Advanced pricing and rationing policies

for large scale multimodal networks

GUIDO GENTILE, NATALE PAPOLA, LUCA PERSIA

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

Via Eudossiana, 18 – 00184 Roma, [guido.gentile, natale.papola, luca.persia]@uniroma1.it

Abstract

The applying of simplified schemes, such as cordon pricing, as second-best solution to the toll

network design problem is investigated here in the context of multiclass traffic assignment on

multimodal networks. To this end a suitable equilibrium model has been developed, together with

an efficient algorithm capable of simulating large scale networks in quite reasonable computer

time. This model implements the theoretical framework proposed in a previous work on the toll

optimization problem, where the validity of marginal cost pricing for the context at hand is stated.

Application of the model to the real case of Rome shows us, not only that on multimodal networks a

relevant share (up to 20%) of the maximum improvements in terms of social welfare achievable

with marginal cost pricing can in fact be obtained through cordon pricing, but also that in practical

terms rationing is a valid alternative to pricing, thus getting around some of the relevant questions

(theoretical, technical, social) the latter raises. As a result we propose a practical method to

analyze advanced pricing and rationing policies differentiated for user categories, which enables us

to compare alternative operative solutions with an upper bound on social welfare based on a solid

theoretical background.

Keywords: cordon tolls, rationing, marginal cost pricing, multiclass traffic assignment, multimodal

networks, social welfare.

1
 1 INTRODUCTION

Urban road pricing is currently seen as one of the most powerful tools in managing transport

demand in densely populated areas, in order to reduce traffic congestion and externalities. At

present, however, there are very few examples of large scale applications world-wide (e.g. London

and Singapore, where cordon-based policies have been adopted) leaving many uncertainties over

features and conditions for successful implementation.

Whereas, rationing policies (i.e. access restrictions to particular areas during given hours) have

been applied for many years in a large number of cities, proving to be very effective in reducing

traffic flows in the said areas and increasing the speed and reliability of public transport vehicles

(Filippi, Persia et al., 1996). Some counteracting effects can arise. Congestion and lack of parking

in surrounding neighborhoods may increase, thus inducing unauthorized vehicles to disregard

access restriction. Moreover, in some Italian cities this measure led to a dramatic explosion of

mopeds, with very negative impact on the environment and safety. These questions and the need to

finance the development and maintenance of the transit system have led local authorities to inquire

into cordon pricing policies, where different fare structures are applied to different categories of

users.

In this study we present a methodology simulating on the one side the impact on the transport

system of advanced policies for managing travel demand based on simplified schemes, such as

cordons and barriers, where the appropriate categories of users and the corresponding optimum tolls

are determined using a “what if” approach, then to evaluate on the other side their effectiveness

compared with Marginal Cost Pricing (MCP), which yields an upper bound on the social welfare.

Other significant issues (e.g. fairness, acceptability, legal and institutional barriers, economic

impacts) are not addressed in this paper.

This methodology is based on a specification of the fixed point equilibrium model developed in

Bellei, Gentile and Papola (2002) and aims at making operative the great descriptive potential,

2
expressed there only in formal terms, of simulating the desired urban scenarios through a multiclass

traffic assignment on multimodal networks with elastic demand and non-separable arc cost

functions. The result obtained there and used here is that, subject to the possibility of charging each

arc of the network any real valued toll, a solution to the toll Network Design Problem (NDP) exists,

and coincides with a System Optimum (SO) where user travel choices can be prescriptively

controlled, and determined by calculating a System Equilibrium (SE). This in turn is defined by

replacing in the User Equilibrium (UE) the cost function with the marginal social cost function.

In our framework a stochastic approach based on random utility theory is used for modeling

travel demand with a hierarchical choice structure. Thus, to ensure the validity of the MCP principle

we shall maximize the social welfare (i.e. user surplus, plus toll revenues, minus external costs) as

objective function, which involves user satisfaction, coinciding with their surplus, instead of

(minus) their actual travel costs, coinciding with user satisfaction only in the deterministic case.

In applying this methodology to the real case of Rome, we will show how in a multimodal

context a relevant share of the maximum improvements in social welfare achievable with MCP can

be obtained through cordon pricing and that rationing is a valid alternative. Interestingly, May et.

Al. (2002), Verhoef (2002 b) and Zhang and Yang (2004), all of whom addressed this subject using

different approaches, drew similar conclusions.

The paper is organized as follows. In section 2, we present a discussion about pricing and

rationing policies in the context of multimodal networks. In section 3, we propose an equilibrium

model capable of catching the relevant aspects of multimodal congestion and elastic demand, and

provide a solution for the corresponding toll optimization problem based on MCP. In section 4, we

present some numerical results relative to the case study of Rome and discuss them.

2 ADVANCED PRICING AND RATIONING POLICIES

Pricing and rationing policies generally have very similar aims: to reduce private traffic in

specific areas (usually city center), mitigating environmental impact and improving public transport
3
performance (speed and reliability) and appeal.

Under rationing, automobile access to portions of the road network is allowed only to users with

specific needs (e.g. public utility, residence, work, the disabled) assessed by the Public

Administration. While, on the contrary, under pricing, passage through some arcs of the network is

permitted only to those users willing to pay. Consequently, pricing also generates income for the

Public Administration, revenue that to a large part (a portion is used in collection costs) can be

reinvested to boost the supply of the transit system and, in general, of the more sustainable modes

of transport.

Note that, while road pricing is meant to price out those users whose willingness to pay is just

marginally above the equilibrium cost, rationing policies deny access in given areas to some groups

of users regardless of their marginal utility. Because in practical terms it is extremely difficult to

identify the groups of users marginally benefiting from the access, inevitably some users with high

surplus will also be forced out of the system, and thus rationing will result in a sub-optimal

scenario.

In both cases, there are sizable obstacles to the actual implementation of such policies, the first

among them being those related to “fairness” and “acceptability” in public opinion.

