Tolls and Transit Frequencies Optimization

G. Bellei, G. Gentile and N. Papola 1

Abstract

In this paper a tolls and transit frequencies optimization problem has been
reduced to a specific equivalent equilibrium assignment problem and its local
solution conditions have been stated. A specific fixed-point approach algorithm has
also been devised, introducing a hyper-network consisting of a multi-modal
transportation network combined with a specific investment network. Based on a test
network, a numerical example concerning different scenarios is presented, exhibiting
encouraging results.

Introduction

The Network Design Problem (NDP) consists, as is known, in the seeking of a

transportation network configuration that maximizes some objective function, subject
to the equilibrium constraint. In this paper we address the optimization of an existing
multi-modal network, considering tolls and transit line frequencies as design
variables.
On the supply side, congested networks are considered, utilizing non-separable
link-cost functions in a multi-user and multi-modal context, where, with specific
reference to the transit network modeling, a directed hypergraph (Nguyen, Pallottino
and Gendreau, 1998) will be utilized, thereby enabling to express the relation
between arc and path costs in an additive form. On the demand side, probabilistic
additive strictly positive differentiable choice models, based on the random utility
theory, have been employed, with elastic demand up to the trip generation level. In
seeking the equilibrium, a fixed-point approach will be used as in Cantarella (1997).
Despite the fact that frequencies optimization is a classic transportation design
issue, most contributions found in the literature face the problem by adopting an
heuristic approach, often combined with route generation procedures. A bilevel
optimization approach, consistent with the nature of the problem, has been utilized

1
Università degli Studi di Roma “La Sapienza” , via Eudossiana, 18 00184 Rome, Italy

1 Bellei, Gentile, Papola

instead by LeBlanc (1988) and Yang (1997), employing the Hooke and Jeeves
algorithm and sensitivity analysis, respectively.
In Bellei, Gentile, and Papola (1998), the NDP is formalized within the well
established framework of the neoclassic microeconomic theory, considering network
users as consumers of urban trips (trip consumer approach) and assuming an
objective function of the so-called Utilitarian type: the Social Surplus (S) , which
measures the social welfare summing up the supply modification monetary effects on
all individual. On this basis, the problem of optimizing the tolls on an existing
network, assumed that it is possible to apply any toll to each arc of the network, is
addressed and the conditions under which the solution exists and is unique are
derived. In this work we extend the problem to the case where transit line frequencies
are optimized together with tolls.

Formalization of the problem

Let t be the vector of arc tolls, φ the vector of line frequencies, f the vector of
arc flows, S( f, t, φ) the social surplus function and f UE the User Equilibrium (UE)
arc flow pattern. The NDP can be formally expressed as:
max f, t, φ S( f, t, φ) = [N⋅Γ -1⋅W(C( f , t, φ))] T⋅1 -E( f , φ) -G(φ) +t T⋅f (1)
s.t.
φ min ≤ φ ≤ φ max
G(φ) ≤ B
f = f UE
f UE = f[c( f UE, φ) +t, φ]
where: N is the diagonal matrix whose generic entry is the number of users
belonging to class u∈U, having specific choice set J u; Γ -1 is the inverse of the
corresponding income marginal utility; C the vector of generalized costs; W(C) the
satisfaction function; E( f , φ) the transport externality function; G(φ) the transit
operating costs function.
As a cost cau and a flow fau , specific to class u, are assigned to each arc a∈A, we
have:
Cju = Σa∈A cau⋅πaju j = 1, … , |J u| , u = 1, … , |U| , or, in compact form C = Π T⋅c, (2)
fau = Σj∈Ju Fju⋅πaju a = 1, … , |A| , u = 1, … , |U| , or, in compact form f =Π ⋅F, (3)
where Cju e Fju are, respectively, the generalized cost and the flow of the alternative
j∈J u and πaju is the probability that arc a is used within the hyperpath associated
with the alternative j. The design variables are then the arc tolls tau , a = 1, … , |A| ,
specific to each class u = 1, … , |U| , and the line frequencies φ l , l = 1, … , |L|. The
following relations hold:
c = c( f , t, φ) = c( f , φ) +t , (4)
Π = Π(φ) , (5)
where c( f , t, φ), c( f , φ), and Π(φ) denote the arc cost function, the arc performance
function, and the share function, respectively.
Let us define the Social Cost (SC) as:
SC = C +E +G -T = C +E +G , (6)

