Transportation Research Part C 43 (2014) 267–279

Contents lists available at ScienceDirect

Transportation Research Part C

journal homepage: www.elsevier.com/locate/trc

Trade-offs between mobility and equity maximization

under environmental capacity constraints: A case study
of an integrated multi-objective model
Tao Feng ⇑, Harry J.P. Timmermans
Urban Planning Group, Department of the Built Environment, Eindhoven University of Technology, PO Box 513, Vertigo 8.16, 5600MB Eindhoven, The Netherlands

a r t i c l e i n f o a b s t r a c t

Article history: This paper investigates the performance of a policy decision tool proposed for multi-objec-
Received 10 July 2013 tive decision under different policy interventions. This tool deals with the trade-off
Received in revised form 20 March 2014 between mobility and equity maximization under environmental capacity constraints.
Accepted 21 March 2014
Two system objectives, maximization of mobility and equity, are formulated in terms of
the sum of total car ownership and number of trips, and the differences in accessibility
between zones. Environmental capacities are based on production efficiency theory in
Keywords:
which the frontier emission under maximum system efficiency is taken as environmental
Policy decision making
Environmental capacity
capacity. To examine the performance of the proposed model, three types of hypothetical
Mobility policies (network improvement, population increase and urban sprawl) are formulated.
Accessibility-based equity Effects are simulated using data pertaining to Dalian City, China. Results show that the pro-
Car ownership posed model is capable of representing the trade-offs between mobility and equity based
on different policy interventions. Compared with two extreme cases with the single objec-
tive of mobility maximization or equity maximization, the Pareto-optimal solutions pro-
vide more interesting practical options for decision makers. Taking the solution based on
the maximum equity as an example, the policy of urban sprawl yields the most significant
improvement in both emission and accessibility of the three scenarios.
Ó 2014 Elsevier Ltd. All rights reserved.

1. Introduction: Policy context

The rapid expansion of car ownership and the increase of car dependency have been main reasons for worldwide urban
problems such as congestion, accidents, and emissions. In most developed and developing cities, automobiles have become
the most important source of urban pollution (e.g. Cai and Xie, 2007; Gao and Niemeier, 2008). In spite of the progress in
developing environmental friendly vehicles (i.e., electric vehicles, new type of fuel resource, automatic driving), technology
alone does not provide a sufficient solution. A more integrated approach is needed, especially in developing cities where peo-
ple will continue pursuing the ideal of motorized convenience when their economic situation improves (Schafer, 1998).
Urban planning and mobility management should therefore consider the interrelationships between urban development,
land use and transport demand generation.
Especially the situation in large metropolitan areas in developing countries such as China has become increasingly more
problematic. These cities have attracted many new people from the countryside seeking job opportunities. This has resulted

⇑ Corresponding author. Tel.: +31 40 247 2301; fax: +31 40 243 8488.
E-mail addresses: t.feng.@tue.nl (T. Feng), h.j.p.timmermans@tue.nl (H.J.P. Timmermans).

http://dx.doi.org/10.1016/j.trc.2014.03.012
0968-090X/Ó 2014 Elsevier Ltd. All rights reserved.
268 T. Feng, H.J.P. Timmermans / Transportation Research Part C 43 (2014) 267–279

in a rapid growth of population, firms and infrastructure. Although investments in public transport and high density neigh-
borhoods have been important planning ingredients, nevertheless the changing land use requirements have resulted in sub-
stantial urban sprawl. In turn, urban sprawl implies longer distances and this in turn has triggered reliance on cars and the
road network, with congestion and adverse environmental impacts as negative externalities. Obviously, improved economic
conditions have meant that an increasing share of the population can afford cars.
Urban sprawl, increased car ownership and use, and the resulting environmental impact are however not the only policy
issue in this context. Substantial differences in welfare and significant spatial difference in these metropolitan cities in terms
of service and transport provision imply that policy makers also face a distributional problem of the impacts of the policies.
This is the issue of equity (or inequality). Equity is commonly understood as fairness or justice of the distribution of the
impacts (both benefits and costs) of an action on two or more subgroups (Litman, 2002). In the field of urban and transpor-
tation planning, equity can be examined from either a social or a spatial perspective. Social equity basically refers to differ-
ences in income or social welfare between individuals. Spatial equity indicates differences in the spatial distribution of
transportation services (e.g., travel time, cost, distance, and number of transfers) or other services. Different from social
equity, spatial equity represents the differences which could be affected by zonal/regional mobility levels, transport network
conditions, inter-zonal mode choice, land use topology, etc. This paper is concerned with spatial equity which is defined
using a formal indicator based on zonal accessibility.
Recognizing the importance of motorized mobility and spatial equity, a vast literature has contributed to this topic area.
The relationship between urban development and mobility has been discussed from different perspectives, such as car own-
ership (Kitamura, 1987; Meurs, 1993), and car use (Steg et al., 2001; Gardner and Abraham, 2008). Spatial equity has been
mainly discussed in the context of the network design problem where equity has been generally used as one of the constraint
conditions (Meng and Yang, 2002; Santos et al., 2008). To position the present paper, it should be emphasized that these dif-
ferent literatures have been isolated; there is a paucity of research which evaluates the dynamic interactions between urban
development, mobility, environmental impact and equity in a comprehensive manner. Arguing that mobility levels should be
at their maximum while keeping inequitable influences at their minimum, policy development may be assisted by develop-
ing an integrated modeling framework, which incorporates evaluations of both mobility and equity and their trade-offs.
To that end, we have formulated a policy decision support tool that allows decision makers to identify maximum mobility
levels under environmental capacity constraints (Feng et al., 2008). The applicability of the model has been verified in a case
study using real network data. Moreover, model performance under different policy interventions has been examined (Feng
et al., 2010). Previous publications have verified the effectiveness of the modeling framework as a policy decision tool, how-
ever, the model has not been tested when dealing with multiple policy concerns. The purpose of this study therefore is to
investigate the performance of an enhanced integrated model proposed for multi-objective policy decision making. The
model addresses the trade-offs between mobility and equity maximization under environmental capacity constraints. The
idea behind the multi-criteria formulation is to identify the maximum mobility level while keeping the maximum equity
of accessibilities. Specifically, it deals with two strategic targets, maximization of (i) motorized mobility and (ii) accessibil-
ity-based equity, which are respectively formulated in terms of total car ownership and number of trips, and differences
between zonal accessibility. Based on efficiency theory, environmental capacity is derived from frontier emissions under
the most environmentally efficient system.
To investigate the performance of the proposed model, three types of hypothetical policies, network improvement (NI),
population increase (PI) and urban sprawl (URS), are designed and the effects of these policy scenarios are simulated using
data of Dalian City, China.
The remainder of this paper is organized as follows: Section 2 presents the integrated model and details of its compo-
nents. A genetic algorithm designed for the multi-objective optimization problem is developed. A case study and the relevant
designed policy scenarios are introduced in Section 3. Results of analyses are discussed in Section 4. The paper is completed
with a summary and discussion in Section 5.

