SCHEDULE-BASED TRANSIT ASSIGNMENT:

A NEW DYNAMIC EQUILIBRIUM MODEL

WITH VEHICLE CAPACITY CONSTRAINTS

Natale Papola, Francesco Filippi, Guido Gentile, Lorenzo Meschini

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

ABSTRACT

We propose in this paper a new approach for modelling congested transit

networks with fixed timetables where it may happen that there is not enough
room onboard to allow all users waiting for a given line on the arriving
carrier, so that passengers need to queue at the stop until the service becomes
actually available to them.
The traditional approach to reproduce this phenomenon within the
established framework of diachronic graphs, where the supply is represented
through a space-time network, is to introduce volume-delay functions for
waiting arcs, which are meant to discourage passengers from boarding
overcrowded carriers. However, this produces a distortion on the cost
pattern, since passengers who achieve boarding do not suffer any additional
cost, and may also cause numerical instability.
To overcome these limitations we extend to the case of scheduled
services an existing Dynamic Traffic Assignment model, allowing for
explicit capacity constraints and FIFO queue representation, where the
equilibrium is formulated as a fixed point problem in terms of flow temporal
profiles.
The proposed model propagates time-continuous flows of passengers on
the pedestrian network and time-discrete point-packets of passengers on the
line network. To this end, the waiting time pattern, corresponding to a given
flow temporal profile of pedestrians who reach a stop to ride a certain line,
 2 Natale Papola, Francesco Filippi, Guido Gentile, Lorenzo Meschini

has a saw-tooth temporal profile such to concentrate passengers on the

scheduled runs, while satisfying the constraint that the number of boarding
users must not be higher than the onboard residual capacities.
An MSA algorithm is also devised, whose efficiency is tested on the
regional transit network of Rome.

1 INTRODUCTION

The pricing and rationing measures applied to discourage the use of

private cars, in order to alleviate the increasing road congestion and the
consequent worsening pollution, are not always coupled with a consistent
policy aimed at improving the performances, or at least the capacity, of the
transit system. As a result, in many modern cities the problem of full transit
carriers is becoming more and more relevant, both for the urban system and
for the regional system used by commuters. Although this situation should
be avoided through a correct design of the transit network by suitably
increasing the line capacities, it is still important to properly simulate the
current scenario in order to justify the resources needed to carry out
appropriate interventions.
The frequency-based static assignment models commonly used to plan
transit networks are suited to represent urban systems (metro, tramways,
busses) where the service is so irregular or so frequent that there is no point
for passenger to synchronize their arrivals at the stop with the scheduled
time of carriers, if any is published. In this context it is generally assumed
that a passenger, once reached a stop, waits for the first attractive carrier
among some common lines. This leads to the concept of optimal strategy
(Spiess and Florian, 1989) which can be formally expressed in terms of a
shortest hyperpath (Nguyen and Pallottino, 1988). On the contrary, in extra-
urban systems (aeroplanes, trains, coaches) and in general when the
frequency is so low that the timetables must be known in order to make the
service usable, passengers reach the stop with the intension of travelling on a
specific run of a specific transit line. To represent this choice or whenever
we want to obtain from the transit assignment the loads and the
performances of each single run – a detail which is highly desirable when
designing the service exercise – a schedule-based approach is needed.
For a detailed analysis of the literature on schedule-based transit
assignment we refer the reader to the recent book of Nuzzolo et al. (2003)
and to the selection, edited by Wilson and Nuzzolo (2004), of contributions
presented at the First Workshop on Schedule-based approach in Dynamic
Transit Modelling (SBDTM). The most natural and well established
modelling approach involves the representation of transit supply, which is
 3 Natale Papola, Francesco Filippi, Guido Gentile, Lorenzo Meschini

intrinsically discrete in time, as a diachronic graph (Nuzzolo and Russo,

1993), where each run is modelled through a specific sub-graph whose nodes
have space and time coordinates according to the timetable. As an
alternative, it is possible to define a dual graph (Nielsen and Jovicic, 1999),
where each run section is a node, while the arcs represent the connections at
stops satisfying temporal consistency. A third approach, referred to as mixed
line-database, is to describe the topology of the transit network through a
graph analogous to that used in the static assignment, and to characterize its
arcs with the information relative to the timetable (Tong and Wong, 1999;
Hickman and Bernstein, 1997; Nielsen, 2000).
In this paper, we will develop a new approach that resembles to a certain
extent the latter and consists in extending to the simulation of scheduled
services an existing Dynamic Traffic Assignment (DTA) model based on a
macroscopic representation of time-continuous flows. This is specified in
Bellei et al. (2005) for road networks and in Gentile et al. (2003) for
multimodal networks, where the transit system is described in terms of line
frequency temporal profiles, which allows representing the average effect of
time-discrete services on the travel cost pattern. Here, we will introduce an
appropriate arc performance function, yielding saw-tooth temporal profiles
of the waiting times that concentrate passengers on the scheduled runs. This
way the network loading map will propagate, accordingly with a logit route
choice model, time-discrete point-packets on the line network and time-
continuous flows on the pedestrian network. The task of spreading on the
pedestrian network the point-packets travelling on the line network is
conferred to the alighting arcs.
One of the open questions in transit assignment is how to comply with
vehicle capacity constraints, that may induce the formation and dispersion of
queues at stops, where passengers wait for the first run of the chosen line
actually available to them. The traditional approach to reproduce this
congestion phenomenon in a static framework is based on the concept of
effective frequency (DeCea and Fernandez, 1993), stating that the line
frequency perceived by the passengers waiting at a stop decreases as the
probability of not boarding its first arriving carrier increases. Since the
residual capacity of a run available to passengers waiting at the stop depends
on the amount of those already onboard, who do not suffer the cost of
queuing, then to apply properly the effective frequency approach an
asymmetric arc cost function is to be introduced as in Bellei et al. (2003). A
different approach is proposed in Kurauchi et al. (2003), where passengers
mingle on the platform (that is no FIFO rule holds in the queue), while fail-
to-board arcs and probabilities are introduced to discard the flow exceeding
the line capacity.
A similar approach to that of effective frequency is adopted in schedule-
 4 Natale Papola, Francesco Filippi, Guido Gentile, Lorenzo Meschini