Leaving aside deeper considerations on these aspects, it is relatively easy to note how the theme

of equity and fairness is brought up mainly by the introduction of pricing policies. While, in fact,

these determine a selection “by census” of those with authorized access, in the case of rationing

access is granted only in cases of local necessity (residents, the disabled, specific economic

activities) or public utility services.

Equally complex is the theme of acceptability, which involves, in addition to the above

considerations of fairness, a series of other factors, including economy (payment of the tariff,

effects on local economic activities) and behavior (need to modify trip habits). All these imply that

the reactions to these two policies vary according to different groups of users.

4
 Figure 1 illustrates the results of a survey conducted in the center of Rome to evaluate how

acceptable each policy was. In general, rationing was more acceptable than pricing. Residents favor

the system of restriction and control of access more than shop owners, while substitution of

rationing with a pure road pricing scheme would be more acceptable to shop owners.

Rationing is a good idea Substitution of rationing with pricing is useful

60% 50%
Residents Residents
50% 40%
Shop owners Shop owners
40%
30%
30%
20%
20%
10% 10%

0% 0%
Definitely Possibly Just a little Not at all Definitely Possibly Just a little Not at all
bit bit

Figure 1. Acceptability of rationing and pricing in Rome.

This difference in opinion also rises from the difference in efficacy of the two approaches.

Rationing policies drastically decrease vehicle congestion. Such reductions would be very unlikely

achieved with pricing policies (unless tariffs were made unacceptably high). This fact is obviously

appreciated by the residents. Pricing policies, however, offer greater flexibility than rationing,

having a smaller impact on accessibility of the area, which makes the shop owners happy.

The two policies also differ in complexity of planning on the part of the decision makers.

Leaving the technical issues for both cases aside here, the problems related to the introduction of

rationing schemes are principally in the choice of the authorized user groups and in the

identification of “complementary measures” (e.g. improvement of the transit lines accessing the

area) to avoid undesirable effects (e.g. increase of the traffic in the surrounding area). Such issues

are relatively simpler than those that accompany the introduction of pricing policies, which has

been the subject of a technical and political debate for years over how to put them into practice.

For some decades, the cardinal point of the debate on taxation of road use has been the theory of

MCP, starting from the welfare theories of Marshall and Pigou in the beginning of the twentieth

century. However it has gone in and out of favor (Rothengatter, 2003). According to this theory, the
5
social costs (impact on environment and on all users) generated by the single user are to be

“internalized”, levying on the use of each infrastructure a tax equal to the marginal social cost, that

is, the incremental social cost induced by one additional user. As a consequence the tariff imposed

on the user would have to vary in function of the place, traffic, moment, vehicle type, etc.

This idea was first applied to the traffic assignment problem by Beckmann, McGuire and Winstein

(1959). Dafermos and Sparrow (1971) proved the optimality of MCP in a mono-user monomodal

deterministic context with rigid demand and separable arc cost functions. Dafermos (1973) extended

this result to a multiclass context, while Smith (1979) generalized it to the case of elastic demand and

asymmetric arc cost function Jacobian. In Bellei, Gentile and Papola (2002) the validity of MCP is

extended to any stochastic traffic assignment on multimodal networks based on random utility theory.

Though recommended by the European Commission as late as 1998 (in the White Paper “Fair

and efficient pricing of the transport infrastructure”) as a rule of thumb for setting tariffs for the

transport infrastructures, the application of the theory of MCP shows gross limitations for different

reasons (Verhoef, 2002a): the tariff structure would be excessively complicated and extremely

difficult for users to understand; the current technology, while greatly evolved in recent years, is not

yet able to support so complex a tariff structure; it is almost never possible to impose road pricing

on the entire network, and, at least in an initial phase, the pricing scheme should be introduced on a

suitable part of the network; the tariff level may not be politically acceptable.

Such obstacles have given rise to the identification of different second-best solutions which take

account of the limits imposed by the practical necessities of implementation (Verhoef, 2002b; May

and Milne, 2000), even if this means the process will become less efficient (Liu and McDonald,

1999). One of the most common second-best solutions is cordon pricing, where access to a given

portion of the road network is charged a toll. This coincides with a particular configuration of road

pricing where all arcs whose tail and head are respectively out and in the Limited Traffic Zone

(LTZ) are charged a same toll. Despite its theoretical significance, from the simulation point of

6
view rationing can be regarded as a particular instance of cordon pricing where the toll equals the

fine for entering the LTZ multiplied by the probability of getting it.

In the following we present some examples of MCP, aimed at showing that referring to a

multimodal context with elastic demand is necessary in order to grasp the actual potentiality of road

pricing on real transport networks. Figure 2a shows the relative increment of social welfare in

respect to the case of no policy obtainable by MCP with reference to the classical test network of

Sioux Falls for different levels of demand (that is scaling the travel demand uniformly through a

demand multiplier) in three different contexts: rigid demand (only private car is available), elastic

monomodal demand (only private car is available, but users can stay at home), multimodal demand

(users must travel, but can choose between private car and public transit). Figure 2b depicts the ratio

between the toll revenues and the social welfare, giving an idea of the economical impact of MCP.

Figure 2c reports the percentage of users traveling by car. Other numerical examples relative to

different networks, not reported here for brevity, show similar model behavior.

a) welfare increment b) revenue / welfare c) road users

70% 350% 100%
rigid
rigid x 10 90%
60% 300% elastic monomodal
elastic monomodal 80%
multimodal
50% multimodal 250% 70%

60%
40% 200%
50%
30% 150%
40%
rigid
20% 100% 30%
elastic monomodal
20%
10% 50% multimodal
10%

0% 0% 0%
0 0.25 0.5 0.75 1 1.25 1.5 1.75 2 0 0.25 0.5 0.75 1 1.25 1.5 1.75 2 0 0.25 0.5 0.75 1 1.25 1.5 1.75 2
demand multiplier demand multiplier demand multiplier

Figure 2. The effects of MCP in three different demand contexts: rigid, elastic monomodal, multimodal.

We can see that in the rigid demand context MCP is poorly effective (note that its depicted

welfare increment is multiplied by 10) and leads to very high tolls. On the contrary, it is very

effective in the elastic monomodal demand context, although tolls are still high. In the multimodal

demand context, which is the more realistic for cities during peak hour, MPC is still very effective

and the tolls are less high.