2. The integrated model

2.1. Scope

The integrated model proposed here serves two optimization objectives: maximization of mobility which is measured in
terms of total car ownership (i.e. the number of cars) and total number of trips, and maximization of spatial equity which is
defined and measured using a zonal accessibility measure. The modeling process is based on a bi-level programming method
which follows an iterated process to obtain the optimal solutions under the constraint that the environmental load does not
exceed environmental capacity. The Pareto-optimal solutions which relate to a vector of zonal car ownerships are obtained
at the upper level of the bi-level programming problem, subject to the constraint that the environmental load should not be
larger than the corresponding environmental capacity. Zonal car ownership is used to calculate inter-zonal transport mode
choice probabilities which in turn are used to calculate the origin–destination (O–D) trip matrix. The resultant O–D matrix is
then assigned to the road network in the lower level problem, resulting in a distribution of traffic flows and associated emis-
sion levels which provide the inputs for the upper level problem. A flowchart of the whole modeling process is depicted in
Fig. 1.
 T. Feng, H.J.P. Timmermans / Transportation Research Part C 43 (2014) 267–279 269

Fig. 1. Flowchart of the integrated multi-objective model system.

2.2. Definition and notation

Assume a city can be divided into a certain number of traffic zones. Let i be an index for these zones. Each zone has a
certain level of population, popi, and density of road network, di. People travelling by car from origin i to destination j produce
emissions along route rij which can be counted as a distribution in different zones. For each of the zones, it has an associated
environmental threshold or environmental capacity, ECi, which defines the maximum pollution level at the zone. The envi-
ronmental capacity of transportation at zone i partly depends on the level of infrastructure (i.e. road density, di) and trans-
port pattern (travel by car or public transport mode). The environmental pollution at zone i is determined by the level and
the composition of road traffic flow. For some of the zones with a higher density of residents, a higher environmental quality
is required. Reducing the difference between environmental capacity and traffic pollution will always be an important topic
for policy makers who need to smartly balance transport supply (capacity) and mobility demand.
Here, the mobility is expressed in terms of zonal car ownership ui and trip demand between O–D pairs (i, j) by car and
public mode (respectivelyqcij and q cij ). To calculate the emissions, this demand is distributed and assigned to the road network,
using an entropy model. Emissions at the various links of the network are then calculated as a function of link length, traffic
volume and emission factors that depend on average driving speed on each link. In the following subsections, details of these
various components will be explained.

2.3. Upper level problem: Trade-off between mobility and equity

As presented in Fig. 1, the integrated model is composed of two levels of problems, named the upper level problem (ULP)
and lower level problem (LLP). The ULP is a multi-objective optimization problem which can be formulated as below:
Maximize:
X XX
f1 ðui Þ ¼ ku ui þ kv ðucij  qcij þ u
 cij  q
cij Þ ð1Þ
i2I i j

Minimize:
1 XX
f2 ðui Þ ¼ jAccj  Acci j ð2Þ
2N2 Acc j2N i2N

Subject to:
Ei ðui Þ 6 E0i ; i 2 I ð3Þ

0 6 ui 6 u0i ; i 2 I ð4Þ
X
Acci ¼ ðpopj =t ij ðui ÞÞ; i 2 I ð5Þ
j–i
270 T. Feng, H.J.P. Timmermans / Transportation Research Part C 43 (2014) 267–279