based models using diachronic graphs (Crisalli, 1999; Nguyen et al., 2001),
which can be reduced to static assignments on space-time networks, where
the congestion may affect the arc costs, but not the travel times that are
incorporated in the graph structure. Although this way it is possible, as in the
static case, to simulate the priority of passengers onboard, a distortion on the
cost pattern is introduced: at the equilibrium the cost for the passengers who
board the arriving run is equal to that suffered by those who must wait at the
stop for a successive run. Moreover, when using this approach a compromise
is to be made between numerical convergence and accuracy in constraint
satisfaction, because, if the waiting cost increases too strongly when the
onboard flow approaches the residual capacity, then the assignment
algorithm becomes unstable. Instead of introducing an asymmetric arc cost
function, Carraresi et. al. (1996) formulate the transit assignment on the
diachronic graph as a multi-commodity flow problem with explicit capacity
constraints. This way they get rid of the numerical instability, but introduce
the questionable behavioural assumption that all the paths whose cost is
within a threshold from the minimum cost are perceived as equivalent by
passengers. Moreover, this approach did not produce applicative tools,
because multi-commodity flow problems are much more cumbersome to
solve with respect to the shortest path problems that pervade all the other
models.
The main advantage of the Dynamic User Equilibrium (DUE) model
presented here is its capability of reproducing correctly the effects of the
vehicle capacity constraints, both in terms of performance pattern and flow
propagation. Our approach is simply to represent these phenomenon within a
suitable arc performance function that yields waiting time temporal profiles
consistent with a FIFO representation of passenger queues for any given
flow pattern. This way we overcome both the numerical instability and the
cost pattern distortion.
The paper is organized as follows. In section 2 we formalize our
modelling framework starting from the data structure of relevant input. In
section 3 the arc performance function is presented, and in particular we
propose the new waiting model. In section 4 a suitable network loading map
is introduced to handle time-continuous flows on the pedestrian network and
time-discrete point-packets on the line network. In section 5 the dynamic
equilibrium model is formulated as a fixed point problem. In section 6 we
address the main algorithmic issues that relate to time discretization, while
convergence is achieved trivially through the Method of Successive
Averages. Finally, in section 7 we present two numerical applications of the
proposed model, the first one an a toy network to discuss some reproducible
results, the second one an a large network to show the potentialities of the
method.
 5 Natale Papola, Francesco Filippi, Guido Gentile, Lorenzo Meschini

2 TRANSIT NETWORK FORMALIZATION

Converting the input data, usually organized in a GIS database, into the
assignment graph handled by the model is an essential operation, which
however is not usually addressed with much detail in the specific literature.
In the following we tribute the proper relevance to this issue, with the aim of
defining the minimum amount of information needed to apply the proposed
schedule-based transit assignment model.

2.1 Input data

The pedestrian network is represented by an undirected graph H = (V, E),

where V⊂ℵ is the set of vertices (ℵ is the set of positive integer numbers),
and E⊆V×V is the set of edges. The set of origins and destinations of
passenger trips, referred to as centroids, is a subset Z⊆V of the vertices.
The generic vertex v∈V is associated with a location in space that can be
accessed by passengers, which is characterized by geographic coordinates
(λv, θv)∈ℜ2. The generic edge (u, v)∈E is then characterized by a length:
L(u,v) = [(λv - λu)2 + (θv - θu)2]0.5.
The set of the edges can either be given explicitly, or implicitly by each
vertex couple (u, v)∈V×V that satisfies at least one among a given set of
connection rules. For example, the following rules have been used in our
numerical applications:
- connect each centroid with the closest non-centroid vertex:
L(u,v) = min{L(u,z): z∈V, z∉Z }, u∈Z ;
- connect each centroid with each other vertex within a threshold αu∈ℜ+
specific to the centroid (ℜ+ is the set of non-negative real numbers):
L(u,v) ≤ αu , u∈Z, v ≠ u ;
- connect each vertex with each other vertex within a threshold α∈ℜ+:
L(u,v) ≤ α , v ≠ u .
The line network is represented by a set ℑ⊆ℵ of lines. The generic line
ℓ∈ℑ is characterized, from a topological point of view, by a sequence of
σℓ∈ℵ stops, referred to as its route, each one corresponding to a different
vertex: R(ℓ) = {rs(ℓ)∈ℵ: s∈[1, σℓ] ⊆ ℵ} ⊆ V .
For any given vertex v∈V and line ℓ∈ℑ, the function s(v, ℓ) yields, if it
exists, the index s such that rs(ℓ) = v, and 0 otherwise.
Line carriers are characterized by a vehicle capacity Qℓ∈ℜ++, and by an
alighting capacity ηℓ∈ℜ++ (ℜ++ is the set of positive real numbers). The
former is the nominal capacity, usually expressed by the number of available
seats, if standing in the carrier is not allowed. On the contrary case, it
 6 Natale Papola, Francesco Filippi, Guido Gentile, Lorenzo Meschini

expresses the maximum number of passengers that can physically fit in the
carrier. The latter expresses the rate at which passengers get off the carrier.
Each line ℓ∈ℑ is operated with a number κℓ∈ℵ of runs. The generic run
k∈[1, κℓ] ⊆ ℵ is characterized at each stop rs(ℓ)∈R(ℓ) by specific arrival
time ATsℓk∈ℜ and departure time DTsℓk∈ℜ. Not every run makes all the
stops of its route and to enhance efficiency we wish to keep as small as
possible the set of the lines. Therefore, a Boolean make-the-stop variable
MSsℓk∈{0,1} is introduced, that is equal to 1, if the k-th run of line ℓ makes
the s-th stop, and to 0 otherwise. These information represent the timetable
of line ℓ, which must satisfy the following consistency constraints (the
arrival and departure times relative to the runs that don’t make a certain stop
can be set arbitrarily within these constraints):
ATsℓk ≤ DTsℓk ≤ ATs+1ℓk, s∈[1, σℓ-1] , k∈[1, κℓ] ,
ATsℓk < ATsℓk+1, DTsℓk < DTsℓk+1 , s∈[1, σℓ] , k∈[1, κℓ-1] .
Regarding the fare schema, we attach to the s-th section of the k-th run of
line ℓ∈ℑ, from the stop rs(ℓ) to the stop rs+1(ℓ), with s∈[1,σℓ-1] and k∈[1,
κℓ], a specific section fare SFsℓk∈ℜ+ so as to obtain purely additive path
costs, which allows implicit path enumeration in route choice computations.
It often happens that one is interested in the main transit network, while
also the secondary transit network is to be taken into account in order to
correctly evaluate the overall performances. To this end, it can work very
well to simulate the secondary network through fast pedestrian edges, whose
speed is assumed equal to the transit operating speed. To this end, we allow
to specify a speed S(u,v)∈ℜ++ for each edge (u, v)∈E.