2 Bellei, Gentile, Papola

where: C = C T⋅F = c T⋅f = total costs; C = C T⋅F = c T⋅f = total performance costs;
T = T T⋅F = t T⋅f = pricing revenues.
As C and T are related to the corresponding arc variables by relations like the
(2), the SC can be expressed in terms of the arc flows f or the alternative flows F as:
SC = c( f , z)T⋅f +E( f , z) +λ⋅G( f , z) = sc( f , z) , (7)
SC = C(F , z) ⋅F +E(Π(z)⋅F, z) +λ⋅G(Π(z)⋅F, z) = SC(F, z) .
T
(8)
We assume the following hypotheses:
B > G(φ min) , (9)
∇G(φ) ≥ ε > 0 ∀φ : φ min
≤φ ≤φ . max
(10)
As a consequence, the feasible set of problem (1), with respect to φ , is compact
and non-empty, while the constraints are assumed regular. Then, for any feasible φ ,
the problem (1) is equivalent to the optimal tolls NDP in Bellei, Gentile e Papola
(1998), whose solution has been proved to determine a System Equilibrium (SE) flow
pattern. We recall that SE is a UE where the individual average costs c are replaced
by the social marginal costs msc , defined as msc = ∇f sc( f , z). Assuming that the
functions appearing in the NDP (1) are all continuous, the value of the Social Surplus
in its solution is a continuous function of φ . Moreover, because the feasible set of φ
is compact, then, on the basis of the Weierstrass’ theorem, the solution of the NDP
(1) determines an SE flow pattern as well.

Solutions characterization

As it is C = C +T , problem (1) expressing S in term of C yields:

min C, φ -S(C, φ) = -[N⋅Γ -1⋅W(C)] T⋅1 -C T⋅F(C) +SC[F(C ), φ] (11)
s.t.
φ min -φ ≤ 0
φ -φ max ≤ 0
G(φ) -B ≤ 0
The Lagrangian function of problem (11) is:
L(C, φ, µ, η) = -S(C, φ) +∑l∈L µ minl⋅(φ minl -φ l) +∑l∈L µ maxl⋅(φ l -φ maxl) +η⋅[G(φ) -B]. (12)
The following Karush - Khun - Tucker first order conditions (KKT) :
η≥0, (13)
µ≥0, (14)
η⋅[G(φ) -B] = 0 (15)
µ minl ⋅(φ minl -φ l) = 0 , µ maxl ⋅(φ l -φ maxl) = 0 l = 1, ... , |L| , (16)
∂L(C, φ, µ, η)/∂φ l = -∂S(C, φ)/∂φ l -µ minl +µ maxl +η⋅∂G(φ)/∂φ l = 0 l = 1, ... , |L| , (17)
∇C L(C, φ, µ, η) = -∇C S(C, φ) = 0 , (18)
G(φ) ≤ B , (19)
φ min ≤ φ ≤ φ max , (20)
due to the constraint regularity, are necessary for a local minimum, and, so, for the
solution of the NDP.
The (18) coincides with the necessary condition of the cited optimal tolls NDP ,

3 Bellei, Gentile, Papola

thereby furnishing analytical evidence of the above statement that the solution of the
NDP (1) determines an SE flow pattern. Consequently, it is equivalent to the
following fixed-point problem:
f UE = f[msc( f UE , φ)] . (21)
Since we assume φ min < φ max, the product µ minl ⋅µ maxl is certainly equal to zero.
Then, differentiating (17) , based on (14), we obtain:
∂SC[F(C), φ]/∂φ l +η⋅∂G(φ)/∂φ l ≥ 0 , -∂SC[F(C), φ]/∂φ l -η⋅∂G(φ)/∂φ l ≥ 0 l = 1, ... , |L| . (22)
When both constraints relative to the line l are non-active, then (17) becomes:
∂SC[F.
(C), φ]/∂φ l +η⋅∂G(φ)/∂φ l = 0 . (23)
With reference to the generic line l , we define static efficiency, denoted ρ l, as:
ρl = -∂SC(F, φ)/∂φ l / ∂G(φ)/∂φ l , l = 1, ... , |L|. (24)
Dividing both sides of (22) and (23) by ∂G(φ)/∂φ l , which is positive by
hypothesis (10), and using (24) , (13)-(17) can be expressed as:
if φ minl <φ l < φ maxl, then ρl =η ; if φ l =φ minl, then ρl≤ η ; if φ l= φ maxl, then ρl ≥ η (25)
if G(φ) < B, then η = 0 ; if G(φ) = B, then η > 0 (26)
Differentiating (8) with respect to φ yields in compact form:
∇φ SC(F, φ) = ∑u∈U ∑j∈J u Fju⋅∇Πju(φ)⋅[∇f c(⋅f, φ) ⋅ f +c +∇f E(f , φ)] +
+∇φ c(f, φ)⋅f +∇φ E(f , φ)+∇G(φ) (27)
where Πj (φ) expresses a generic column of Π . Consequently, evaluating ρl needs
u