where f1 and f2 are two optimization objectives, representing mobility and spatial equity, respectively; ui represents the car
ownership in zone i, and u0i represents the maximum car ownership in zone i; qcij and q cij are trip demand between O–D pair
c c
(i, j) by car and public mode, respectively; ku ; kv ; uij and u  ij are the pre-defined parameters relating to the zonal car owner-
ship and the number of trips by car and public transportation modes; N represents the total number of zones; Ei represents
the total emission of zone i, and E0i represents the emission capacity of zone i; Acci and Accj are the accessibility at zone i and
j, respectively; Acc is the average accessibility across all zones; popj is the population of zone j and tij is the average travel
time between zone i and zone j; I is the set of zones.
The objective functions are composed of maximization of mobility (f1) and equity (f2). This multi-objective optimization
problem implies that policy makers wish to maximize the mobility level (here referring to the total number of car ownership
and trips) on the one hand, and maximize the accessibility-based equity (decrease accessibility differences) on the other. The
trade-off effects between these two objectives can be measured by the set of Pareto-optimal solutions. A solution is Pareto-
optimal when each objective component of any solution can only be improved by degrading at least one of its other objective
components. Considering the possible conflicts between different system targets in practice, Pareto solutions provide policy
makers information about the trade-offs between mobility and equity where various weighting schemes can be applied.
The mobility level can be evaluated from various perspectives, e.g. number of available cars per capita, mode availability,
or travel cost. To be consistent mathematically with the modal split model and the environmental capacity model of which
car ownership is one of the variables, we define motorized mobility (f1) by the total of car ownership and number of trips by
car mode and public mode. Such an approximation is consistent with the requirement of mobility management such that
policy makers do not control the vehicle purchases but influence their excessive use by increasing cost through different pol-
icies. The parameters ku, kv, ucij , and u
 cij reflect different weights for policy goals which should be defined based on consensus
building among all stakeholders related to the targeted policies. For instance, parameters ku and kv can also be used to reflect
different emphasis between car ownership and car use (trips). Moreover, larger values of u  cij may result in more trips by pub-
lic mode and less car trips while ensuring that the total number of trips reaches its maximum.
The measurement of spatial equity has been simulated before as the impacts on different individuals or groups caused by
changes in transportation services. Studies have represented the differences between travel time (Yang and Zhang, 2002;
Meng and Yang, 2002) and/or travel cost (Szeto and Lo, 2006, 2008; Lo and Szeto, 2009) in the context of the network design
problem and/or road pricing assumptions. As an alternative measurement, the formal type of indicators (e.g. Gini coefficient
or Theil index), which are capable of representing the distributional differences of transportation resources and services in a
statistical manner, have been used recently in a number of studies (Connors et al., 2005; Santos et al., 2008; Duthie and
Waller, 2008). Considering the popularity and statistical characteristic, we chose the Gini coefficient to represent equity
in this paper.
As shown in Eq. (2), equity (f2) is defined as a function of zonal accessibility (Acci). The reason to use accessibility rather
than other measures is due to the importance and generality of accessibility in urban planning. In general, accessibility can
be defined at either the individual level or zone level. Individual-based accessibility concerns opportunities that an individ-
ual at a given location possesses to participate in a particular activity or a set of activities (e.g., Odoki et al., 2001). However, it
suffers from the disadvantage that it is data-intensive. Conventional location-based accessibility measures would be more
operational. Such accessibility measures have been applied in zonal-based travel demand analysis based on gravity-type trip
distribution models. Here accessibility is measured at the aggregate level in the sense that the overall population at zones is
considered without differentiation. More specifically, accessibility is formulated as a function of zonal population popj and
inter-zonal travel time tij, as shown in Eq. (5).
The Gini coefficient indicates the degree of distributional differences between zonal accessibilities. The values of Gini
coefficient are between 0 and 1, a lower coefficient indicating a more equal accessibility distribution, and a higher coefficient
indicating a less equal distribution. The value ‘‘0’’ corresponds to perfect equity, meaning that each zone has the same level of
accessibility, and ‘‘1’’ corresponds to perfect inequity, indicating that only one zone gets the accessibility, while all other
zones have a zero accessibility level. Of course, the values of 0 and 1 are two extreme theoretical cases which are impossible
to happen in practice.
The optimization problem follows the constraint condition that the emission in each zone i (Ei(ui)) is less than EC (E0i). The
zonal emission is mathematically expressed as a function of zonal car ownership (ui) in which the link volume for emission
calculation is obtained through a traffic assignment procedure where the given trip demand are dependent on the level of car
ownership. In this sense, the outputs of the lower level problem (assigned link volume, emission, speed) are dependent on
the level of zonal car ownership which is the optimal solution of upper-level problem. Therefore, the travel time between
zones (tij) can be also represented as a function of zonal car ownership (ui). A detailed mathematical illustration on the inher-
ent mechanism and interactions between the decision variables of the upper and lower level problems can be clearly rep-
resented by transforming the traffic assignment into a variational inequality problem (Tobin and Friesz, 1988).
In addition, the decision variables, zonal car ownership (ui) at zone i, are specified within the range of [0, u0i]. The upper
bound of the constraint (u0i) is given by, for example, taking into account the actual limitations of the zonal population. This
limit could be equal to or larger than the maximal car ownership derived from the bilevel model. This limit was introduced
due to the assumption that car ownership per capita should be within a certain range. The number of trips can be calculated
from the travel demand between O–D pairs (Qij) and mode choice probability (Pcij ), which are shown as follows:
qcij ðui Þ ¼ Q ij  P cij ðui Þ ð6Þ
 T. Feng, H.J.P. Timmermans / Transportation Research Part C 43 (2014) 267–279 271

cij ðui Þ ¼ Q ij  ð1  Pcij ðui ÞÞ

q ð7Þ

where Qij represents the total trips between O–D pair (i, j); Pcij
represents the probability of choosing car mode from zone i to
zone j.
Here, the mode choice probability between zone i and j is also a function of zonal car ownership. To link the lower and
upper levels, an aggregate Logit model is proposed to reflect the fact that mode choice probabilities are determined by car
ownership levels at the origins, inter-zonal travel times and zonal land use characteristics. The probability of car trips
between zones i and j is measured as follow:
expðU cij Þ
Pcij ðui Þ ¼ ð8Þ
1 þ expðU cij Þ

U cij ¼ V cij þ ei ð9Þ

V cij ¼ b0 þ b1  t cij þ b2  ui þ b3  indui þ b4  commi ð10Þ

where V cij and V

 c are deterministic term of the utility, U c , of choosing and not choosing the car from zone i to j, respectively; ei
ij ij
represents the error term; t cij represents the travel time by car between zone i and j; indui and commi are dummy variables of
land use for industry and commerce, respectively; b0 and bi are parameters need to be estimated.
Unlike disaggregate choice models that require choice data at the individual level, aggregate models are used to represent
the accumulated results of individual choices at the zonal level where the traffic analysis zones (TAZ) are taken as the choice
alternatives. Consequently, the dependent variable of the aggregate model under study is the average modal share of differ-
ent travel modes. Given the probability of a car trip, the trip probability by public mode can be simply obtained by using Eq.
(7). Note that we differentiate between car and public mode only because car trips make up a big share of the pollution. Traf-
fic flows and emissions associated with public mode are not dealt with in this study. Thus, only the car trip matrix is incor-
porated in the traffic assignment in the lower level problem.
The emission from each link is calculated as the product of link length, traffic volume and emission factors that depend on
the average driving speed on each link, calculated as follows:
X
Ei ðui Þ ¼ e a ; Ai 2 A ð11Þ
a2Ai

ea ¼ cak  v a  la ; a 2 A; k 2 K ð12Þ
where ea represents the emission on link a; cak represents the emission factor of category k on link a and k indicates travel
speed category; va represents the link volume; la represents the length of link a.
The emission factors are based on the existing literature. The emission Ei from zone i equals the sum of emissions from the
links (ea) which belong to zone i. Therefore, the area and spatial location of each zone affect the emission and concentration
level. For instance, the zones within the central business district (CBD) mostly have a high density road network and traffic
flow, and consequently show high emission/pollution concentrations, although their areas are small. In contrast, suburban
zones are usually large in scale, but have lower road network densities and less pollution than CBD zones. Considering the
complexity of pollutant diffusion and spatial differences, only emission constraints are included in this study.
Calculation of environmental capacity (EC) at the zonal level would be helpful to effectively control pollution considering
the specific characteristics of different areas. Although EC can be understood in terms of either emissions or concentrations,
here we define EC by emissions. More specifically, the theory of efficiency analysis is used to specify the emission capacity by
assuming there is no inefficiency in the ideal state of transportation system (Feng et al., 2008). Here, we adopt the stochastic
frontier analysis (SFA) model (Kumbhakar and Lovell, 2000) with multiple inputs and single output. The logarithm form of
Cobb–Douglas equation is adopted for measuring system efficiency:
X
ln yi ¼ h0 þ hm ln xim þ gi ð13Þ
m2M