2.2 Assignment graph

In the following we will convert the above data structure of the

topological input into a directed graph G = (N, A) that represents the formal
transit network handled by the assignment model, where N is the set of the
nodes, and A is the set of the arcs.
The generic node x∈N is identified by an ordered couple, whose first
element is the node line, denoted NL(x) ⊆ -ℑ∪{0}∪ℑ, and the second
element is the node vertex, denoted NV(x) ⊆ V, that is x = (NL(x), NV(x)).
This lets us to distinguish 3 different types of nodes, as depicted in Figure 1:
PN = {(0, v): v∈V} pedestrian nodes;
AN = {(-ℓ, rs(ℓ)): ℓ∈ℑ, s∈[2, … , σℓ] ⊆ ℵ} arrival nodes;
DN = {(+ℓ, rs(ℓ): ℓ∈ℑ, s∈[1, … , σℓ-1] ⊆ ℵ} departure nodes,
so that we have: N = PN ∪ AN ∪ DN .
As usual, the generic arc a∈A is identified by an ordered pair of nodes,
referred to respectively as the tail, denoted TL(a) ⊆ N, and the head, denoted
HD(a) ⊆ N; that is a = (TL(a), HD(a)). As depicted in Figure 1, we
 7 Natale Papola, Francesco Filippi, Guido Gentile, Lorenzo Meschini

distinguish 5 different types of arcs:

PA = {( (0, u) , (0, v) ): (u, v)∈E} pedestrian arcs;
WA = {( (0, rs(ℓ)) , (+ℓ, rs(ℓ)) ): ℓ∈ℑ, s∈[1, σℓ-1] ⊆ ℵ} waiting arcs;
RA = {( (+ℓ, rs(ℓ)) , (-ℓ, rs+1(ℓ)) ): ℓ∈ℑ, s∈[1, σℓ-1] ⊆ ℵ} running arcs;
DA = {( (-ℓ, rs(ℓ)), (+ℓ, rs(ℓ))): ℓ∈ℑ, s∈[2, σℓ-1] ⊆ ℵ} dwelling arcs;
AA = {( (-ℓ, rs(ℓ)), (0, rs(ℓ)) ): ℓ∈ℑ, s∈[2, σℓ] ⊆ ℵ} alighting arcs,
so that we have: A = PA ∪ WA ∪ RA ∪ DA ∪ AA .

(-ℓ, rs(ℓ)) (+ ℓ, rs(ℓ)) (-ℓ, rs+1(ℓ))

(0, rs(ℓ)) (0, rs+1(ℓ))

Pedestrian Node ∈PN Pedestrian Arc ∈PA

Dwelling Arc ∈DA

Arrival Node ∈AN Running Arc ∈RA
Waiting Arc ∈WA
Departure Node ∈DN Alighting Arc ∈AA
Figure 1 - The generic portion of the transit network between two consecutive stops of a line.

Each arc a ∈ A\PA is univocally associated with a line ℓ(a)∈ℑ;

specifically, if a∈AA, then ℓ(a) = NL(TL(a)), otherwise ℓ(a) = NL(HD(a)).
Then we can denote as s(a) = s(NV(TL(a)), ℓ(a)) the index of the associated
route stop.
To be noticed that more than one line may stop at the same pedestrian
node, and that the structure of the pedestrian network can be very simple or
very complex, depending on the modeling choices; Figure 1 does not
illustrate these facts.
 8 Natale Papola, Francesco Filippi, Guido Gentile, Lorenzo Meschini

3 THE ARC PERFORMANCE MODEL

The arc performance function aims at determining the travel time

temporal profile and the generalized cost temporal profile on each arc of the
transit network as a function of the inflow temporal profiles of the adjacent
arcs. By definition, only the waiting and alighting times are congested, due
to queuing, since the running and dwelling times are given by the timetable,
while the pedestrian times are assumed as usual to be fixed. Moreover,
onboard comfort is strictly related to the occupation rate, so that also running
and dwelling costs are congested.
We will consider a time-discrete flow model when referring to running
and dwelling arcs, where all the passengers on board of a run are assumed to
cross any section along the line at the same instant (point-packet). On the
contrary, we will consider a time-continuous flow model when referring to
the pedestrian arcs. Waiting and alighting arcs concentrate continuous flows
into discrete flows and spread discrete flows into continuous flows,
respectively.
Below we introduce the notations for flow and performance variables:
fa(τ) number of passengers that entered arc a∈PA∪WA until time τ;
fa k number of passengers on board of run k∈[1, κℓ(a)] that enter arc
a∈RA∪DA∪AA;
ta(τ), ca(τ) exit time and cost for passengers entering arc a∈PA∪WA at
time τ;
k k
ta , ca exit time and cost for passenger on board of run k∈[1, κℓ(a)] that
enter arc a∈RA∪DA∪AA.
Note that to handle both models in the same framework we express the flow
variables in terms of cumulative temporal profiles and introduce run indexed
vectors only for convenience, since these can actually be represented in the
former form. The first derivative of the generic cumulative flow, when
defined, represents the instantaneous flow and is denoted as f ′a(τ) = ∂fa(τ)/∂τ.
Introducing the functional Γ, the arc performance function can be
expressed, in compact form, as:

[c, t] = Γ(f) , (1)

where the arc components of c, t and f are temporal profiles or run indexed
vectors, depending on the arc type.

3.1 Waiting arcs

Waiting times are affected by the vehicle capacity constraints.

Specifically, at the stop s(a) of line ℓ(a) associated to the generic waiting arc
 9 Natale Papola, Francesco Filippi, Guido Gentile, Lorenzo Meschini

a∈WA, the maximum number of passengers that can actually board on each
run k∈[1, κℓ(a)] is equal to the residual capacity Qℓ(a) - fd(a)k , where if s(a) > 1
d(a) = ( (-ℓ(a), NV(TL(a))) , (+ℓ(a), NV(TL(a))) ) is the corresponding
dwelling arc, otherwise fd(a)k = 0.
To evaluate the effects of the above capacity constraints, we shall
determine for each run k∈[1, κℓ(a)] the instant ρak when the last passenger
that achieves boarding it (or would achieve to do it, in case of null inflows)
enters the waiting arc. By definition the exit time of all passengers that board
run k coincides with the departure time DTs(a)ℓ(a) k, while the arc cost is given
by multiplying the travel time by the value of waiting time μ∈ℜ+. Then we
have:

ta(τ) = DTs(a)ℓ(a) k, k | ρak-1 < τ ≤ ρak ; ta(τ) = ∞, τ > ρaκℓ(a) , (2)

ca(τ) = μ ⋅ (ta(τ) - τ) . (3)

If there was no room for a passenger reaching the stop at time τ to board any
run of the line till the end of the simulation period, then his travel costs
would be infinite; therefore, he will chose a different path.

run k-1 run k run k+1

Qℓ(a)
capacity

fd(a)k
vehicle

DTs(a)ℓ(a) k-1 DTs(a)ℓ(a) k DTs(a)ℓ(a) k+1

waiting

ta(τ) - τ
time

ρak-1 ρak ρak+1

σak
waiting

f ′a(τ)
inflow

time
Figure 2 - Waiting time for given residual capacities and inflow

The instants ρak can be determined recursively following the run order.
 10 Natale Papola, Francesco Filippi, Guido Gentile, Lorenzo Meschini

To this end lets assume ρa0 = -∞. Starting from the instant ρak-1 the
passengers that arrive at the stop later than that willing to ride line ℓ(a) shall
board the successive run k until their number overcomes the residual
capacity Qℓ(a) - fd(a)k, which happens at a specific time denoted σak :

Qℓ(a) - fd(a)k = fa(ρak-1) - fa(σak) , (4)

or the carrier of run k departs from the stop, which happens at time DTs(a)ℓ(a)k:

ρak = min{σak, DTs(a)ℓ(a) k} . (5)

The proposed waiting model satisfies the FIFO rule, but yields
discontinuities in the travel time pattern, although this is coherent with the
real phenomenon. Indeed, the waiting time temporal profile has the saw-
tooth shape depicted in Figure 2, where each run k will be taken by the
passengers that entered the waiting arc during the time interval (ρak-1, ρak].