7
 As expected the real target of road pricing is mode choice and not only route choice. In this

respect, the question of which arcs of the network should be charged, which is fundamental for

directing route choice, becomes less crucial, so that simplified schemes, such as cordon pricing can

be investigated, and the classical road traffic assignment models with rigid demand become an

improper tool for this investigation.

3 MODELING METHODOLOGY

3.1 The demand model

In modeling travel demand we follow the behavioral approach based on random utility theory

(see, for instance, Ben Akiva and Lerman, 1985), where it is assumed that the user is a rational

decision-maker who, when making his travel choice: considers a finite number of mutually

exclusive travel alternatives constituting his choice set; associates with each travel alternative of his

choice set a utility, not known with certainty and thus regarded as a random variable; selects a

maximum utility travel alternative. On this basis, user’s behavior is described in terms of the

probability that each of his travel alternatives has maximum utility.

The users are grouped into demand components. The Ni > 0 users constituting the demand

component i∈D, where D is the set of demand components, are assumed to be identical to each

other in respect to any characteristic influencing travel behavior and externalities, and share the

same choice set J(i) of travel alternatives. Specifically, these users travel between the origin o(i)∈C

and the destination d(i)∈C, where the set of the centroids C is a subset of the node set.

The demand components are grouped into user classes. All the demand components belonging to

a same user class u∈U, where U is the set of the user classes, share the same set of individual

attributes characterizing the user as a trip consumer (e.g. age, value of time, and purpose of trip) and

the same set of attributes specifying the production of trip externalities such as congestion and

pollution (e.g. vehicle type and occupancy rate). Moreover, they have available the same subset
8
M(u) ⊆ M of modes, where M is the set of the transport modes.

In the following we refer to the generic demand component i∈D, whose user class is denoted

u(i). The utility Uji of the generic travel alternative j∈J(i) is given by the sum of a finite systematic

utility term Vji, and a zero mean random residual εji. The choice model is assumed to satisfy the

following properties: the random residuals have finite variance and their joint probability density

function is independent of the systematic utilities, continuous, and strictly positive; the systematic

utility of the generic travel alternative is given by subtracting the generalized cost Cji of the

associated route on the transport network to a constant utility term Xji independent of congestion:

Vji = Xji-Cji. (1)

The choice probability Pji of the generic travel alternative is by definition the probability of j

being a maximum utility travel alternative among the choice set J(i):

Pji = Pr[Vji+εji ≥ Vki+εki: k∈J(i)].

The above equation expresses the choice probability of the generic travel alternative as a function of

the systematic utilities of all the travel alternatives of the choice set:

Pji = Pji(Vki: k∈J(i)). (2)

The flow Fji of the travel alternative is given by multiplying its choice probability by the number of

users of the demand component:

Fji = Pji ⋅ Ni ; (3)

then, by definition, we have:

∑ j∈J(i) Fji = Ni .

On the basis of (2) and (1), equation (3) defines the demand function, expressing the travel

alternative flows in terms of the travel alternative generalized costs:

Fji = Ni⋅Pji(Xji-Cji: k∈J(i)), in compact form: F = F(C), (4)

where F and C are the (n × 1) vectors of the travel alternative flows and costs, with n = ∑i∈D |J(i)|.

The satisfaction Wi is by definition the mean value of the maximum utility:

9
Wi = E[max{Vji + εji: j∈J(i)}].

On the basis of (1), the above equation defines the satisfaction function, expressing the satisfaction

of the demand component in terms of the travel alternative generalized costs:

Wi = Wi(Cji: j∈J(i)), in compact form: W = W(C).

where W is the (|D| × 1) vector of satisfactions.

As usual, we assume that the travel choice process can be decomposed into a sequence of

mobility choices (in our case: to travel or not to travel, by which mode and following which route)

represented by a choice tree (Oppenheim, 1995). A travel alternative is then a path on the user

choice tree, specifying the choice made at each level (in our case: trip generation, modal split and

route assignment), and its utility is the sum of the utilities associated with each arc of this path. In

particular, each travel alternative (except for the “not-to-travel” alternative, when present) is

associated with a single route connecting on a specific mode the origin of the trip to its destination.

The hierarchic structure of the choice model (i.e. the correlations between the utilities of the

travel alternatives) is implicit in the form of the joint probability density function of the travel

alternative random residuals. However, in practice the choice made at each level is described by

means of a sub-model where the utility due to the choices made at lower levels is synthetically

taken into account through a satisfaction variable. The utility associated with each alternative of a

sub-model is given by summing up a specific systematic utility term, a specific random error term,

and the maximum utility obtainable from lower levels, whose mean value is the satisfaction.

Each travel alternative j∈J(i) is identified by a mode m(j)∈M(u(i)) and a route r(j)∈R(o(i), d(i), m(j)),

that is j = (m(j), r(j)), where R(o, d, m) is the set of routes available on mode m∈M for traveling

between the centroids o∈C and d∈C. In order to model trip generation, the choice set J(i) is given

as: J(i) = ∪g∈G(u(i)) J(i, g), where G(u) is {0,1} if the user class u∈U can choose whether to make the

trip or not, and {1}, otherwise. The set J(i, 1) is the set of the actual travel alternatives

{(m, r): m∈M(u(i)) , r∈R(o(i), d(i), m)}, while the set J(i, 0) is the “not-to-travel” alternative (0, 0).

10
 choice levels

trip generation modal split route assignment

g∈G(u(i)) m∈M(u(i)) r∈R(o(i), d(i), m)

0 -Cro(i)d(i)m
Wi Wo(i)d(i)m
Xm/i1
W1/i
X1/i

V1/i = X1/i + W1/i Vm/i1 = Xm/i1 + Wo(i)d(i)m

Figure 3. The choice tree.