gi ¼ xi þ li ð14Þ
where yi represents the total amount of car emissions in city or zone i; xim represents the mth input variable in city or zone i;
M represents the set of input variables; h0, hm are unknown parameters; gi is a composite error term; li is the non-negative
technical inefficiency component; xi is the two-sided random-noise component.
Different from the deterministic frontier model, here, the error term in stochastic frontier model has two components.
This makes it possible to measure the random effect outside of the control of producers. The noise component of xi is
assumed to be distributed independently of ui. The parameters in Eqs. (12) and (13) can be estimated by the maximum like-
lihood method.
Because we expect the emission to be lower giving the same level of inputs, the sign of li in the error term is positive. The
measure for environmental efficiency CEi is given by
272 T. Feng, H.J.P. Timmermans / Transportation Research Part C 43 (2014) 267–279

CEi ¼ expðli Þ ð15Þ

Here, CEi reflects the grade of inefficiency with a value between zero and one. The smaller the value is, the more efficient
the transportation system. In the context of environmental efficiency, the emitted pollutants can be viewed as the produc-
tion output. The transportation environment efficiency indicates the technical efficiency of the transportation system from
the viewpoint of traffic pollution. The input variables can be factors related to emission, such as car ownership, road network
construction, population increase, etc. Then, the frontier emission, ECi, indicating environment capacity, can be calculated as
EC i ¼ yi  CEi ð16Þ
The system inputs used in SFA are indicators of transport mobility and network density while the outputs are the
observed car emissions. Consequently, the frontier emission calculated by current explanatory variables and validated
parameters is regarded to specify EC. In this sense, the maximum of the system target is specified at the most efficient level
rather than the maximum level. The system which has the highest efficiency will have the worst environmental performance
where the emission reaches its maximum. Therefore, the concept of maximum can be also understood as the environmen-
tally most efficient level.

2.4. Lower level problem: A combined distribution and assignment model

Given the updated O–D trip matrix, the traffic flow distribution can be observed by a traditional traffic assignment pro-
cess. Due to the consideration of trip distribution among zones, and the fact that the O–D trip demand is fixed, the lower
level problem represents a combined distribution and assignment model.
Minimize:
X Z va 1 XX c
ca ðxÞdx þ ðqij ln qcij  qcij Þ ð17Þ
a 0 n i j

Subject to:
X ij
fh ¼ qcij ; i 2 I; j 2 J ð18Þ
h2H

X
qcij ¼ Oi ; i 2 I ð19Þ
j2J

X
qcij ¼ Dj ; j 2 J ð20Þ
i2I

X
va ¼ fh dah ; a 2 A; h 2 H ð21Þ
h2H

fh ; qcij P 0; h 2 H; i 2 I; j 2 J ð22Þ

where ca represents the travel time on link a; fh represents the traffic flow on path h; Oi represents the trip generation by car
in origin zone i and Dj represents the trip attraction by car in destination j; n is the dispersion parameter for the trip distri-
bution model; J and H are the set of zones and paths, respectively.
The flow distribution is constructed by using an entropy model where the path flows can be any combination of traffic
flows between O–D pairs and each combination is called a state. All states are equally likely to occur, and the flow with the
highest likelihood is the set with the maximum number of states. The model is a doubly constrained model in that both the
total flow generated at origins and the total flow attracted to destinations are fixed and known. The solution of this model is
a set of O–D trip rates and link flows which satisfy both the principle of user equilibrium and entropy maximization. We
solve the lower level problem by using the algorithm described by Sheffi (1985).

2.5. Solution method

A number of algorithms have been proposed in the literature to solve the bilevel programming problem. However,
because of the intrinsic non-convex characteristics associated with the bilevel programming problem (Bard, 1999), the algo-
rithms commonly suffer from the difficulty to theoretically ensure solutions as the global optimum. Considering the oper-
ational simplicity in actual applications, here we therefore adopt the framework of genetic algorithms (GA). The effectiveness
of GA has been verified in many studies (e.g., Leblanc, 1975; Karlaftis et al., 2004; Vlahogianni et al., 2005, 2007; Duthie and
Waller, 2008; Feng et al., 2008; Yu et al., 2010).
As the upper level problem is a multi-objective optimization issue, a multi-objective GA is specifically adopted here. Even
though there is no single best solution with respect to both objectives, the non-dominated solutions or Pareto-optimal solu-
tions can be obtained. We use the non-dominated sorting genetic algorithm, NSGA-II, as the main engine. NSGA-II is the
 T. Feng, H.J.P. Timmermans / Transportation Research Part C 43 (2014) 267–279 273