3.2 Alighting arcs

Strictly speaking, the travel time of run k∈[1, κℓ(a)] at the stop s(a) of line
ℓ(a) associated to the generic alighting arc a∈RA depends on the position of
the passenger in the alighting queue. Therefore, from the first to the last user
in the queue it varies linearly from 0 to the ratio fak/ηℓ between the number of
alighting passengers and the alighting capacity. This way, passengers
starting their trip at the same time from the same stop and travelling along
the same path may reach the same destination at different times.
However, if we should assume such a rigorous formulation to hold also in
determining the arc cost, we would end up with undetermined path costs,
which are highly undesirable from a modelling point of view. Moreover,
from a behavioural point of view, passengers are unlikely to make their path
choice relying on their alighting order. For these reasons we assume a risk
adverse behaviour, such that all the alighting passengers perceive a same
travel time equal to the maximum fak/ηℓ , while the arc cost is given by
multiplying this travel time by the value of alighting time γ∈ℜ+:

tak = fak / ηℓ(a) + ATs(a)ℓ(a) k , (6)

cak = γ ⋅ (tak - ATs(a)ℓ(a) k) . (7)

On the other hand, when propagating alighting passengers on the

pedestrian network, we will spread them uniformly from time ATs(a)ℓ(a) k to
time tak, coherently with actual facts.
 11 Natale Papola, Francesco Filippi, Guido Gentile, Lorenzo Meschini

3.3 Running arcs

The travel time of run k∈[1, κℓ(a)] on the section s(a) of line ℓ(a)
associated to the generic running arc a∈RA is simply given by the difference
between the arrival time ATs(a)+1ℓ(a) k and the departure time DTs(a)ℓ(a) k:

tak = ATs(a)+1ℓ(a) k . (8)

Moreover, we assume that the value of riding time is a linear function of

the occupancy rate fak / Qℓ(a) , whose coefficient ξℓ(a)∈ℜ+ and constant
ζℓ(a)∈ℜ+ are line specific in order to take into account the ergonomic
characteristics of the line carrier :

cak = (ζℓ(a) + ξℓ(a) ⋅ fak / Qℓ(a)) ⋅ (tak - DTs(a)ℓ(a) k) + SFs(a)ℓ(a) k . (9)

The comfort function can be easily generalized by introducing an

exponent of the occupancy rate, so as to take into account the nonlinearity of
the crowding costs. However, in some applications it is important to
reproduce the different discomfort suffered by passengers who have to stand
up with respect to those who have a seat. To this end we can introduce a
fictitious duplicate of each line, whose vehicle capacity is the maximum
number of standing-up passengers that can physically fit in the carrier and
whose coefficient of the comfort function is typically high, while the
capacity constraint of the original line is set equal to the number of seats and
the coefficient of the comfort function can be set to zero. The only drawback
of this representation is that standing-up passengers already on board will
have to alight and board the carrier again in order to take an available seat,
thus loosing any priority with respect to the passengers waiting at the stop,
so that we don’t know which of them would sit down after places become
available.
When the occupancy rate gets close to 1 the crowding discomfort may
become so high that some passengers would not dare boarding, although
there is still some residual capacity on the carrier. The different attitude of
passengers could be well simulated with a multiclass assignment. In the
proposed model we don’t introduce multiclass only for simplicity, and we
will rely on logit random errors in the travel choices to reproduce the
distribution of users’ values of time, included the coefficients of the comfort
function. On the contrary, the longer wait the more the passengers will
accept to board a crowded carrier. The equilibrium mechanism is well
capable of reproducing this phenomenon.
 12 Natale Papola, Francesco Filippi, Guido Gentile, Lorenzo Meschini

3.4 Dwelling arcs

The travel time of run k∈[1, κℓ(a)] at the stop s(a) of line ℓ(a) associated
to the generic dwelling arc a∈DA is simply given by the difference between
the departure time DTs(a)ℓ(a) k and arrival time ATs(a)ℓ(a) k, while the cost is
obtained similarly to (9). Then we have:

tak = DTs(a)ℓ(a) k , (10)

cak = (ζℓ(a) + ξℓ(a) ⋅ fak / Qℓ(a)) ⋅ (tak - ATs(a)ℓ(a) k) . (11)

3.5 Pedestrian arcs

The travel time on the generic pedestrian arc a∈PA is simply given by the
ratio between the length Lb(a) of the edge b(a) = (NV(TL(a)), NN(HD(a)))
associated to a and its speed Sb(a). The arc cost is given by multiplying the
travel time by the value of walking time ω∈ℜ+. Then we have:

ta(τ) = τ + Lb(a) / Sb(a) , (12)

ca(τ) = ω ⋅ (ta(τ) - τ) . (13)

4 THE NETWORK LOADING MAP

The network loading map is complementary to the arc performance

function in the sense that it aims at determining the inflow temporal profiles
as a function of the travel time temporal profiles and of the generalized cost
temporal profiles.
In the following we present an extension of the model proposed in Bellei
et al. (2005) for DTA on road networks to deal with time-discrete flows on
the line network and time-continuous flows on the pedestrian network. Thus,
we consider a logit route choice model, which has the desirable feature of
spreading more realistically the passenger flows on the different supply
elements. However, nothing prevents from adopting in this framework a
deterministic or a probit route choice model based on dynamic shortest path
computations, as it is done in Gentile and Meschini (2006) for the case of
traffic assignment.

4.1 Route choice

Dealing with implicit path enumeration we have to introduce:

 13 Natale Papola, Francesco Filippi, Guido Gentile, Lorenzo Meschini

- the node satisfaction, which is the opposite of the expected value of

the minimum perceived cost to reach the destination being on that
node at a given instant;
- the arc conditional probability, which is the probability of choosing
that arc to continue the trip towards the destination, being on its tail at
a given instant.
To be noticed that, while it is possible to be on a pedestrian node at any
instant τ of the simulation period, with reference to arrival and departure
nodes this is possible only in correspondence of the scheduled runs.
Therefore, we can formalize the arc conditional probabilities and node
satisfactions as follows:
pad(τ) conditional probability of arc a∈PA∪WA for passengers directed
to destination d∈Z and being on node TL(a) at time τ;
pad k conditional probability of arc a∈RA∪DA∪AA for passengers
directed to destination d∈Z and being at TL(a) on run k∈[1, κℓ(a)];
wxd(τ) satisfaction of node x∈PN at time τ to reach destination d∈Z;
wxd k satisfaction of node x∈AN∪DN on run k∈[1, κNL(x)] to reach
destination d∈Z.
As in any Dial-like model we assume that passengers travel only on
efficient arcs, that is they always near the destination with respect to a given
node topological order Txd, with x∈N, d∈Z. Let then FSE(x, d) = {a∈A:
TL(a) = x, Txd > THD(a)d} and BSE(x, d) = {a∈A: HD(a) = x, TTL (a)d > Txd} be
the efficient forward and backward star of node x with respect to destination
d, respectively.
Based on the results achieved in the referred paper it is possible to
express the node satisfactions through the following recursive equations,
where θ∈ℜ++ is a parameter to be calibrated:

⎛ ⎛ −ca (τ) + wHD( a ) d (ta (τ)) ⎞ ⎞

⎜
⎜
∑ exp ⎜⎜
θ
⎟⎟ + ⎟
a∈FSE ( x , d ) ∩ PA ⎝ ⎠ ⎟
wxd (τ) = θ⋅ ln ⎜ ⎟
⎜ ⎛ −ca (τ) + wHD ( a ) d k ⎞ ⎟ (14.1)
⎜⎜ + ∑ exp ⎜
⎜ θ
⎟⎟ ⎟⎟
⎝ a∈FSE ( x,d )∩WA ⎝ ⎠ ⎠
k : ρa k −1 < τ ≤ ρa k , x ∈ PN \ (0, d ), w(0,
d
d ) (τ) = 0

wx d k = −ca k + wHD ( a ) d k
(14.2)
k ∈ [1, κ A ( a ) ], a = FSE ( x, d ), x ∈ DN
 14 Natale Papola, Francesco Filippi, Guido Gentile, Lorenzo Meschini

⎛ ⎛ −ca k + wHD ( a ) d k ⎞ ⎛ −cb k + wHD (b ) d (tb k ) ⎞ ⎞

wxd k = θ⋅ ln ⎜ exp ⎜ ⎟ + exp ⎜ ⎟⎟ ⎟
⎜ ⎜ θ ⎟ ⎜ θ ⎟ (14.3)
⎝ ⎝ ⎠ ⎝ ⎠⎠
k ∈[1, κA ( a ) ], a = FSE( x, d ) ∩ DA, b = FSE( x, d ) ∩ AA, x ∈ AN

The conditional probabilities are then given by the following equations:

⎛ −ca (τ) + wHD ( a ) d (ta ( τ)) − wTL ( a ) d (τ) ⎞

pa d ( τ) = exp ⎜ ⎟⎟ a ∈ PA (15.1)
⎜ θ
⎝ ⎠

⎛ −ca (τ) + wHD(a)d k − wTL(a)d (τ) ⎞ k −1

pa (τ) = exp ⎜
d
⎟⎟ k : ρa < τ ≤ ρa , a ∈WA (15.2)
k
⎜ θ
⎝ ⎠

pa d k = 1 k ∈[1, κA( a) ], a ∈ RA (15.3)

⎛ −ca k + wHD( a ) d k − wTL( a )d k ⎞

pa d k = exp ⎜ ⎟⎟ k ∈[1, κA( a ) ], a ∈ DA (15.4)
⎜ θ
⎝ ⎠

⎛ −ca k + wHD ( a ) d (ta k ) − wTL ( a ) d k ⎞

pa d k = exp ⎜ ⎟⎟ k ∈[1, κA ( a ) ], a ∈ AA (15.5)
⎜ θ
⎝ ⎠

When in equations (14.1) and (15.2) there is no run k satisfying the required
property, then the corresponding satisfaction term shall be considered equal
to -∞.
Since users choose only efficient paths, the system of equations (14) can
be solved by processing the nodes in topological order, while time instants
and runs may be processed in any order for each node. Then we express its
solution in compact form through the following functional:

w = w(c, t) . (16)

Equation (15) can also be expressed in compact form by a functional:

p = p(w, c, t) . (17)

The node components of w and the arc components of p are temporal

profiles or run indexed vectors, depending on the arc and node type.
 15 Natale Papola, Francesco Filippi, Guido Gentile, Lorenzo Meschini

4.2 Departure time choice

Each passenger may anticipate or delay his departure time from the origin
in order to benefit from an higher satisfaction to reach his destination.
However, in doing this he suffers a disutility, which is here assumed to be
proportional to the advance or delay, respectively.
Following Bellei et al. (2006) we reproduce this travel decision through a
logit model such that the generic passenger desiring to depart at time τ from
the origin o∈Z toward the destination d∈Z has a continuous choice set of
alternatives from time τ-δA to time τ+δD, where δA∈ℜ+ and δD∈ℜ+ denote
the maximum feasible advance and delay, respectively. The systematic
utility ϖod(σ/τ) of the actual departure time σ∈[τ-δA ,τ+δD], that constitutes
the generic alternative, is assumed to be equal to:

ϖod(σ/τ) = w(0,o)d(σ) - max{βA ⋅ (τ-σ), βD ⋅ (σ-τ)} , (18)

where βA∈ℜ+ and βD∈ℜ+ denote the value of departure advance and delay.
On this basis, the departure time choice produces a transformation of the
desired demand flow temporal profile into an actual demand flow temporal
profile. Let Dod(τ)∈ℜ+ and Φod(τ)∈ℜ+ be the desired and the actual
cumulative demand flow at time τ, respectively. In the referred paper it is
proved that the above assumptions lead to the following formulation, where
ϑ∈ℜ++ is a parameter to be calibrated:

⎛ ϖod (τ / σ) ⎞
τ+δA exp ⎜ ⎟
⎝ ϑ ⎠
Φo '(τ) = ∫ Do '(σ) ⋅ σ+δD
d d
⋅ dσ . (19)
⎛ ϖ d
(λ / σ) ⎞
∫ exp ⎜⎝ ϑ ⎟⎠ ⋅ dλ
τ−δD o

σ−δA

Since the network performance patter is strictly related to the timetable,

the temporal profiles of the node satisfactions are patently discontinuous.
Then the departure time choice model tends to concentrate the desired
demand flows, since passengers can coordinate their departure from the
origin with the departures of the runs at the stops. Nevertheless, the above
logit formulation is capable of preserving the continuity of the cumulative
flow temporal profile in the transformation from the desired demand to the
actual demand. This is essential to let us adopting a time-continuous flow
model on the pedestrian network.
Equations (19) is expressed in compact form by the following functional:

Φ = Φ(w; D) . (20)
 16 Natale Papola, Francesco Filippi, Guido Gentile, Lorenzo Meschini

