uD] the
time RTD(τ) that remains to wait and to ride in order to reach the destination, given that no
attractive arrival has occurred during the interval [0, τ). The condition defining the attractivity of
a line ℓi at time τ is given by:
si ≤ RTD (τ ) ; (1)

in this case, in fact, it is more convenient to board an arriving carrier of the line than to keep
waiting at the stop, while in the opposite case it is not convenient to board. We say that D(t) is
instantaneously attractive at time τ if it is constituted by all lines ℓi ∈ L that satisfy condition (1);
moreover D(t) is globally attractive if and only if it is instantaneously attractive for each
τ ∈ [0, uD].
In [Gentile, Nguyen and Pallottino, 2003b] it is shown that, assuming the monotonicity of the
remaining waiting time of each line (that is, when the time spent waiting in vain at the stop
increases, the remaining waiting time of each line does not increase), then the remaining time
function of a globally attractive dynamic set is monotonically not increasing. From this property
of monotonicity derives that, if a line ℓi ∈ L is attractive, it is such for a time interval [0, ti], with
0 < ti ≤ uD (a line which is not attractive at time τ = 0 is never attractive). Moreover, the globally
attractive dynamic set exists, is unique and varies (reducing itself) a finite number of times. It is
defined, by construction, as:
D (τ ) = {A i ∈ L : si ≤ RTD (τ )}, ∀0 ≤ τ ≤ uD . (2)

Without loss of generality, let us consider that the lines are ordered based on their travel times and
that these are different from each other, that is s1 < s2 < … < sn, with n = |L|; moreover, we will
denote by Lj = {ℓ1, ℓ2, … , ℓj} the set of the j fastest lines in terms of travel times. Consider the time
tj when RTD(tj) = sj. Because RTD(t) is a monotonically non increasing function, from (2) we have
that D(tj) = Lj and that there exists a time tj+1 < tj at which RTD(tj+1) = sj+1 (or tj+1 = 0 and
RTD(0) ≤ sj+1). In the time interval (tj+1, tj] spent waiting (or [0, tj] if tj+1 = 0) the dynamic set does
not change (we say that it is “static”). Moreover, the dynamic of D(t) is described at the most by n
time intervals where D(t) is static. In general, we denote by Lq the set of lines that are attractive
during the whole definition interval [0, uD] and by Lr, r ≤ q, the set of lines that are attractive at
Le Gosier, Guadeloupe, June 13–18, 2004
TRISTAN V : The Fifth Triennial Symposium on Transportation Analysis 5

time τ = 0; thus the lines ℓr+1, …, ℓn are never attractive.

5 Travel time expected value and line probabilities

In [Gentile, Nguyen and Pallottino, 2003b] is developed the theoretical framework required to
formulate the function RTD(t) and the probabilities to board the different lines. Here we provide
only the most important elements; let ui denote the upper bound of the definition interval relative
to the function fi(w) for line ℓi (+∞ in the case of unbounded distribution); while Fi(w) is the
cumulative distribution function of the waiting time. In particular, we are interested in its
complement:
ui

Fi ( w) = 1 − Fi ( w) = ∫ f i (t ) dt .
w

The probability density function of the waiting time of line ℓi, conditional to having already
waited in vain until time τ is:
 f i ( w)
 , if w ∈ [τ , ui ];
f i|τ ( w) =  Fi (τ )
 0, otherwise;

the complement of the corresponding cumulative distribution function is:
 Fi ( w)
ui
 , if w > τ ;
Fi|τ ( w) = ∫ f i|τ (t ) dt =  Fi (τ )
w  1, otherwise.

Recalling the meaning of the instants tj for the different attractive lines, let us consider the case
when the user has waited in vain until time τ, with tj+1 < τ ≤ tj, when the lines ℓ1, …, ℓj are
attractive, while the lines ℓj+1, …, ℓr are not anymore attractive. If the first attractive arrival occurs
at time w , the (density of) probability γi|τj(w) that it is relative to a vehicle of line ℓi, with i ≤ j, is
given by:
j
f i ( w) j Fk ( w)
γ i|jτ ( w) = f i|τ ( w) ∏
k =1,k ≠i
Fk |τ ( w) = ∏
Fi ( w) k =1 Fk (τ )
.

On this basis, the share of travel time expected value relative to the case where the first attractive
arrival occurs within the time interval [τ, tj], is given by:
t
j
 j f ( w)  j Fk ( w)
Ψ j (τ , t j ) = ∫  ∑ ( w + si − τ ) i ∏ dw . (3)
τ  i =1 Fi ( w)  k =1 Fk (τ )

Le Gosier, Guadeloupe, June 13–18, 2004

6 TRISTAN V : The Fifth Triennial Symposium on Transportation Analysis

Knowing that RTD(tj) = sj, by means of the latter formula it is possible to provide a “local”
description of the remaining time function RTD j(t, tj) relative to the generic interval (tj+1, tj] where
the static set of attractive lines is Lj:
j
Fk (t j )
RTDj (τ , t j ) = Ψ j (τ , t j ) + s j ∏ , j = q,..., r . (4)
k =1 Fk (τ )