second version of the famous non-dominated sorting genetic algorithms for solving non-convex and non-smooth single and
multi-objective optimization problems. It implements a procedure where all individual are sorted according to the level of
non-domination. It implements elitism which stores all non-dominated solutions, which enhances the convergence capabil-
ity. In addition, it adapts a suitable automatic mechanics based on the crowding distance in order to guarantee diversity and
spread of solutions (Deb, 2001).
The elitism in NSGA-II can speed up the performance of the GA significantly, and can also prevent the loss of good solu-
tions once they are found. In addition, it can provide various Pareto-optimal solutions in a single run and consequently the
burden of performing multiple runs for various values of weights can be reduced. NSGA-II also has the capability of con-
straint handling, which is useful given our requirements. The NSGA-II has been applied by scholars in different disciplines
(Murugan et al., 2009; Han et al., 2014). In the numerical example, a traffic assignment modular is embedded in the
NSGA-II program to calculate fitness values.
The initial zonal car ownership is used as an input for the aggregate modal split model. Before carrying out the calculation
of the BL programming problem (for details see for example, Bard (1999)), the current car O–D matrix is initially assigned to
the road network. The combined trip distribution and assignment problem can be solved using Frank-Wolfe approach
besides the additional effort of dealing with the Hitchcock problem which is commonly described as a minimum-cost flow
problem. The Hitchcock problem can be described as a minimum-cost flow problem over the network where the aim is to
specify the necessary links and related flows. The number of links carrying flow, at optimality, equals the minimum number
of links that can connect origins to destination.
The traffic volume and average travel speed on links are taken as the basis of estimating link emissions and the accessi-
bility-based equities. Then, the emission in zones can be calculated by summing all relevant links within the zone. The zonal
travel time were calculated using the link based travel time. To find the optimal car ownership, an iterative process is imple-
mented until the convergence criterion is reached.
Regarding inequality constraints in the multi-objective optimization, the fitness function incorporates the constraint vio-
lations by means of a penalty method. The lower level problem is solved by conventional optimization techniques. The
NSGA-II algorithm is incorporated in solving the upper optimization problem where the penalty function was considered
as an indicator of dominance. The penalty function was designed based on the principle: when the zonal emission jumps
over the related capacity in iteration n, the difference will be multiplied by an infinite number, h. Then, the objective value
of fitness function will be negatively transformed because of the large penalty. The fitness function with penalty can be
expressed as follows:
X XX 1 XX X
gðnÞ ¼ ku ui;n þ kv ðucij  qcij;n þ u
 cij  q
cij;n Þ þ 2
jAccj  Acci j þ h  ðE0i;n  Ei;n Þ ð23Þ
i2I i j 2N Acc j2N i2N i2I

where n indicates the iteration time, and h is the infinite number. The penalized excess of zonal emission should be nega-
tively larger than the sum of other elements in Eq. (20) to ensure the fitness value be increased sequentially.

3. Case study

3.1. Data and parameter settings

The data was collected in Dalian City in 2004 in a research project about the comprehensive planning of the Dalian trans-
portation system, funded by the Dalian government. Dalian is one of the central cities in the Ring-Bohai economic region in
China and a leading city in stimulating the economic development of the whole province. In the central area of Dalian City,
more than seventy percent of daily trips are served by public transport modes such as bus, light rail and tram. Motorcycles
are rare and few bicycles are used. In recent years, the popularization speed of automobile in Dalian City has dramatically
increased. Between 2002 and 2005, the disposable average income increased by 11.8%, while the normal price for a car
decreased by 26.5% (DMBS, 2001–2005). The annual growth rate of private passenger cars increased fast, from 28% to
34.8%. In 2007 particularly, private passenger cars accounted for approximately 60% of the total increase in the number of
vehicles.
Fig. 2 shows the road network of the central urban area of Dalian City. The road network, which was simplified for the
sake of model simulation, includes 33 zones, 895 links and 544 nodes. The central area with a dense road network shown
by grey line covers a few zones, including zone 24, 25, 26 and 31. Recently, the region located near zone 5 has become
another business center with road and building construction for multiple functions.
Regarding the specification of EC, we adopted the Millennium Cities Database (Kenworthy and Laube, 2001) because it is
hard to get historical zonal data for Dalian City. The database covers the data of 100 cities worldwide concerning demograph-
ics, economics, urban structure and a large number of transport-related data. Since the emission in this study is defined at
the zonal level, whereas the database is at the city level, we validated the parameters in Eq. (13) using city data and esti-
mated the ECs for zones. The final EC model incorporated the variables number of cars and road network density. Details
regarding the calibration and parameter estimation using SFA can be found in Feng et al. (2008).
Due to limited data quality, the modal split model estimated from PT data is not suitable for prediction. Therefore, a sim-
plified calculation procedure was adopted where the trip probability from zone i to zone j was set equivalent to the ratio
274 T. Feng, H.J.P. Timmermans / Transportation Research Part C 43 (2014) 267–279

Fig. 2. Road network of Dalian City.

between car ownership and population. This calculation method imposes the assumption that people having cars would def-
initely choose the car mode for travel, ignoring any effects of possible congestion.
The trip matrix for cars is assigned to the road network, ignoring non-car trips. The link impedance function in the traffic
assignment is defined using the Bureau of Public Roads (BPR) function of the following form:
(  4 )
va
ta ðv a Þ ¼ t 0  1:0 þ 0:15 ð24Þ
Va

where ta and t0 are travel time on link a and travel time under free flow, respectively. Va is the volume capacity on link a.
Regarding the various emitted pollutants from road traffic, only CO was taken into account in this research project
because it is responsible for the largest share of urban pollution (Cai and Xie, 2007). In spite that emissions are dependent
on different factors, like the type of cars, speed, loading rate, etc., we do not have the data regarding the specifics of cars.
Therefore, we use the emission factors based on average travel speed are from the existing literature (Feng et al., 2003),
as shown in Table 1.
The parameters related to car ownership and trip frequency (ku , kv , ucij and u  cij ) were set to 1, which means that the
weighted importance between total car ownership and total number of trips, and trips between public and private mode,
are ignored. To further simplify, the dispersion parameter, n, for the trip distribution was also set to 1. Considering that

Table 1
Emission factors by travel speed (g/km).

Pollutant (g) Speed (km/h)

c < 15 15 6 c > 20 20 6 c > 25 25 6 c > 35 35 6 c > 45 c P 50
CO 84.7 58.8 51.6 40.1 29.8 26.2
 T. Feng, H.J.P. Timmermans / Transportation Research Part C 43 (2014) 267–279 275

theoretically there are a huge number of possible solutions, a restricted solution space is generated in advance for simulation
purposes. For each zone, a vector of possible solutions composed of integers with fixed intervals between two values for each
variable was generated at random. The solutions for all zones were then randomly combined without duplication.