The choice model then has the following structure:

Pji = P1/i ⋅ Pm(j)/i1 ⋅ Pr(j)o(i)d(i)m(j) j∈J(i, 1) , Pji = (1 - P1/i) j∈J(i, 0),

P1/i = P1/i(V1/i), Wi = Wi(V1/i), V1/i = X1/i + W1/i ,

Pm/i1 = Pm/i1(Vm/i1: m∈M(u(i))), W1/i = W1/i(Vm/i1: m∈M(u(i))), Vm/i1 = Xm/i1 + Wo(i)d(i)m,

Prodm = Prodm(Crodm: r∈R(o, d, m)), Wodm = Wodm(Crodm: r∈R(o, d, m)),

where the satisfactions are denoted W, the specific systematic utilities X, the systematic utilities V,

and the probabilities P. For j∈J(i, 1), it is: Xji = X1/i + Xm(j)/i1 and Cji = Cr(j)o(i)d(i)m(j) .

In the above elastic demand model, the traditional partial share approach for representing travel

choices is integrated into a factorized random utility model (Cascetta, 2001, pag. 181), ensuring

that, not only route assignment, but the entire sequence of sub-models, including generation and

modal split, reflect any change in the level of service pattern through the satisfaction variables.

3.2 The supply model

The multimodal network of infrastructures and services which constitutes the transport supply is

modeled here through an oriented hyper-graph G = (N, A), where N is the set of the nodes, each

11
node representing a spatial location and possibly a trip state (e.g. arriving at a road intersection, start

or end waiting at a transit stop), and A is the set of the hyper-arcs, each hyper-arc representing a

specific trip phase (e.g. driving throughout a road link, waiting at a transit stop). The generic hyper-

arc a∈A is identified by one initial node TL(a)∈N, referred to as tail, and by a set of ending nodes

HD(a)⊆N, referred to as head: a = (TL(a), HD(a)). An arc is a hyper-arc whose head is a singleton.

In this framework, the route, if any, associated with a given travel alternative (no route is

associated with the “not-to-travel” alternative) is represented through a hyper-path connecting on G

the origin of the trip to its destination. An hyper-path on G with initial node o∈N and final node

d∈N, is a hyper-graph r = (Nr , Ar) having the following properties: 1) r is an acyclic sub hyper-

graph of G; 2) o has no entering hyper-arcs in Ar and d has no exiting hyper-arcs in Ar ; 3) each

node i∈Nr\{d} has exactly one exiting hyper-arc in Ar ; 4) for every node i∈Nr\{o, d} there is at

least one path on r from o to d traversing it.

In the following we assume that G is constituted by a road graph GR = (NR , AR) and by a transit

hyper-graph GT = (NT , AT), with NR ∩ NT = C. The road graph represents the base graph. The transit

hyper-graph includes a copy of the road graph which acts as a pedestrian network. The generic line

l∈L, where L is the set of the transit lines, is constructed starting from its line routes LR(l)∈AR,

expressed in terms of a connected sequence of support arcs of the base graph, consistently with the

scheme depicted in Figure 4. An access arc representing a transit stop is introduced for each arc of

the base graph which acts as support arc for at least one line.

Note that the generic hyper-arc is defined as a one-to-many relationship between nodes.

However, all the hyper-arcs introduced here are ordinary one-to-one relationships, that is simple

arcs, except for the transit waiting hyper-arcs, where the tail is a waiting node and the head is a set

of line nodes (Nguyen, Pallottino and Gendreau, 1998).

12
 line 2 line arc

boarding arc
boarding node

line 1

waiting hyper-arc
line node
alighting arc

waiting node
access arc

pedestrian arc

pedestrian node

Figure 4. The transit hyper-graph.

The hyper-arcs are a useful tool to represent the adaptive choice occurring during a transit trip

when the user waiting at a stop boards each line with a certain probability. Specifically, each

waiting hyper-arc identifies a set of lines on which the passenger is willing to board and a

probability is associated with each boarding node belonging to the head of the hyper-arc. For

example, if the passengers’ behavior is to board the first arriving carrier of a given set of lines, the

probability associated with each line node is the probability that the line is served as first among the

set of lines identified by the hyper-arc at hand. The presence of many hyper-arcs at the same stop

(as in Figure 4) represents the possibility of choosing the most convenient set of lines on which the

passenger is disposable to board, which is usually referred to as the attractive line set. In theory, a

waiting hyper-arc should be introduced for any disposition of the boarding arcs at a given stop.

Fortunately, there is no need to enumerate the hyper-arcs, as they are regarded implicitly at the

algorithm level, where the network is implemented through a standard graph by eliminating from

the hyper-graph described in Figure 4 all waiting hyper-arcs and boarding nodes and then
13
connecting the boarding arcs directly to the corresponding waiting node.

Starting from the road graph and the transit hyper-graph, several mode-specific networks can be

constructed in order to simulate: different pricing and rationing policies for specified categories of

users; different performances characterizing the various transport modes and vehicle classes;

different behaviors in terms of cost perception. In general, the network associated to the generic

mode m∈M represents trips of users who perceive and produce costs and congestion in the same

way and thus share the same route choice conditional probabilities.

This approach implies an extension of the concept of mode, which embodies elements of both

the supply and the demand models and consists of an aggregation process aimed at optimizing the

algorithm. Indeed, this permits, in the context of a multiclass equilibrium, to accomplish in practice

a thick segmentation of users on the demand side, as this is not expensive in terms of computational

effort, while reducing to a minimum the segmentation on the supply side, which requires specific

shortest paths and network loading procedures.

The generic mode m∈M is referred to a specific hyper-graph G(m) = (N(m), A(m)) which

coincides with GR or GT . Let cam and fam denote, respectively, the arc cost and the arc flow on the

generic arc a∈A(m). The following relations hold:

Cji = ∑ a∈A(m(j)) cam(j)⋅πam(j)ij , in compact form: C = Π T⋅c , (5)

fam = ∑ i∈D, j∈J(i) Fji⋅πamij , in compact form: f = Π ⋅F , (6)

where: πamij is the arc-alternative probability that the arc a∈A(m), with m = m(j), is utilized by the users

belonging to the demand component i∈D as an element of the hyper-path associated with the travel

alternative j∈J(i, 1), otherwise πamij is equal to zero; Π is the (ν × n) matrix, with ν = ∑ m∈M |A(m)|,

whose elements are πamij; c and f are the (ν × 1) vectors whose generic component is respectively cam

and fam.