models working in terms of alternative variables, a difficulty we overcome using

∂sc( f, φ)/∂φ l as an approximation of ∂SC(F, φ)/∂φ l . In fact, differentiating (7) yields:
∇φ sc( f, φ) = ∇φ c( f, φ)⋅f +∇φ E( f , φ) +∇G(φ) , (28)
then, substituting (28) into (27) yields the simple function of f :
ρl =-[∑u∈U ∑a∈A fau⋅∂cau( f , φ)/∂φ l +∂E( f , φ)/∂φ l] / ∂G(φ)/∂φ l -λ, l = 1, ... , |L| . (29)
Finally, the KKT become:
f = f[cms( f , φ)] , (30)
if φ minl <φ l < φ maxl , then ρl( f , φ)= η ; if φ l =φ minl, then ρl(f , φ)≤ η;
if φ l =φ maxl , then ρl( f , φ) ≥ η (31)
if G(φ) < B, then η = 0 ; if G(φ) = B, then η > 0 , (32)
G(φ) ≤ B , (33)
φ min ≤ φ ≤ φ max . (34)

Assignment equivalent problem and solution algorithms

Conditions (31)-(34) coincides with the KKT of the problem:

min φ sc( f, φ) (35)
s.t.
G(φ) ≤ B
φ min ≤ φ ≤ φ max
The feasible set of problem (35) and NDP (1) are the same. Then, for any f∈Sf a
solution exists for which the KKT hold. Let φ( f ) be the map giving for any f∈Sf the

4 Bellei, Gentile, Papola

set of solutions for which the KKT hold, and let f SE(φ) be the map giving for any
feasible φ the set of f solutions of the SE. Then (30)-(34) can be reduced to:
f = f SE(φ) , (36)
φ = φ( f ) . (37)
To solve this system block-Jacobi or Gauss-Seidel algorithms (Bertsekas, 1995) are
usually used. More conveniently, a fixed-point approach is used in this paper to
devise two different algorithms, each having advantages and disadvantages. The first,
omitted here for lack of space, more robust but less efficient, furnishes directly the
line frequencies φ as an output; the second, described below, consisting of a hyper-
network with investment arcs is more efficient but requires the inversion of the
investment function.

Hyper-network with investment arcs algorithm

The hyper-network approach has been widely used when evaluating traffic flows
equilibrium with elastic demand (Sheffy, 1985). This approach, due to the wide
analogy between the Wardrop first principle and the KKT (31)-(34), can be extended
to our NDP, introducing an investment network with only one od pair, whose origin
and destination centroids represent, respectively, the budget to be allocated and the
budget already allocated. This pair is connected by as many investment arcs as the
cardinality of L plus one arc associated to the non invested budget. Actually, unless
one thinks of using an assignment algorithm with constraint capacity, thereby
introducing more theoretical and practical difficulties, to make workable such a
procedure it is necessary to reduce the constraint φ min ≤ φ ≤ φ max, to a non negativity
constraint. This can be done by, first, identifying φ min with the origin, after having
subtracted G(φ min) from the budget B, and, second, suppressing the constraint
φ ≤ φ max, simply replacing it with arc cost functions dramatically increasing in
proximity of the capacity point.
The approach we propose consists in jointly assigning the investment flows x
and the traffic flows f to a hyper-network obtained by combining the investment
network and the multi-modal multi-user transportation network. The procedure can
be formalized as a fixed-point problem in both variables f and x as:
( f T, x T) T = (f[msc( f , G -1(x))] T , x[ρ( f , G -1(x))] T)T . (38)
Because both the user costs on the transportation network and static efficiency
increase in f and decrease in x, while the costs of the investment network decrease in
ρ, the Jacobian of the arc cost functions is patently non symmetric and generally non-
positive semi-definite. Therefore uniqueness conditions for equilibrium do not hold
in general, which means that the problem is generally non convex.
The solving algorithm, detailed below, exhibits a number of steps connected with
the specific problem, and calculate, essentially, an equilibrium assignment of the
System Equilibrium type, through an MSA.