In the above formula it is to be intended tq = uD; moreover, in the case of unbounded headways, we
have in general q = 1, except for the case of exponential distribution where the classical results are
valid and replace the above formulation.
Equation (4) expresses the remaining travel time as the sum of its share Ψj(τ, tj), relative to the
case where the first attractive arrival occurs within the time interval [τ, tj], and the complementary
share sj, opportunely weighted by the probability that the user has to wait in vain until time tj.
Using (4) it is possible to verify if line ℓj+1 is attractive for at least a part of the waiting process. In
fact, it is sufficient to check that RTD j(0, tj) ≤ sj+1. In this case, line ℓj+1 is never attractive, that is
r = j and tj+1 = 0. Otherwise we determine the time tj+1 such that RTD j(tj+1, tj) = sj+1, and then we
have RTD(τ) = RTD j(τ, tj), for each τ∈(tj+1, tj]. The expected travel time is given by:
ET = RTD(0);
while the probability of utilizing the vehicles of line ℓi ∈ Lr is given by:
tj
r r
f i ( w) j
πi = ∑ ∏
j = max{i , q} k = j +1
Fk (tk ) ∫ F ( w) ∏ k =1
Fk ( w) dw . (5)
t j +1 i

Clearly πi = 0 for the non attractive lines ℓi, i = r + 1, …, n.

6 Computational aspects

The local form of RTD j(t, tj) suggests a “backward construction” of the remaining time function.
The determination of q can be achieved starting from q = 1 and updating q with j+1 whenever
during the backward construction of the remaining time function we find tj+1 ≥ uj+1.
The above formulation requires in general the computation of many integrals, which might be
somewhat time consuming unless introducing approximations whose propagation is difficult to
control. In the case of Erlang functions this can be avoided, noting that all the functions to be
integrated are a sum of different terms whose generic form is:
v
v! xk
∫ β e x dx = − β e ∑
-α x v -α x
+c;
k =0 k ! α
v − k +1

where α and β are real parameters, while v is a nonnegative integer parameter and c in an additive
constant. On this basis it is possible to determine the expected value of the travel time and the line
Le Gosier, Guadeloupe, June 13–18, 2004
TRISTAN V : The Fifth Triennial Symposium on Transportation Analysis 7

probabilities through computations that are easy to organize in order to reduce the number of
operations and to control the propagation of approximation errors. This approach is inspired to the
one proposed by Gendreau [1984].

7 Necessity of a dynamic model

Consider the following example with only two lines whose waiting time has a uniform
distribution: u1 = u2 = 50 min, s1 = 30 min and s2 = 50 min. Applying the proposed model we
obtain the following globally attractive dynamic set: D(t) = L2 for 0 ≤ t ≤ 10 and D(t) = L1 for
10 < t ≤ 50, with an expected value of the travel time RTD(0) = 54.53 and line probabilities
π1 = 0.82, π2 = 0.18.
Applying instead the classic static approach, which is valid only for exponential distribution, we
obtain L1 as optimal attractive set with a travel time expected value of 55 (the travel time expected
value of the set L2 is 56.67). However, in this case at the beginning of the wait the user renounces
to board the vehicles of the second line, although this would be convenient in terms of travel time;
that is, the assumption of passenger’s rational behaviour fails.

8 Conclusions

We have described a new, well founded, model for the line choice at the stops of an urban transit
network, where it is possible to take into account the level of service regularity for each line and
for each one of its segments. We have also suggested a computational technique that guaranties
the efficiency of the model and the possibility of plugging it into any transit assignment model
based on hyperpaths.

References

C. Billi, Sui modelli matematici per il trasporto pubblico urbano, Graduation thesis, Dipartimento
di Informatica, Università di Pisa, Italy (2003).
C. Chriqui and P. Robillard, “Common bus line”, Transportation Science 9, 115-121 (1975).
M. Gendreau, Une étude approfondie d'un modèle d'équilibre pour l'affectation des passagers
dans les réseaux de transport en commun, Ph.D. thesis, Département d'informatique et de

Le Gosier, Guadeloupe, June 13–18, 2004

8 TRISTAN V : The Fifth Triennial Symposium on Transportation Analysis

recherche opérationnelle, Université de Montréal, Canada (1984).

G. Gentile, S. Nguyen and S. Pallottino, “Route choice on transit networks with on-line
information at stops”, TR-03-14, Dipartimento di Informatica, Università di Pisa, Italy,
accepted for publication on Transportation Science (2003a).
G. Gentile, S. Nguyen and S. Pallottino, “Passengers’ behaviour in transit networks with common
lines: a dynamic attractive set model”, Manuscript, Dipartimento di Informatica, Università di
Pisa, Italy (2003b).
R.C. Larson and A.R. Odoni, Urban Operations Research, Prentice-Hall, Englewoods Cliffs, NJ
(1981)
P.H.J. Marguier, Optimal strategies in waiting for common bus lines, Master's thesis, Department
of Civil Engineering, M.I.T., Cambridge, MA (1981).
P.H.J. Marguier and A. Ceder, “Passenger waiting strategies for overlapping bus routes”,
Transportation Science 18, 207-230 (1984).
S. Nguyen and S. Pallottino, “Equilibrium traffic assignment for large scale transit networks”,
European Journal of Operational Research 37, 176-186 (1988).
H. Spiess, Contributions à la théorie et aux outils de planification des réseaux de transport urbain,
Ph.D. thesis, Département d'informatique et de recherche opérationnelle, Université de
Montréal, Canada (1984).

Le Gosier, Guadeloupe, June 13–18, 2004