3.2. Policy design

Because data related to future policies are scarce, we designed three hypothetical policy scenarios: road network
improvement, population change (total population in zones) and urban sprawl, as shown in Table 2.
The first scenario S1 represents the benchmark case of no policy intervention and is used as a reference to compare the
results of other scenarios. S2 relate to network improvement and assumes that the capacity of each link will be increased by
20%. S3 assumes an increase in the population in each zone by 10%. The scenario of urban sprawl (S4) is based on the assump-
tion that residents living in the central area move to suburban areas due to high land prices and costs of living in the city
center. We assume a 30% decrease and 30% increase for downtown and suburban areas, respectively. Population in transi-
tional zones between center and suburban areas is assumed not to change.
Since actual data related to future policies are scarce, we arbitrarily set the levels for all policies. The network
improvement (S2) is a common policy to alleviate congestion. The scenario related to population is important in that
the number of people moving from rural to urban areas in the progress of urbanization has been increasing in the past
years in most of the major cities in China. The increment of the proportion of the mobile population challenges the
capacity of the urban infrastructure and transportation system. Scenarios S2 and S3 both consider future increases in con-
gestion and resident population, while S4 indicates a natural re-distribution of population with the economic develop-
ment. Even if the location of living is a kind of result of people’s choices, the under-developed land around city area
can be carefully planned in advance, where the central or local government plays a role in the process of policy decision
making.
In evaluating model performance, the key indicators need to be defined. Here, we are primarily concerned with variations
in mobility and equity levels, and the associated traffic emissions and accessibilities. Therefore, emissions, car ownership and
accessibility were selected as the most relevant performance indicators. Before applying the model under different policy
intervention, the parameters in calculating the environmental capacity and modal split models were estimated in advance,
respectively. Details can be found in the previous paper (Feng et al., 2010).

4. Results and analyses

Before investigating the performance of the multi-objective model under policy scenarios, we implemented two
additional models, one with the single objective of mobility maximization and one with the single objective of equity
maximization. These two models are treated as two extreme cases compared with the multi-objective model. The single
objective model should indicate the maximum level of car ownership under the environmental capacity constraint,
provided that the system intends to keep the objectives of mobility maximization and equity maximization, respectively.
Table 3 gives the results for total car ownership and the Gini coefficient for these two single-objective models. Values
in the main diagonal cells in bold correspond to the optimal solutions with respect to respectively mobility and equity
maximization. In that sense, the car ownership value from Case 1 indicates the theoretically maximum level of car own-
ership that the transportation system can accommodate when taking the environmental capacity constraint into account.
It shows that under environmental capacity constraints, total car ownership for Case 1 approximately doubles that for
Case 2 where equity maximization is taken as the optimization objective. This means that, to keep the most equitable
distribution of accessibility across zones, car ownership needs to dramatically decrease relative to the Case 1. Since car
ownership in Case 1 depends on zonal pollution constraints, the optimal solution is equivalent to a redistributed level
of car ownership across zones. This may result in an inequitable accessibility distribution. For instance, a dramatic increase
in car ownership in some areas may induce extra congestion leading to decreasing accessibility in particular zones. There-
fore, the level of car ownership should be allocated such as to not only consider traffic pollution but also the accessibility-
based equity.
Although the difference in car ownership values between the two cases is obvious, the equity values (0.19033 and
0.18986) based on accessibility differ only marginally. This is due to the fact that a small change in the equity measure
may be indicative of a significant difference in network congestion. The scale of the equity values depends to some extent

Table 2
Design of policy scenarios.

ID Policy scenarios
S1 No policy
S2 Network improvement (NI)
S3 Population increase (PI)
S4 Urban sprawl (URS)
276 T. Feng, H.J.P. Timmermans / Transportation Research Part C 43 (2014) 267–279

Table 3
Total car ownerships and Gini coefficients for the single-objective models.

Cases Total car ownership Gini coefficient

Case 1 Mobility maximization 1,045,875 0.19033
Case 2 Equity maximization 508,350 0.18986

on the absolute value of accessibility which is in general here at a level of tens of thousands numerically. As a result,
differences between zonal accessibilities which are at a level of thousands cannot significantly affect the weighted Gini
value. In principle, the equity can be further formulated by attaching a vector of weight parameters to the accessibilities
of specific zones to represent different policy emphasis. This will certainly result in different numerical values for equity.
The details of the sensitivity of this type of equity measures have been investigated in the context of the network design
problem (Feng and Zhang, 2012).
The proposed multi-objective model is implemented based on the hypothetical policy scenarios. The Pareto-optimal solu-
tions obtained from the four cases are presented in Fig. 3. It shows a similar correlation trend between equity and total car
ownership among different cases. For each case, the total car ownership changes inversely with equity in the sense that the
increase of total car ownership relates to the decrease of accessibility-based equity. This uncovers the fact that policy makers
cannot ensure one objective as the optimal while without losing other benefits. In the specific correlation between equity
and motorized mobility, decision makers need to trade-off between the two systematic concerns according to the context
of specific requirements.
Although different scenarios have a similar pattern of correlations between mobility and equity, however, the equity val-
ues along horizontal axis vary among the four cases. This means that under the same level of total car ownership, different
policies may have different impacts on system equity. Oppositely, in keeping the same level of equity, different possible solu-
tions of car ownership exist according to different policy scenarios. Therefore, under the same level of environmental capac-
ity constraints, a more efficiently designed policy may result in a higher level of total car ownership. This further motivates
the necessity to examine the policy effectiveness in combination with the integrated model.
In order to examine the different impacts of policy scenarios, we present the descriptive statistics on Gini coefficients for
all cases, as shown in Table 4.
Total number of car ownership

Total number of car ownership

GINI coefficient GINI coefficient

(a) (b)
Total of number of car ownership
Total number of car ownership

GINI coefficient GINI coefficient

(c) (d)
Fig. 3. Pareto-optimal solutions by different policy scenarios (a) Solutions without policies, (b) solutions by policy of NI, (c) solutions by policy of PI, and (d)
solutions by policy of URS.
 T. Feng, H.J.P. Timmermans / Transportation Research Part C 43 (2014) 267–279 277

Table 4
Results of Gini coefficients by different policy scenarios.