4.3 Network flow propagation

To formulate the network flow propagation model, it is useful to

introduce cumulative inflow and outflow variables referred to passengers
travelling toward a specific destination d∈Z:
fad(τ) number of passengers that entered arc a∈PA∪WA until time τ;
fa d k number of passengers on board of run k∈[1, κℓ(a)] that enter arc
a∈RA∪DA∪AA.
ead(τ) number of passengers that exited arc a∈PA∪AA until time τ;
ead k number of passengers on board of run k∈[1, κℓ(a)] that exit arc
a∈WA∪RA∪DA.
The instantaneous flow entering a given arc a∈A at time τ is equal to its
conditional probability multiplied by the instantaneous flow exiting from its
tail TL(a); the latter flow is given, in turn, by the sum of the instantaneous
outflows from the backward star of TL(a), and of the actual demand flow
from TL(a) to d, which is null when TL(a)∉Z. Then, for the different arc
types of the transit network we have:

f ′ad(τ) = pad(τ) ⋅ [Φ′TL(a)d(τ) + ∑ b∈BSE(TL(a), d) e′bd(τ)] a∈PA∪WA (21.1)

fad k = ∑ b∈BSE(TL(a), d) ebd k k∈[1, κℓ(a)], a∈RA (21.2)

fad k = pad k ⋅ ebd k k∈[1, κℓ(a)], b = BSE(TL(a), d), a∈DA∪AA (21.3)

The outflows for the different arc types are determined as follows:

e′ad(τ) = f ′ad(τ - Lb(a) / Sb(a)) a∈PA (22.1)

ead k = fad(ρak) - fad(ρak-1) k∈[1, κℓ(a)], a∈WA (22.2)

ead k = fad k k∈[1, κℓ(a)], a∈RA∪DA (22.3)

e′ad(τ) = ηℓ(a) ⋅ fad k / fak k: τ∈[ATs(a)ℓ(a) k, tak], a∈AA (22.4)

The arc cumulative inflow is simply given by the sum of the cumulative
flows relative to all the destinations:

fa(τ) = ∑ d∈Z -∞∫τ f ′ad(σ)⋅dσ a∈ PA∪WA (23.1)

fak = ∑ d∈Z fad k k∈[1, κℓ(a)], a∈RA∪DA∪AA (23.2)

Since users choose only efficient paths, the system of equations (21), (22)
 17 Natale Papola, Francesco Filippi, Guido Gentile, Lorenzo Meschini

and (23) can be solved by processing the nodes in reverse topological order,
while time instants and runs may be processed in any order for each node.
Then we express its solution in compact form through the following
functional:

f = f(p, t, Φ) . (24)

5 THE DYNAMIC USER EQUILIBRIUM MODEL

Extending to the dynamic-stochastic case Wardrop’s first principle, DTA

is here regarded as an equilibrium where no user can reduce his perceived
travel cost by unilaterally changing path, under the assumption that the path
cost is that actually experienced by the passenger while travelling on the
network consistently with time-varying travel times and generalized costs.
The formulation based on implicit path enumeration of the DUE model is
synthetically depicted in Figure 3, which immediately highlights the
possibility of formulating the model as a fixed point problem in terms of the
cumulative arc inflow temporal profiles f.

network loading map

D p(w, t, c)

Φ(w; D) w

Φ p

f(p, t, Φ) w(c, t)

f t c

Γ(f)

arc performance model

Figure 3 - Formulation of the Dynamic User Equilibrium with implicit path enumeration.

Specifically, combining the route choice model (16)-(17) and the

departure time choice model (20) with the flow propagation model (24)
yields the formulation of the logit network loading map based on implicit
 18 Natale Papola, Francesco Filippi, Guido Gentile, Lorenzo Meschini

path enumeration. Then, combining the latter with the arc performance
function (1) we have:

f = f(p(w(Γ(f)), Γ(f)), Γ(f), Φ(w(Γ(f)); D)) . (25)

Note that the vehicle capacity constraints will be satisfied only at the
equilibrium, since they affect directly only the arc cost function, while the
network flow propagation is transparent to them.

6 SOLUTION ALGORITHM

To implement the proposed model, the simulation period is divided into n

time intervals identified by the sequence of instants τ = {τi∈ℜ: i∈[1, n]⊆ℵ},
with τi < τj for any 0 ≤ i < j ≤ n.
In the following we approximate the generic temporal profile g(τ)
through a piecewise linear function, defined by the values gi = g(τi) taken at
each instant τi∈τ. Then, for τ∈(τi-1, τi], with i∈[1, n], we have:

g(τ) = gi-1 + (τ - τi-1) ⋅ (gi - gi-1) / (τi - τi-1) . (26)

This way, the generic temporal profile g(τ) can be represented numerically
through the (1 × n+1) row vector g = (g0, … , gi, … , gn).
Note that the generic instantaneous flow, being the derivative of a
piecewise linear cumulative flow temporal profile has a piecewise constant
temporal profile.
Time discretization is always a crucial compromise between accuracy and
efficiency. We don’t have explicit limitations to satisfy in our model.
However, if we expect reliable results relative to the passenger loads on each
run, which is usually the case in schedule-based transit assignment, we
cannot avoid introducing at least one instant τi∈τ between each couple of
consecutive runs of a same line (or of the same group of common lines)
departing from a stop.

6.1 Arc performances

Given the cumulative inflows f, the computation of the arc exit times and
costs is straightforward, except for waiting times.

function Γ(f)
for each arc a∈PA
for each instant τi∈τ
 19 Natale Papola, Francesco Filippi, Guido Gentile, Lorenzo Meschini

compute tai and cai based on (12)-(13)

for each arc a∈RA∪DA∪AA
for each run k∈[1, κℓ(a)]
compute tak and cak based on (6)-(11)
for each arc a∈WA
for each instant τi∈τ , tai = ∞ , cai = ∞
i = 0, F = 0
for each run k∈[1, κℓ(a)] in the natural order
F = F + Qℓ(a) - fd(a)k
until fai > F or τi > DTs(a)ℓ(a) k do
tai = DTs(a)ℓ(a) k , cai = μ ⋅ (tai - τi)
i = i+1
if τi > DTs(a)ℓ(a) k
ϕ = fai-1 + ( fai - fai-1) ⋅ (DTs(a)ℓ(a) k - τi-1) / (τi - τi-1)
if ϕ < F then F = ϕ

The exit times of the generic waiting arc a∈WA are computed in
chronological order, by examining which passengers achieve boarding each
run one after the other. Specifically, let F be the number of passengers that
boarded the runs preceding the current one, denoted k, plus the residual
capacity of the latter. Until the cumulative inflow fai is higher than F or τi is
later than DTs(a)ℓ(a) k, the exit time at τi is DTs(a)ℓ(a) k. When one of these two
conditions is met a new run is considered. Before iterating, we shall check
whether not all the residual capacity was used, and in this case update F to
ϕ = fa(DTs(a)ℓ(a) k).

6.2 Network loading

Given the exit times t and the costs c, also the computation of the node
satisfactions and of the arc conditional probabilities is straightforward.

functions w(c, t) and p(w, c, t)

for each destination d∈Z
for each node x∈N in increasing topological order Txd
for each instant τi∈τ or run i∈[1, κNL(x)]
depending on the node type
compute wxd i based on (14)
for each efficient arc a∈FSE(x, d)
compute pad i based on (15)

Note that, by processing nodes in topological order, when wxd i and pad i
are computed the node satisfactions relative to the arcs belonging to the
 20 Natale Papola, Francesco Filippi, Guido Gentile, Lorenzo Meschini

efficient forward star have already been determined.