The main advantage of using this hyper-graph based network representation is expressing in a

purely additive form the travel alternative generalized costs in terms of the arc costs through (5) and
14
the arc flows in terms of the travel alternative flows through (6), also when taking into

consideration the adaptive choices taking place at the stops of the transit network, thus avoiding the

introduction of non-additive alternative costs as in Cascetta (2001).

It is to be noted that, where no on-line information is available at the stops, Π is fixed as it

depends exclusively on the line headway distributions (Gentile, Nguyen and Pallottino, 2003),

while if there is no need to represent adaptive choices, Π becomes a classical arc-path incidence

matrix.

A desirable feature of the equilibrium model is the possibility of simulating the simultaneous

effect on traffic flows of applying different toll patterns on different modes. To this end it is

convenient to separate the tolls ρ from the performances ĉ, so that the arc costs are defined as

follows:

c=ĉ+ρ, (7)

where the vectors ρ and ĉ have both the same structure as c. The congestion phenomenon is then

represented formally through the arc performance function:

ĉ = ĉ(f) , (8)

which is in general non-separable, and thus permits to take into account the interaction among

different transport modes sharing the same road links and the transit congestion.

3.3 User Equilibrium, System Equilibrium and toll optimization problem

The UE and the SE are formalized here as fixed point problems. On this basis, we address the

problem of determining an optimal arc toll vector, assuming that it is possible to charge each arc of

the network any real valued toll.

By combining (5), (4) and (6), we obtain the network loading map (Cantarella, 1997), which

yields the arc flows as a function of the arc costs:

f(c) = Π ⋅F(Π T⋅c) . (9)

15
 Then, for a given value of the arc tolls ρ, a UE flow and cost pattern is determined by solving the

fixed point problem obtained by combining (9), (7) and (8):

f = f[ĉ(f) + ρ] . (10)

The social cost and its gradient, referred to as the marginal social cost, can be expressed in terms

of the mode-specific arc flows respectively as follows:

sc(f) = ĉ(f)T⋅f + E(f) ,

msc(f) = ∇f sc(f) = ĉ(f) + ∇f ĉ(f)⋅f + ∇f E(f) ,

where the function E(f) expresses the monetization of the externalities.

By definition (see Bellei, Gentile and Papola, 2002), a SE flow pattern is determined by solving

the following fixed point problem:

f = f[msc(f)] , (11)

obtained by replacing in the UE problem (10) the cost function ĉ(f) + ρ with the marginal social

cost function msc(f).

A discussion about the existence and uniqueness of the solution to the UE and SE problems is

beyond the scope of this paper.

The analysis of road pricing from the point of view of economic efficiency leads to the well-

known result that the optimal flow pattern is an SE, which enables us to generalize the notion of SO

current in the literature to the case of stochastic equilibrium. Consider the following toll

optimization problem:

max f, ρ S(f, ρ) = ∑ i∈D di ⋅ Wi(Π T⋅[ĉ(f) + ρ]) - E(f) + ρ T⋅f

s. to: f = f[ĉ(f) + ρ] , (12)

where the objective function is social welfare, the UE constraint is expressed by means of the fixed

point problem (10), and we are assuming that the transfer of money from private to public takes

place with a unitary weight. With reference to the above bilevel formulation, in Bellei, Gentile and

Papola (2002) it is proved that “for each arc toll vector that solves the problem there is another

16
solution vector mcp(f) = ∇f ĉ(f)⋅f + ∇f E(f) which leads to the same SE flow pattern f ” (proposition

5), and that “for an arc toll vector to solve the problem it must yield an SE flow pattern; if the SE is

unique, then the condition is also sufficient and the solution is unique in terms of flow pattern”

(proposition 6).

If the SE is unique, then the solution is unique in terms of flow pattern, but not in terms of tolls.

However, in practical terms, the non-uniqueness of the pricing solution can occur only in particular

situations having no relevance from an operational point of view. In fact, the bicriteria approach

(which consists in exploiting the multiplicity of the solution in terms of arc tolls in order to meet a

second criterion, such as minimizing the toll revenue or the number of charged arcs) adopted by

Hearn and Ramana (1998) and Dial (1999) in a deterministic route choice framework, while helpful

for defining the structure of a second-best pricing policy, is not essential when solving the problem

in the stochastic case (Bellei, Gentile and Papola, 2002).

A solution satisfying the necessary conditions of the toll optimization problem (12) can then be

determined by solving a UE problem, where the tolls ρ are replaced with the MCP toll function

mcp(f). Clearly, if the SE is unique, then the solution obtained this way is a global solution.

Note that MCP is a kind of discriminatory toll scheme, that charges on each arc a distinct toll for

every mode, which aggregates here trips sharing the same cost production and perception. However,

in practice, it is very difficult to isolate the different groups of users in the system.

4 NUMERICAL APPLICATION TO A LARGE SCALE NETWORK

Access control in Rome was first applied in 1989, when restrictions were placed on vehicle

entrances to the old city center (the LTZ), an area of about 6 km2 containing an enormous historical

and artistic heritage. Permission to enter was given free of charge to residents of the area (including

the disabled and public service vehicles) and to certain categories of users (workers retaining a

private parking slot in the LTZ, some categories of craftsmen and technicians, and freight vehicles)

in the following referred to as the authorized. In 1998 a pricing policy for the authorized was
17
introduced. They are now required to pay for their permits an annual fare of about 300 euros (the

equivalent of 12 monthly public transport passes). Both measures are now enforced through an

automated system based on 23 electronic gates able to identify the vehicle and to calculate the

applicable fare or fine for vehicle entrance into the restricted area.