Policy scenarios Average Min Max Max–Min

S1: No policy 0.19010 0.18986 0.19042 0.00056
S2: Network improvement 0.19004 0.18988 0.19024 0.00036
S3: Population increase 0.19006 0.18987 0.19032 0.00045
S4: Urban sprawl 0.18846 0.18830 0.18864 0.00034

It shows that the average Gini coefficients with policy interventions become smaller than that without policy. This means
all policies (network improvement, population increase and urban sprawl) contribute to equity improvement in different
extent. This can be also demonstrated by comparing the value difference between the minimum and the maximum where
S1 has the largest difference among solutions (0.00056). Regarding the policies relevant to population, the S3 has the least
equity improvement (0.19004 compared with 0.19010) in average among three policy scenarios. The population increase
in certain zones might result in high accessibility, but also higher car ownership in general which results in possible high
inter-zonal travel time because of traffic congestion. As an alternative, the policy of urban sprawl results in the most signif-
icant improvement in terms of equity among the three scenarios. This may indicate that urban sprawl would have a signif-
icant effect on decreasing the distributional difference of equity across zones.
Further, comparisons of the minimum equity values between S2, S3 and S1 found that the equity without policy impact
(0.18986) is not the least value among all cases. This suggests that policy scenarios would not always result in a more equi-
table accessibility distribution than the original situation without policies. However, there is a possibility that a feasible pol-
icy design may benefit to equity improvement in terms of the values in average. For the same type of policies, its
effectiveness partly depends on the specific design. In the assumed policy of S2 and S3, we treat all road segments and zones
as the same (the same level of capacity enhancement for all links, and the same proportion of population increase). Such type
of policy is helpful to validate the model performance, however, is not operational in reality. It will be expected in practice
that the policies designed by incorporating regional specifics proceed more efficient and effective to improve the difference
among zonal accessibilities.
In case of policy making in practice, decision makers need to determine the most feasible alternative from the vector of
Pareto-optimal solutions in terms of the trade-offs between mobility and equity. For the purpose of analysis, we choose, for
example, the solution related to the minimum equity value for each of the policy scenarios, as the optimum in subsequent
analyses. As a result, the relevant emission and accessibility are shown in Table 5.
It shows that both population related policies (S3 and S4) result in positive improvement according to emission and acces-
sibility. It is evident that S4 (urban sprawl) yields the most significant improvements on traffic emission, and the biggest
accessibility level among all scenarios. This means that a re-distribution of population and mobility can contribute not only
the emission alleviation, but also the increase of zonal accessibility. This is understandable in terms of the characteristics of
road network topology where some zones, i.e., zone 20 and zone 13 in the case study, have small population and network
density. The population moving from the central area to suburban area can be attributed to the improved accessibility in
suburban areas. The population decrease in central areas may on the other hand benefit from decease in travel time which
consequently results in the equitable balance before and after. Therefore, the distributional pattern of zonal population
which mainly influences the amount of motorized travel demand has a significant effect on equity.
As a comparison with S4, although population increase is expected to lead to the increased travel demand which may
impose much environmental burden, S3 does not result in higher emission than S1. This can be attributed to the simplicity
of modal split model where the over increase of population relative to car increase result in a low level of O–D trip demand
and the associated emissions. In this sense, the performance of the aggregated modal split model relevant to the change of
car ownership and population at zones plays a significant role in the context of population related policies. Moreover, the
outputs by S1 and S2 show slight difference in accessibility rather than emission, which means that the resulted inter-zonal
travel time between the two policy scenarios are different. It also indicates that network improvement scenarios should be
designed effectively according to the specific context. In addition, the introduction of environmental constraint in the pro-
posed model may have a stronger impact than network improvement on the traffic flow distribution in the sense that emis-
sion at zones should be less than the corresponding capacity, ignoring the effect of link capacity enhancement.

Table 5
Results of emission and accessibility by different policy scenarios.

Policy scenarios Emission (kg) Accessibility

Average Total Average Total
S1: No policy 54 1771 233,734 7,713,223
S2: Network improvement 54 1771 233,784 7,714,875
S3: Population increase 49 1610 257,141 8,485,648
S4: Urban sprawl 46 1532 260,508 8,596,748
278 T. Feng, H.J.P. Timmermans / Transportation Research Part C 43 (2014) 267–279

5. Conclusions and discussions

Policies regarding urban strategic planning and infrastructure investment are in general designed to be consistent with
the requirements of a sustainability framework. The issue of how to sustain the optimum of motorized mobility under the
environmental capacity constraints should be important to policy makers. Since emphasizing a single system target under
policy interventions may lose benefits from other aspects, to ability to deal with the trade-offs between multiple objectives,
provides more flexibility and improved sensitivity.
Therefore, this paper investigated the performance of an integrated model under different policy interventions, support-
ing the policy decision making with multiple concerns of motorized mobility and spatial equity. The model was designed to
specifically deal with the multi-objective optimization problem of mobility maximization and equity maximization under a
quantitatively specified environmental capacity constraint. Results from two extreme cases with single optimization objec-
tive (mobility maximization and equity maximization) show that, under the environmental capacity constraints, the total car
ownership calculated based on the mobility maximization model is approximately double of that obtained from the equity
maximization model, indicating the necessity of multi-criteria evaluation.
Simulation on different policy scenarios verified that the proposed integrated multi-objective model is capable of iden-
tifying the trade-offs between motorized mobility and accessibility-based equity for policy decision making. For each policy
scenario, the total car ownership increasingly changes with the decrease of equity. The Pareto-optimal solutions provide a
group of alternatives which can be adopted as a practical reference for planners according to their specific requirements con-
sidering environment evaluation, mobility management and equity assessment.
Evaluation on the accessibility-based equity showed that all policies achieved an improvement in terms of equity. Among
all scenarios, the urban sprawl has shown the highest contribution to increase equity. Taking the solution with the minimum
equity value as an example, the urban sprawl scenario also yielded the most significant improvement among all policy sce-
narios according to both emission and accessibility. This indicates that a policy relevant to population migration may con-
tribute to not only alleviate emissions, but also to improve zonal accessibility.
In addition to above findings, it should be noted that the current modeling framework is targeted to evaluate the max-
imum or optimal level of mobility with the environmental capacity as the constraints. The environmental capacity was set at
the zone level. One can expand it to a link level constraint. In addition, in practice, the environment indexes and/or level of
environmental pollution cannot really interact with the dynamics of mobility expansion directly. However, policy makers
can justify and trigger long-term environmentally friendly strategies by using various policy measures, according to the level
of mobile emission relative to the environmental capacity. In this regards, the proposed modeling framework would suffice
the requirement of decision making to provide not only the theoretical support, but also quantified assessment in the space
where policy makers design and approve their scenarios.
As discussed above, the proposed integrated model incorporates multiple theoretically founded model specifications,
which gives rise to the interconnection problem between different components when applied to real cases. Limited by
the quality of the personal trip survey data in this case study, there are various potential to obtain better outlook on the
effects of different policies. For instance, a more advanced modal split model representing aggregated mode choice as well
as good data would improve the significance of the results. In addition, the model can be also extended in future to address
the dynamics of land use issue within this modeling framework.