Algorithm:
1) initialization
2) k = k +1 update the counter

5 Bellei, Gentile, Papola

3) φ k = G -1(x k) calculate the frequencies
4) ρ k = ρ(v k, τ k, φ k) calculate the static efficiencies
5) yxk = yx(ρ k) assign the investment flows
6) x k+1 = yxk update the investment flows
7) s k = s(v k, φ k) calculate the arc times
8) t k = t(v k, τ k, φ k) calculate the arc tolls
9) s k = s k +η k perturb the arc running times (Probit)
10) (w k, p k) = r(s k, t k, φ k) calculate node satisfaction and probabilities
11) d k = d(w k) calculate the demand
12) yvk = yv( p k, d k, φ k) , yτk = yτ( p k, d k, φ k) assign the demand to the network
13) v k+1 = v k +1/k +(yvk -v k) , τ k+1 = τ k +1/k +(yτk -τ k) update the demand flows
14) x k+1 = x k +1/k +(yxk -x k) update the investment flows
15) if max a∈A |va - va | / va < ε then stop
k k-1 k
stop criterion
16) go to 2
As step 3 shows, calculating the frequencies needs G(φ) to be inverted. To simplify
the calculation procedure, we invert its linear approximation so obtaining:
φ lk = φ lk-1 +(xlk-xlk-1)/[∂G(φ)/∂φ l]k-1 l = 1, ... , |L| . (39)
In conclusion, a practical consideration. Although the most natural way to assign
the budget to the investment network could appear to be an all-or-nothing network
loading algorithm, where the opposite of each static efficiency function is assumed as
the corresponding investment arc cost function, since in this case the investments in
frequencies are assigned one at time, dealing with real-size networks, this is not
efficient. Instead, a Logit model and the inverse of static efficiency plus 1 can be
conveniently used, as we do in this paper.

Numerical results

In performing the numerical applications, some simplifications not affecting the

validity of the results have been introduced to the simulation model. In particular,
only one class of users has been considered and the elasticity with respect to
destination has been suppressed. The well-known Sioux Falls network in its bimodal
version (LeBlanc, 1988), consisting of 24 nodes, all centroids, 76 road arcs, 76
pedestrian arcs, and 5 transit lines, has been used as test network. Indeed, it is an
aggregate network, exhibiting a high congestion level in correspondence of the
considered 180000 passengers per hour total demand.
The simulation has been replicated for the following scenarios: User Equilibrium
(UE) and System Equilibrium (SE) with no investment; Frequencies Design with a
budget constraint of 50 millions $ (FD B) and with no budget constraint (FD)
combined with tolls optimization. In all the cases two extreme hypotheses regarding
the transit lines have been considered: mixed traffic and dedicated lanes with φmin = 0
and φmax = 60 veic/h in both cases. No external costs, like noise or emission
pollution, is considered in numerical examples, so underestimating the social surplus
variations.
The effectiveness of the method has been tested considering the following
output: demand trip generation; modal shift, road and transit average speed, road and

6 Bellei, Gentile, Papola

transit pricing, road and transit revenue, social surplus, and investment. The results
of the different scenarios are synthesized in Figure 1. Although the data are
sufficiently self-explicative, some remarks can be useful.
• a pricing scheme that leaves unaltered the transit capacity (SE scheme) has a
modest efficacy in terms of modal split and a negative impact on the demand due
to the strong reduction of the trip generation;
• the effectiveness of the intervention decisively improves when acting on tolls and
transit frequencies jointly. Already in the FDB case, a dramatic increase of the
modal split and a significant increase of the generation is obtained;
• not surprising, in the FD case the investment is three times higher than that of the
FBD case. This is because in the UE case an high congestion level has been
hypothesized, jointly with a scarcely efficient transit mode.

Conclusions

In this paper a network design problem consisting in tolls and transit lines
frequencies optimization, which is by its nature a bilevel problem, has been reduced
to a specific equivalent equilibrium assignment problem, which allows stating the
existence of a global solution given the continuity of the involved functions. On the
other hand, the Jacobian of the arc cost functions is non-positive semi-definite and
the problem may have more than one local solution.
After having introduced the concept of investment network, a specific fixed-
point approach algorithm has been also devised, based on a hyper-network consisting
of both transportation and investment networks. The algorithm, particularly efficient,
needs the transit operating cost function G(φ) to be inverted, a difficulty overcome
hereby using its linear approximation.
The simulation results, though relative to a test network, are decisively
encouraging from two interrelated points of view. Because they are definitely
plausible, they ensure the reliability of our approach in terms of both the modeling
and the calculation procedure. For this same reason, they constitute, as they are, a
useful insight into the real transportation problems in those situation, where a high
congestion level calls for some coordinated scheme, involving private and transit
mode, to ensure the best use of an existing network.