Due to several problems occurring with the current regime (high level of congestion in the

surrounding area and very high number of mopeds), the introduction of alternative cordon-based

policies has been envisaged by the local authorities including road pricing. Referring to the groups

of users identified by the Public Administration, we are interested in comparing cordon pricing and

rationing with MCP, in order to establish how far these simplified schemes are from the SO upper

bound in terms of social welfare and to which extent may thus represent a second-best solution of

the toll NDP. With the aim of avoiding arbitrary assessments, the external costs have not been taken

into account in the social welfare; we limited ourselves to evaluate each factor, leaving it up to a

politic decision the task of choosing their weight, as a multi-objective analysis is beyond the scope

of this work. Moreover, in evaluating the impact of the cordon-based policies we ignored the

negative effects on the activity system related to the reduction of accessibility in the charged area.

Despite the relatively limited area involved in the pricing intervention, the strong appeal of the

LTZ has induced us to consider in the traffic simulations a much larger area of study (the whole

Lazio Region). A graph with 16,878 directed arcs and 5,902 nodes, 854 of which are centroids, was

used as the base network on which 680 transit lines with 19,626 line arcs (including the regional

railway and bus systems) have been defined. The demand has been segmented in terms of trip

purpose (systematic and non-systematic), personal attributes (5 combinations of age, sex,

profession, car and motorcycle availability) and destination (LTZ and non-LTZ) giving rise to

1,534,788 demand components grouped in 20 user classes. Exploiting the concept of mode

introduced here, this complex demand structure has been simplified at supply level introducing only

3 modes (resident cars, authorized cars, non-authorized cars) to simulate the different policies, and

18
2 additional modes to simulate multimodality (mopeds, transit).

Consistently with the aim of the simulations, the choice of the destination is assumed to be

relevant only when it involves the access to the LTZ. In fact, should road pricing be implemented,

we expect that very few drivers not heading towards the LTZ will cross it; on the other hand, the

availability of the car mode for traveling towards the LTZ, is likely to attract trips previously

directed to other destinations. Then, one additional user class has been introduced representing non-

systematic trips with destination in the LTZ, for which the demand is considered elastic up to the

generation, while for all other user classes the generation is assumed fixed. On this basis, the

distribution model is implemented as a generation model toward the LTZ. Then, three choice levels

have been actually considered: generation, modal split and route assignment.

The demand model has been specified as a Nested Logit with reference to generation and modal

split, while as a Probit with reference to route choice, and has been calibrated using RP-SP data, as

the introduction of a road pricing scheme involves changing the choice set of road users. In the

supply model, besides road congestion, we have taken into account that transit vehicles are slowed

down by private traffic when they use the same lanes, the discomfort of overcrowding, the times

needed for boarding and alighting at the stops as a function of the transfer flows, and the effect of

line capacities. Exploiting implicit path enumeration in the assignment algorithm, the computation

of the above instance of the equilibrium problem takes about 10 minutes on a 1500 Mhz PC. The

interested reader may find all the model specifications and details on the solution algorithm in

Gentile. Papola and Persia (2004).

Some preliminary results showed that, even with a high level of fares, no substantial changes can

be expected, in terms of social welfare and modal split, should the access limitation remain

operational and only authorized cars be charged an increasing toll. This is due to the limited number

of vehicles potentially subject to charges, compared to the whole number of users accessing the

area, and to the usually high surplus level of the user categories charged. More relevant

19
modifications occur when the access restriction is extended to mopeds, that represent a relevant

share (19%) of private traffic. This policy also reduces the through traffic generated by such a

category of private vehicles (some 40% of the motorbike users currently accessing the LTZ have

their destination outside this area).

In the following we present and discuss some results relative to the introduction of pricing

policies where the access restriction for non-authorized car users is removed: residents may travel

free of charge through the LTZ, authorized cars pay a fixed toll of 1 €, non-authorized cars pay a

variable toll up to the amount of the fine (here we assume that, thanks to the automated control

system, the probability of getting the fine for the non-authorized vehicles violating the LTZ is 1).

All the relative evaluations will refer to the no-policy scenario (LTZ toll = 0 €).

Figure 5 depicts the percentage of the social welfare increment achievable through MCP on the

whole network (which amounts to 52% of overall social welfare) captured by means of different

cordon pricing policies with an increasing toll to enter the LTZ. Although the LTZ covers only a very

small part of the whole area of analysis and includes the destination for only 14% of the travel demand,

up to 18% of the theoretical welfare increment due to MCP can be achieved through a cordon pricing

policy. In this regard we must recall that implementing MCP would involve very high construction and

maintenance costs, that are much higher than those required to implement a cordon policy. Therefore,

the benefits obtained from the cordon-based schemes when compared with MCP are on the low side

and a bit conservative. In light of this consideration 18% becomes quite a good result.

The optimal toll appears to be 12 €, which is consistent with the toll actually charged in other

European cities (in London about 8 € is charged). The rationing policy (LTZ toll = 100 €) is also

effective, capable of capturing one third of the welfare increment due to the best cordon pricing

policy, while it may require much smaller implementation costs depending on the control system

adopted. The toll yielding the break-even in terms of social welfare between pricing and rationing

appears to be around 2 €.

20
 20%
18%
16%
14%
12%
10%
8%
6%
4%
2%
0%
1.00 2.00 3.00 5.00 8.00 12.00 18.00 100.00
LTZ toll (€)

Figure 5. Portion of the MCP social welfare increment achieved by the policy, versus cordon toll.

Figure 6 depicts the ratio between the toll revenue and the social welfare for a number of

different cordon pricing tolls. Interestingly, the policy achieving the maximization of the toll

revenue (around 500.000 € per day) coincides with the optimal one in terms of social welfare;

although, clearly, this is not a general result, and by increasing the number of the simulations this

coincidence would vanish. Recall that MCP implies a very high revenue-welfare ratio (295% in our

case). For these reasons, MCP is probably not acceptable from a social point of view, while the

cordon pricing policies, although concentrating their impact on a restricted group of users, have a

much softer impact on the network travel cost pattern (the revenue-welfare ratio is at the most 12%).

12%

10%

8%

6%

4%

2%

0%
1.00 2.00 3.00 5.00 8.00 12.00 18.00 100.00
LTZ toll (€)

Figure 6. Ratio between toll revenue and social welfare, versus cordon toll.