One of the two relevant algorithmic issues in this computation is the
evaluation through equation (26) of the satisfaction relative to pedestrian
nodes at specific exit times which are not necessarily included in τ. To this
end, for the generic pedestrian arc a∈PA, we define Ψai to denote the time
index j such that tai∈(τj-1, τj], τi∈τ. Analogously, for the generic alighting arc
a∈AA, we define Ψak to denote the time index j such that tak∈(τj-1, τj], k∈[1,
κℓ(a)]. The second issue is related to the identification, for the generic waiting
arc a∈WA, of the run index k such that τi∈(ρak-1, ρak], τi∈τ. We extend the
notation Ψai to denote this index with reference to waiting arcs. The indices
Ψ can be easily determined within the arc performance function.
With reference to the departure time choice, the computation of the
integral (19) relative to the generic o-d couple, with o∈Z and d∈Z, reduces
here to a series of logit models, one for each desired departure time interval
(τi-1,τi], τi∈τ\τ0. There we compute the contribution of Dod i-Dod i-1 to the
actual demand Φod j of each feasible alternative (τj-1, τj], whose systematic
utility is taken as ϖod(τj/τi).
Time discretization affects the network flow propagation more than any
other procedure. Indeed, equation (21.1) refers to instantaneous flows at a
given instant which shall here be handled as number of passengers relative to
a given interval, as follows:
fad i = fad i-1 + pad i ⋅ [(ΦTL(a)d i - ΦTL(a)d i-1) + ∑ b∈BSE(TL(a), d) (ebd i - ebd i-1)] .
It is worth noting that both the departure time choice and the route choice
relative to the time interval (τi-1,τi] are based on the cost pattern at time τi, as
we considered, respectively, ϖod(τj/τi) and pad i. This is due to the fact that in
general passengers shall perceive the effects of their choices.
To propagate the inflow to HD(a) we need the share Ωa j/i of inflow
relative to i that falls in j, where i and j denote a time interval or a run
depending on the arc type, which shall also be computed within the arc
performance function. Then, based on equations (22.1), (22.2) and (22.4) the
contribution to the outflow component j is respectively:
( fad i - fad i-1) ⋅ Ωa j/i , j∈[Ψai-1, Ψai] , Ωa j/i = [τj-1, τj]∩[tai-1, tai] / [tai-1, tai]
fad i ⋅ Ωa j/i , j∈[Ψai-1, Ψai] , Ωa j/i = [ρa j-1, ρad j]∩[τi-1, τi] / [τi-1, τi]
fad i ⋅ Ωa j/i , j∈[h, Ψai] , Ωa j/i = [τj-1, τj]∩[ATs(a)ℓ(a) i, tai] / [ATs(a)ℓ(a) i, tai]
where h: ATs(a)ℓ(a) i∈(τh-1, τh]. We have used again the indices Ψ to determine
which outflow components j are affected by the inflow component i.

function f(p, t, Φ)
f=0
for each destination d∈Z
fd = 0, ed = 0
for each node x∈N in reverse topological order
 21 Natale Papola, Francesco Filippi, Guido Gentile, Lorenzo Meschini

for each instant τi∈τ or run i∈[1, κNL(x)]

depending on the node type
for each efficient arc a∈FSE(x, d)
compute fad i based on (21)
propagate this inflow to HD(a) computing
its contribution to each ead j based on (22)
d
f=f+f

Note that, by processing nodes in reverse topological order, when fad i is

computed the outflows relative to the arcs belonging to the efficient
backward star have already been determined.

6.3 Equilibrium

The DUE, expressed as a fixed point problem, can be solved through the
following MSA algorithm, where ε∈ℜ+ and mmax∈ℵ are, respectively, the
maximum cumulative flow difference and the maximum number of
iterations.

function DUE
m = 0, f = 0, y = ∞ initialization
until ||f-y||∞ < ε or m > mmax do stop criterion
m=m+1 new iteration
[c, t] = Γ(f) arc performance function
w = w(c, t) node satisfactions
p = p(w, c, t) arc conditional probabilities
Φ = Φ(w; D) departure time choice
y = f(p, t, Φ) network loading map
f = f + 1/m ⋅ (y - f) update arc flows with MSA

7 NUMERICAL APPLICATIONS

The first application presented here aims at testing the effectiveness of

the proposed model in reproducing the formation and dispersion of queues at
transit stops, due to the vehicle capacity constraints, and the consequent
delay suffered by passengers. To this end we consider the simple network
represented in Figure 4, consisting in a single line with two sections serving
3 stops and operated with 4 runs. The 2 hours of simulation, from 7:00 to
9:00, are divided into 120 time intervals of 60 seconds. Two DTA where
performed; the first one without vehicle capacity constraints, the second one
with a vehicle capacity Q equal to 80 passengers.
 22 Natale Papola, Francesco Filippi, Guido Gentile, Lorenzo Meschini

transit network timetable

run stop time
A B
1 2 3 1 1 07:05
1 2 07:10
1 3 07:14
2 1 07:20
2 2 07:25
2 3 07:29
demand 3 1 07:35
orig dest from to flow [pax/h] 3 2 07:40
1 3 07:00 07:30 200 3 3 07:44
1 3 07:30 07:50 100 4 1 07:50
2 3 07:00 07:30 200 4 2 07:55
2 3 07:30 07:50 100 4 3 07:59
Figure 4 - Test application

Figure 5 shows the onboard flows obtained with the two DTA; in the first
case the number of passengers that board on the second run at stop 2 exceeds
the residual capacity, while in the second case, as expected, the exceeding
users are forced to wait for the third run. This way also the third run
becomes full, so that some users are forced to wait for the fourth run.

onboard passengers [n°] - WITHOUT capacity constraint onboard passengers [n°] - WITH capacity constraint
100 100
run 1
80 run 2 80
run 3
60 60
run 4
40 40
Q
20 20

0 0
A B A B

section section

Figure 5 - Onboard flows without and with the vehicle capacity constraint.

This correct behaviour of the model is confirmed by the waiting time and
queue temporal profiles at stop 2 reported in Figure 6. In particular, the
passengers arrived at the stop between 7:19 and 7:25, which in the first case
could board the second run, in the second case have to wait for the third run,
since the residual capacity on the second run is used by the passengers
arrived between 7:10 and 7:19. Note that the latter users can still board the
same run, and thus their waiting time does not change.
 23 Natale Papola, Francesco Filippi, Guido Gentile, Lorenzo Meschini

stop 2 - waiting time [sec] stop 2 - passengers in queue [n°]

2000 60

1600 WITHOUT 50
capacity
40
1200 constraint
30
800 WITH
capacity 20
constraint
400 10

0 0
07.00 07.10 07.20 07.30 07.40 07.50 08.00 07.00 07.10 07.20 07.30 07.40 07.50 08.00

Figure 6 - Waiting time and queue temporal profiles.