References

Bard, J.F., 1999. Practical Bilevel Optimization: Algorithms and Applications. Springer Press, USA.
Cai, H., Xie, S., 2007. Estimation of vehicular emission inventories in China from 1980 to 2005. Atmos. Environ. 41, 8963–8979.
Connors, R., Sumalee, A., Watling, D., 2005. Equitable network design. J. Eastern Asia Soc. Transport. Stud. 6, 1382–1397.
Deb, K., 2001. Multiobjective Optimization Using Evolutionary Algorithms. John Wiley & Sons, New York.
DMBS, 2001–2005. Dalian statistical yearbook. Dalian Municipal Bureau of Statistics. China Statistical Press, China.
Duthie, J., Waller, S.T., 2008. Incorporating environmental justic measures into equilibrium-based network design. Transp. Res. Record: J. Transp. Res. Board
2089, 58–65.
Feng, T., Zhang, J., 2012. Multicriteria evaluation on accessibility-based transportation equity in road network design problem. J. Adv. Transp. http://
dx.doi.org/10.1002/atr.1202.
Feng, X., Chen, S., Zhao, Q., 2003. Evaluation Technique and Method for Pollution by Road Vehicular Traffic. China Communications Press, China.
Feng, T., Zhang, J., Fujiwara, A., 2008. An integrated modeling framework for environmentally efficient car ownership and trip balance. IATSS 32, 95–108.
Feng, T., Zhang, J., Fujiwara, A., Timmermans, H.J.P., 2010. Integrated model system and policy evaluation tool for maximizing mobility under environmental
capacity constraints: a case study in Dalian City, China. Transp. Res. D 15, 263–274.
Gao, H.O., Niemeier, D.A., 2008. Using functional data analysis of diurnal ozone and NOx cycles to inform transportation emissionsnext term control. Transp.
Res. Part D 13, 221–238.
Gardner, B., Abraham, C., 2008. Psychological correlates of car use: a meta-analysis. Transp. Res. Part F 11 (4), 300–311.
Han, Y., Gong, D., Sun, X., Pan, Q., 2014. An improved NSGA-II algorithm for multiobjective lot-streaming flow shop scheduling problem. Int. J. Prod. Res. 52
(8), 2211–2231.
Karlaftis, M.G., Kepaptsoglou, K., Stathopoulos, A., 2004. Optimal location of transit repair vehicles on a large urban network: a genetic-algorithm based
approach. J. Transp. Res. Board: Transp. Res. Record 1879, 41–50.
Kenworthy, J., Laube, F., 2001. The Millennium Cities Database for Sustainable Transport, International Union of Public Transport (UITP) and Institute for
Sustainability and Technology Policy (ISTP), Brussels.
Kitamura, R., 1987. A panel analysis of household car ownership and mobility. Proc. Jpn. Soc. Civil Eng.: Infrustruct. Planning Manage. 383, 13–27.
Kumbhakar, S.C., Lovell, C.A.K., 2000. Stochastic Frontier Analysis. Cambridge University Press, Cambridge, UK.
Leblanc, L.J., 1975. An algorithm for the discrete network design problem. Transp. Sci. 9, 183–199.
Litman, T., 2002. Evaluating transportation equity. World Transp. Practice 8, 50–65.
 T. Feng, H.J.P. Timmermans / Transportation Research Part C 43 (2014) 267–279 279

Lo, H., Szeto, W.Y., 2009. Time-dependent transport network design under cost-recovery. Transp. Res. Part B 43 (1), 142–158.
Meng, Q., Yang, H., 2002. Benefit distribution and equity in road network design. Transp. Res. Part B 36 (1), 19–35.
Meurs, H., 1993. A panel data switching regression model of mobility and car ownership. Transp. Res. A 27 (6), 461–476.
Murugan, P., Kannan, S., Baskar, S., 2009. NSGA-II algorithm for multi-objective generation expansion planning problem. Electric Power Syst. Res. 79, 622–
628.
Odoki, J.B., Kerali, H.R., Santorini, F., 2001. An integrated model for quantifying accessibility-benefits in developing countries. Transp. Res. A 35, 601–623.
Santos, B., Antunes, A., Miller, E.J., 2008. Integrating equity objectives in a road network design model. Compendium of Papers CD-ROM, the 87th Annual
Meeting of the Transportation Research Board, Washington D. C., January 13–17.
Schafer, A., 1998. The global demand for motorized mobility. Transp. Res. Part A 32, 455–477.
Sheffi, Y., 1985. Urban Transportation Networks: Equilibrium Analysis with Mathematical Programming. Prentice Hall, Englewood Cliffs, NJ.
Steg, L., Geurs, K., Ras, M., 2001. The effects of motivational factors on car use – a multidisciplinary modeling approach. Transp. Res. Part A 35 (9), 789–806.
Szeto, W.Y., Lo, H., 2006. Transportation network improvement and tolling strategies: the issue of intergeneration equity. Transp. Res. Part A 40 (3), 227–
243.
Szeto, W.Y., Lo, H., 2008. Time-dependent transport network improvement and tolling strategies. Transp. Res. Part A 42 (2), 376–391.
Tobin, R.L., Friesz, T.L., 1988. Sensitivity analysis for equilibrium network flow. Transp. Sci. 22 (4), 242–250.
Vlahogianni, E.I., Karlaftis, M.G., Golias, J.C., 2005. Optimized and meta-optimized neural networks for short-term traffic flow prediction: a genetic approach.
Transp. Res. Part C 13, 211–234.
Vlahogianni, E.I., Karlaftis, M.G., Golias, J.C., 2007. Spatio-temporal short-term urban traffic volume forecasting using genetically-optimized modular
networks. Comput.-aided Civil Infrast. Eng. 22 (5), 317–325.
Yang, H., Zhang, X., 2002. Multi-class network toll design problem with social and spatial equity constraints. J. Transp. Eng. 128, 420–428.
Yu, B., Yang, Z., Yao, J., 2010. Genetic algorithm for bus frequency optimization. J. Transp. Eng. 136, 576–583.