Figure 7 depicts the increment of user surplus. Note that not all the cordon policies considered

are convenient both in terms of social welfare and in terms of user surplus. Regarding the latter, in

21
reference to the specific case study at hand, rationing turns out to be the best policy. On the other

hand, referring to the systematic mobility with destination in the LTZ, the monetary cost increase

due to the tolls overcomes the time saving. Indeed, the advantages of the pricing policies are to be

found in the reduction of congestion and externalities for all users.

3%

2%

1%

0%
1.00 2.00 3.00 5.00 8.00 12.00 18.00 100.00

-1%

-2%
LTZ toll (€)

Figure 7. Increment of user surplus, versus cordon toll.

Figure 8 depicts the percentage of transit share increment achieved through MCP (equal to 4%)

captured by means of increasing cordon pricing tolls. Clearly, the policy with the best modal shift is

pure rationing, yielding here a 59% relative improvement.

70%

60%

50%

40%

30%

20%

10%

0%
1.00 2.00 3.00 5.00 8.00 12.00 18.00 100.00
LTZ toll (€)

Figure 8. Portion of the MCP modal shift increment achieved by the policy, versus cordon toll.

Figure 9 depicts the percentages of speed increments achievable through MCP (equal to 66% for

cars, 10% for mopeds, and -6% for transit) captured by means of increasing cordon pricing tolls.

The high speed increase achieved through MCP enjoyed by cars, which corresponds to a travel time

22
decrease of 40%, is due to a decrease of traffic flows of almost 20% (the car modal share

diminishes from 48% to 40%; 4% switches to transit and 4% to the moped mode, whose relative

weight on the main traffic stream is low) on a highly congested road network, and reflects the non-

linearity of the supply model. As expected, the policy with the best private speeds is rationing,

yielding a 41% and a 23% relative increment for cars and mopeds, respectively. The speed

reduction characterizing the transit mode is due to the side effects of the cordon. In fact, in Rome’s

city centre the bus lines benefit from dedicated lanes, so that the LTZ does not induce on transit

vehicles any relevant speed improvement. At the same time, just outside the LTZ, where the

congestion increases due to the discontinuity in the network cost pattern produced by the toll

cordon, the line carriers share the road links with private vehicles. The effectiveness of road pricing

would be higher (in terms of social welfare and modal split) and the social impacts softer (in terms

toll values), should the policy be coupled with an improvement of the transit performances. Indeed,

in the case of Rome, this is a priority and in particular the protection of the bus lines from road

traffic outside the LTZ is a highly recommended intervention.

50%

40% car
moto
30% transit
20%

10%

0%
1.00 2.00 3.00 5.00 8.00 12.00 18.00 100.00
-10%

-20%

-30%

-40%

-50%
LTZ toll (€)

Figure 9. Portion of the MCP speed increments achieved by the policy, versus cordon toll.

Figure 10 depicts the increments of the externalities corresponding to the different cordon

pricing tolls. All polluting emissions decrease relevantly when the toll cordon increases. Instead,

when the toll is high enough to switch a considerable share of users to the moped mode, the number

of Killed and Seriously Injured (KSI) increases, because motorbikes are really unsafe.
23
 105%
FUEL CO VOC NOx

CO2 PM KSI

100%

95%

90%

85%
1.00 2.00 3.00 5.00 8.00 12.00 18.00 100.00

LTZ toll (€)

Figure 10. Increments of the externalities, versus cordon toll.

Figure 11 depicts the modal shares and the travel times for different cordon pricing tolls, with

reference to non-authorized users traveling towards the LTZ. This is the category of users at which

the pricing policies are mainly directed, and non surprisingly their effects are remarkable.

Specifically, the transit share, starting from 43% in the non-intervention option, goes up to 79% in

the case of rationing, while the car share decreases from 49% to zero.

100 90
90 transit 80 car
80
moto moto
car 70 transit
70
60
60
50
50
40
40
30
30
20 20

10 10

0 0
0.00 1.00 2.00 3.00 5.00 8.00 12.00 18.00 100.00 SO 0.00 1.00 2.00 3.00 5.00 8.00 12.00 18.00 100.00 SO
modal split (%) LTZ toll (€) travel time (min/us) LTZ toll (€)

Figure 11. Mode shares and the travel times for non-authorized users, versus cordon toll.

On the other hand, the motorcycle share increases from 7% to 21% with a relative increment of

200% and thus a high negative impact on safety. The main effect on travel times involves the car

mode, for which the commercial speed increases up to 50% in the 18 € solution (the value of car

24
travel time for the rationing policy is not significant, as the modal share is too small); the

improvement for mopeds, which are less affected by the congestion, is not as significant.

Figure 12 depicts the mode travel times for different cordon pricing tolls, with reference to non-

LTZ users. As the speed of private modes increases substantially with the toll, the initial conjecture

regarding the side effects of the rationing policy turns out to be partially incorrect in the sense that

the increment of congestion experienced right outside the LTZ is more than compensated by the

benefits on the rest of the network.

45
40
35
30
25
20
15
car
10 moto
5 transit
0
0.00 1.00 2.00 3.00 5.00 8.00 12.00 18.00 100.00 SO
travel time (min/us) LTZ toll (€)

Figure 12. Mode travel times for non-LTZ users, versus cordon toll.

Figure 13 depicts the share of non-systematic users potentially traveling to the LTZ who actually

make the trip. This user type is introduced to simulate implicitly the effects on the demand

distribution pattern as previously explained. Note that the rationing policy induces all users to

choose other destinations than the LTZ.

100
90
80
70
60
50
40
30
20
10
0
0.00 1.00 2.00 3.00 5.00 8.00 12.00 18.00 100.00 SO
LTZ toll (€)

Figure 13. Trip generation for non-systematic users directed to the LTZ, versus cordon toll.

25
 5 CONCLUSIONS

Starting from a theoretical framework proposed in a previous paper, in this work we present an

effective multiclass equilibrium model for multimodal networks. We have also devised an efficient

algorithm, omitted here for brevity but available on request, and applied it to the case study of

Rome.