References
Bellei G. , Gentile G. , Papola N. (1998) Ottimizzazione del Trasporto Urbano in
Contesto Multiutente e Multimodo Mediante l’Introduzione di Pedaggi (Urban
Network Optimization in Multi-User and Multi-Modal Contexts Through the
Introduction of Tolls). Primo Convegno su Metodi e Tecnologie dell’Ingegneria dei
Trasporti, Reggio Calabria, Italia.
Bertsekas D. P. (1995) Nonlinear Programming. Athena Scientific, Belmont.
Cantarella G. E. (1997) A General Fixed-Point Approach to Multimode Multi-User
Equilibrium Assignment with Elastic Demand. Transpn. Sci. 31, 107-128.
LeBlanc, L. J. (1988) Transyt System Network Design. Transpn. Res. 22B, 383-390.
Nguyen S. , Pallottino S. , Gendreau M. (1998) Implicit Enumeration of Hyperpaths

7 Bellei, Gentile, Papola

in a Logit Model for Transit Networks. Transpn. Sci. 32, 54-64.
Sheffi Y. (1985) Urban Transportation Networks: Equilibrium Analysis with
Mathematical Programming Methods. Prentice-Hall, Englewood Cliffs, NJ.
Yang H. (1997) Sensitivity Analysis for the Elastic-Demand Network Equilibrium
Problem with Applications. Transpn. Res. 31B, 55-70.

8 Bellei, Gentile, Papola

 Generation Modal Shift
85,0% 45,0%
40,0%
80,0% 35,0%
30,0%
75,0% 25,0%
20,0%
70,0%
15,0%
65,0% 10,0%
5,0%
60,0% 0,0%
UE SE FD B FD UE SE FD B FD
Mixed Traffic 79,7% 69,9% 75,2% 80,6% Mixed Traffic 12,4% 15,4% 28,6% 40,3%
Dedicated Lane 80,3% 70,6% 76,8% 82,7% Dedicated Lane 14,4% 15,6% 28,7% 39,9%

Average Road Speed Average Transit Speed

35,0 25,0
30,0
20,0
Km / h

25,0

km / h
20,0 15,0

15,0 10,0
10,0
5,0
5,0
0,0 0,0
UE SE FD B FD UE SE FD B FD
Mixed Traffic 20,3 27,0 29,8 33,1 Mixed Traffic 10,0 13,2 15,1 17,6
Dedicated Lane 20,6 26,7 28,9 31,8 Dedicated Lane 16,3 19,4 20,4 22,7

Average Road Pricing Average Transit Pricing

0,70 1,00
0,60
0,80
0,50
$ / Km
$ / km

0,40 0,60

0,30 0,40
0,20
0,20
0,10
0,00 0,00
UE SE FD B FD UE SE FD B FD
Mixed Traffic 0,00 0,62 0,55 0,47 Mixed Traffic 0,00 0,72 0,32 0,12
Dedicated Lane 0,00 0,59 0,48 0,36 Dedicated Lane 0,00 0,87 0,38 0,12

Road Network Revenue Transit Network Revenue

1000 250
Millions of $ / year

Millions of$ / year

800 200

600 150

400 100

200 50

0 0
UE SE FD B FD UE SE FD B FD
Mixed Traffic 0 883 732 559 Mixed Traffic 0 117 160 104
Dedicated Lane 0 850 652 441 Dedicated Lane 0 150 200 118

Social Surplus Investiment

4500 180
Millions of $ / year
Millions of $ / year

4000 160
3500 140
3000 120
2500 100
2000 80
1500 60
1000 40
500 20
0 0
UE SE FD B FD UE SE FD B FD
Mixed Traffic 3455 3727 3950 4082 Mixed Traffic 0 0 50 147
Dedicated Lane 3511 3771 4036 4176 Dedicated Lane 0 0 50 163

Figure 1. Numerical results mixed traffic dedicated line

9 Bellei, Gentile, Papola