Moreover, the model correctly represents the priority of passengers

onboard. In fact, the users boarded at stop 1 are not affected by those willing
to board at stop 2, and the travel time temporal profile from node 1 to node
3, represented on the left side of Figure 7, is the same for the two cases.

OD (1-3) travel time [sec] OD (2-3) travel time [sec]

1500 1800

1200 1500
WITHOUT
capacity 1200
900 constraint
900
600 WITH
capacity 600
constraint
300 300

0 0
07.00 07.10 07.20 07.30 07.40 07.50 08.00 07.00 07.10 07.20 07.30 07.40 07.50 08.00

Figure 7 - OD travel time temporal profiles.

In order to test the efficiency of the proposed model and its applicability
to real instances, we performed a DTA on the regional transit network of
Rome. This very large network, depicted in Figure 8, consists of 3,963
transit lines (3,879 buses and 84 railways) with 45,682 stops (44,764 bus
stops and 918 railway stations) served during a work day by 10,447 runs.
The corresponding assignemnt graph consists of 122,747 arcs and 51,293
nodes. The demand consists of 288,007 morning trips to work or school and
is represented through 854 traffic zones, on which basis are defined 4 hourly
origin-destination matrices (06:00-07:00; 07:00-08:00; 08:00-09:00; 09:00-
10:00). The 5 hours of simulation, from 06:00 to 11:00 AM, are divided into
60 time intervals of 5 minutes.
Again, two DTA where performed without and with vehicle capacity
constraints, assuming as stop criterion mmax = 20. The computation, carried
out on a 3.0 GHZ PCU with 2 GB of RAM, in the first case required 18.7
minutes and just one iteration, since no congestion affects the travel choices,
while in the second case required 379 minutes, that is about 18.9 minutes for
each iteration, showing that the calculation speed is independent of
congestion.
 24 Natale Papola, Francesco Filippi, Guido Gentile, Lorenzo Meschini

Figure 8 - Number of passengers that cannot board an arriving carrier of the chosen line due
to the vehicle capacity constraints at the stops of the regional transit network of Rome.

To discuss the accuracy of the two models we have measured the

maximum occupancy rate, that is:
max{fak/Qℓ(a): a∈RA, k∈[1, κℓ(a)]},
obtaining in the two cases 22.96 and 3.09, and the average oversaturated
occupancy rate, that is:
1/|ORAK| ⋅ ∑ (a, k)∈RA* fak/Qℓ(a) ,
where ORAK = {(a, k)∈RA×[1, κℓ(a)]: fak > Qℓ(a)}, obtaining in the two cases
3.09 and 1.33. The fact that both indicators are significantly higher than 1 in
the first case, means that the capacity constraints are indeed highly effective
in this application, which unfortunately reflects reality. This is confirmed by
Figure 8, showing at each stop the number of passengers that cannot board
an arriving carrier of the chosen line. Moreover, the fact that in the second
case both indicators are consistently reduced towards the ideal value of 1,
means that our model is able to reproduce the effects of the capacity
constraints. However, 20 iterations were not enough to achieve a perfect
satisfaction of the constraints.

REFERENCES
1. Bellei G., Gentile G., Papola N. (2003) Assegnazione alle reti multimodali in presenza di
congestione. In Metodi e Tecnologie dell’Ingegneria dei Trasporti. Seminario 2001, ed.s
G. Cantarella and F. Russo, Franco Angeli, Milano, Italy, 117-134.
 25 Natale Papola, Francesco Filippi, Guido Gentile, Lorenzo Meschini

2. Bellei G., Gentile G., Papola N. (2005) A within-day dynamic traffic assignment model
for urban road networks. Transportation Research B 39, 1-29.
3. Bellei G., Gentile G., Meschini L., Papola N. (2006) A demand model with departure time
choice for within-day dynamic traffic assignment. To be published in European Journal
of Operation Research.
4. Carraresi P., Malucelli F. and Pallottino S. (1996) Regional mass transit assignment with
resource constraints. Transportation Research B 30, 81-98.
5. Cascetta E. (2001) Transportation systems engineering: theory and methods, Kluwer
Academic Publisher, 384-386.
6. Cristalli U. (1999) Dynamic transit assignment algorithms for urban congested networks.
In Urban Transport and the Environment for the 21st century V, ed. L.J. Sucharov,
Computational Mechanics Publications, 373-382.
7. DeCea J. and Fernandez E. (1993) Transit assignment for congested public transport
systems: an equilibrium model. Transportation Science 27, 133-147.
8. Gentile G., Meschini L., Papola N. (2003) Un modello logit di assegnazione dinamica
intraperiodale alle reti multi modali di trasporto urbano. In Metodi e Tecnologie
dell’Ingegneria dei Trasporti. Seminario 2001, ed.s G. Cantarella and F. Russo, Franco
Angeli, Milano, Italy, 33-55.
9. Gentile G. and Meschini L. (2006) Fast heuristics for the continuous dynamic shortest
path problem in traffic assignment. Submitted to Networks.
10. Hickman M. and Bernstein D. (1997) Transit service and path choice models in stochastic
and time-dependent networks. Transportation Science 31, 129-146.
11. Kurauchi F., Bell M. and Schmocker J.-D. (2003) Capacity constrained transit assignment
with common lines. Journal of Mathematical Modelling and Algorithms 2, 309-327.
12. Nielsen O. and Jovicic G. (1999) A large scale stochastic timetable-based transit
assignment model for route and sub-mode choices. Proceedings of 27th European
Transportation Forum, Seminar F, Cambridge, England, 169-184.
13. Nielsen O. (2000) A stochastic transit assignment model considering differences in
passengers utility functions. Transportation Research B 34, 377-402.
14. Nguyen S. and Pallottino S. (1988) Equilibrium traffic assignment for large scale transit
networks. European Journal of Operational Research 37, 176-186.
15. Nguyen S., Pallottino S. and Malucelli F. (2001) A modelling framework for the
passenger assignment on a transport network with time-tables. Transportation Science 35,
238-249.
16. Nuzzolo A. and Russo F. (1993) Un modello di rete diacronica per l’assegnazione
dinamica al trasporto collettivo extraurbano. Ricerca Operativa 67, 37-56.
17. Nuzzolo A., Russo F. and Crisalli U. (2003) Transit network modelling. The schedule-
based dynamic approach. Collana Trasporti, Franco Angeli, Milan, Italy.
18. Spiess H., Florian M. (1989) Optimal strategies: a new assignment model for transit
networks. Transportation Research B 23, 83-102.
19. Wilson N. and Nuzzolo A. (ed.s) (2004) Schedule-based dynamic transit modelling:
theory and applications. Kluwer Academic Publishers.
20. Tong C. and Wong S. (1999) A stochastic transit assignment model using a dynamic
schedule-based network. Transportation Research B 33, 107-121.