The model is capable of simulating the simultaneous effect on traffic flows of applying different

fare structures to different categories of users in urban networks with a variety of transport modes

interacting among each other and, in particular, where transit line vehicles sharing road links with

private traffic are affected by the same overall congestion. In order to make a thick segmentation of

users practically possible, on the demand side, and the introduction of many transport modes, on the

supply side, while reducing to a minimum the calculation of specific shortest paths and network

loading procedures, we provide an extension of the concept of mode which embodies elements of

both the supply and the demand models and consists of an aggregation process aimed at optimizing

the algorithm.

The proposed model has then been utilized for analyzing cordon pricing and rationing policies

simultaneously on different modes as a second-best solution of the toll network design problem.

Specifically, in Rome there is an automated control system based on 23 electronic gates that

constitute the cordon of a Limited Traffic Zone.

In cases like these where the aim of the policy is to improve the modal split, it is not possible to

evaluate any intervention by means of a rigid demand traffic assignment model, while a classical

elastic monomodal demand model does not reproduce correctly the context at hand. Indeed, a high-

performance equilibrium model is required, which allows simulating in practical terms the

variations in generation and modal split caused by broad-ranging measures, such as those

considered here, and is sensitive to the different reactions of the different categories of users.

In evaluating the different operative policies, Marginal Cost Pricing has been taken into account

26
as an upper bound, constituting a helpful point of reference, despite its being purely theoretical and

not achievable in practice. When examining the effectiveness of a specific policy by comparison

with the upper bound, two main aspects are to be considered: the implementation costs of MCP are

prohibitive with the available technologies; the extent of the LTZ is very small compared with the

whole network. In light of the above, the results obtained can be considered definitely satisfactory:

although the LTZ includes the destination for only 14% of the travel demand, up to 18% of the

theoretical social welfare increment due to MCP can be achieved through a 12 € toll generating a

revenue of about 500.000 € per day and 1/3 of the modal shift produced by MCP; rationing is

capable of yielding 1/3 of the welfare increment due to the above cordon pricing policy and 60% of

the modal shift produced by MCP, while it may require much smaller implementation costs

depending on the control system adopted, so being a valid alternative to road pricing, at least in the

specific case of Rome.

The results are also promising for future investigation, where, removing the limits of the case

study at hand (position, size and shape of the toll cordon, user categories identified), more general

policies and schemes can be considered, and so making optimal use of the power and versatility of

the model.

REFERENCES

Beckmann M., McGuire C.B., Winstein C.B. (1956) Studies in the economics of transportation,

Yale University Press, New Haven, CT.

Bellei G., Gentile G., Papola N. (2002) Network pricing optimization in multi-user and

multimodal context with elastic demand, Transportation Research B 36, 779-798.

Ben Akiva M., Lerman S. (1985) Discrete choice analysis: theory and application to travel

demand, MIT Press, Cambridge, MA.

Cantarella G.E. (1997) A general fixed-point approach to multimode multi-user equilibrium

assignment with elastic demand, Transportation Science 31, 107-128.

27
 Cascetta E. (2001) Transportation systems engineering: theory and methods, Kluwer Academic

Publishers, Dordrecht, The Netherlands.

Dafermos S.C. (1973) Toll patterns for multiclass-user transportation networks, Transportation

Science 7, 211-223.

Dafermos S.C., Sparrow F.T. (1971) Optimal resource allocation and toll patterns in user-

optimized transport network, Journal of Transport Economics and Policy 5, 198-200.

Dial R.B. (1999) Minimal-revenue congestion pricing part I: a fast algorithm for the single-

origin case, Transportation Research B 33, 189-202.

Filippi F., Persia L. et al. (1996) Public transport prioritization. Transport Research APAS Urban

Transport Report, DG VII, Office for Official Publications of the European Commission,

Luxembourg.

Gentile G., Nguyen S., Pallottino S. (2003) Route choice on transit networks with on-line

information at stops, accepted for publication on Transportation Science, Technical Report TR-03-

14, Università di Pisa, Dipartimento di Informatica, Pisa, Italy.

Gentile G., Papola N., Persia L. (2004) Comparing marginal cost pricing with toll cordon

policies for large scale multimodal networks, in Proceedings of the 10th Word Conference on

Transport Research - WCTR'04, Istanbul, Turkey.

Hearn D.W., Ramana M.V. (1998) Solving congestion toll pricing models, in Equilibrium and

Advanced Transportation Modeling, ed.s P. Marcotte and S. Nguyen, Kluwer Academic Publishers,

Boston Hardbound, MA, 109-124.

Liu L.N., McDonald J.F. (1999) Economic efficiency of second-best congestion pricing schemes

in urban highway systems, Transportation Research B 33, 157-188.

May A.D., Liu R., Shepherd S.P., Sumalee A. (2002) The impact of cordon design on the

performance of road pricing scheme, Transport Policy 9, 209-220.

May A.D., Milne D.S. (2000) Effect of alternative road pricing systems on network performance,

28
Transportation Research A 34, 407-436.

Nguyen S., Pallottino S., Gendreau M. (1998) Implicit enumeration of hyperpaths in a logit

model for transit networks, Transportation Science 32, 54-64.

Oppenheim N. (1995) Urban Travel Demand Modeling, Wiley-Interscience, New York, NY.

Rothengatter W. (2003) How good is first-best? Marginal cost and other pricing principles for

user charging in transport, Transport Policy 10, 121-130.

Smith M.J. (1979) The marginal cost taxation of a transportation network, Transportation

Research B 13, 237-242.

Verhoef E.T. (2002 a) Marginal cost based pricing in transport: key implementation issues from

the economic perspective, paper prepared for the IMPRINT Project.

Verhoef E.T. (2002 b) Second-best congestion pricing in general networks. Heuristic algorithms

for finding second-best optimal toll levels and toll points, Transportation Research B 36, 707-729.

Zhang X., Yang H. (2004) The optimal cordon-based network congestion pricing problem,

Transportation Research B, forthcoming.

29