A Scheme to Improve Network Performance Based on Traffic Restriction and Pricing in the Presence of Carpooling

. For the purpose of optimizing the trafc network in the presence of carpooling, a trafc demand management scheme was proposed with the consideration of trafc restriction, carpooling, and road pricing. We establish a variational inequality describing the travelers’ mode and route choice following the logit mode split and multimodal user equilibrium principle in the condition of elastic demand. A bilevel programming was designed with the variational inequality as the lower level and optimization objective function taking into account trafc congestion and social welfare as the upper level. Te results show that this scheme is possible to be efective in improving network performance. Te higher average occupancy of carpooling vehicles and the carpooling demand may not bring better network performance due to the topology, parameters of the network, and travel demand. Te fare of carpooling travelers is possible to impact the network very slightly with this scheme. Besides, the group of travelers who is able to aford more than one vehicle is hardly afected by this scheme.


Introduction
In terms of the increasingly serious situation of trafc congestion, carpooling, which is defned as two or more persons sharing one vehicle on the same trip, has gradually been focused on due to the superiorities of reduction of the total number of vehicles, releasing the travel demand, and alleviating trafc congestion. Some previous studies show that it has positive impacts on mitigating trafc congestion [1,2], while another issue is that some travelers are reluctant to carpool since they care about the privacy and security [3]. Vanoutrive et al. issues that compulsory measures are able to enhance carpooling more than mild measures [4]. Terefore, compulsory measures are better choices for us to encourage carpooling and thereby alleviate congestion and elevate social welfare.
As a widely-used trafc management policy, road pricing is implemented in several places all over the world and followed by a large body of studies in theory and practice for the purpose of analyzing the impacts on the trafc system and exploring optimization [5][6][7][8]. For instance, Yang and Huang proceed with systematic research to reveal the mathematical and economical elements in the road pricing feld [9]. Zhang and Yang propose a model and algorithm to determine the pricing district and the toll level in the case of cordon-based congestion pricing [10]. Te studies show that road pricing is an efective measure to alleviate trafc congestion and improve system efciency by adjusting travel demands and travel fow patterns. Simultaneously, trafc restriction can be seen in worldwide cities such as Milan, Beijing, San Paulo, Santiago, etc., and brings positive feedback in mitigating trafc congestion. From the perspective of theoretical research, some studies detect the impacts of given trafc restriction schemes on the networks [11,12], while others design the optimal restriction district and the proportion [13,14]. However, there exist debates with respect to road pricing and trafc restriction. For example, both road pricing and trafc restriction lead to objections from the public since, by some populations, the roads are considered public resources, so the two managements are violations of people's rights. Terefore, in many cities, the policies are difcult to implement. Besides, trafc restriction is valid in managing congestion with the price of restraining the travel demand. Considering the value of the time, Nie establishes a model to describe the user equilibrium and system optimal in general road networks when implementing frst-best pricing and trafc restriction, respectively, and analyzes the impacts on a network. Te fndings show that trafc restriction is almost impossible to achieve the results of road pricing, even not benefcial to the trafc system when using trafc restrictions since travelers are able to buy extra cars to avoid the trafc restriction, which is usually license plate rationing. Terefore, trafc restriction is supposed to be implemented with other policies matching up to optimize the system [15]. Nie explores the efects of the schemes taking tradable permits, vehicle quota, and tradable credits as the remedies for trafc restriction, respectively, and issues that tradable credit remedy brings the best efects [16]. Terefore, taking into account the superiorities and the drawbacks of road pricing and trafc restriction, the remedy policies are feasible choices.
Song et al. propose a Pareto-improving optimal scheme combined with road pricing and trafc restriction, which stipulates that the travelers who are restricted could pay to enter the restriction districts and provide the model and algorithm to capture the optimal restriction area, proportion, and payment. Te result shows that it is capable of obtaining better efects with this scheme compared to the trafc restriction scheme and the existing Pareto-improving scheme [17]. Nevertheless, in this scheme, Song et al. have not considered the impacts of the case that some travelers are able to aford extra cars to avoid the trafc restriction. Neither has they set the restriction as a connected district, which brings inconvenience for implementing the scheme. In addition, carpooling is not supposed to be neglected, accounting for the gradual popularity of this travel mode.
Triggered by this scheme, this paper proposes a cordon-based trafc restriction scheme in the unrestricted solo-driving vehicles and carpooling vehicles are permitted to enter the connected district, and the restricted solo-driving travelers could either pay to enter the restriction area or bypass the area. Terefore, restricted solo-driving travelers could choose to carpool to save the travel cost without the willingness to pay. Up to now, the studies concerning carpooling can be divided into two streams: simulation and mathematical modeling. In the feld of simulation, Balac et al. use MATSim to estimate the carsharing demand in Zurich and investigate the impacts of parking pricing on carsharing [18,19]. Ciari et al. use simulation technology to detect the impacts of several road pricing schemes on carsharing in Zurich [20]. Trough a complementarity model, Xu et al. made the frst attempt to describe the ridesharing network with ungiven occupancy of the ridesharing vehicles and analyze the solution of the model [21]. Li et al. investigate the impacts of OD-based surging pricing on the ridesharing market [22]. Li et al. propose a convex programming approach to describe the ridesharing equilibrium network and provide the algorithm [23].
Few works of literature discuss the comprehensive impacts of carpooling, road pricing, and trafc restriction. Tis paper takes the case that some travelers are able to aford multiple vehicles and elastic demand into consideration and use a multimode model to describe the trafc system. In addition, bilevel programming is adopted to determine the restriction district, the proportion, and the toll level. Te rest of this paper is organized as follows: Section 2 introduces the preliminaries of this paper. Section 3 introduces the basic multimode model of trafc restriction. Section 4 establishes the carpooling system equilibrium taking into account road pricing and trafc restriction. Te optimal scheme and the algorithm are designed in Sections 5 and 6, respectively, and applied to numerical examples in Section 7. Section 8 concludes this paper.

Preliminaries
Considering a general network G(Q, A), Q is the set of nodes, and A is the set of links. W is the set of OD pairs, M is the set of travel modes, and a ∈ A, w ∈ W, m ∈ M, W r is the set of the OD pairs which have been encompassed in the restriction district. Te set of path is L and l ∈ L. f m w,l represents the path fow with mode m on path l between OD pair w. d m w denotes the travel demand with mode m between OD pair w and d w denotes the travel demand between OD pair w. v a is the link fow on link a.

Travel Mode Split and Route Choice.
In this paper, we use the logit model to split the modes of travelers. Each traveler determines the travel mode based on the utility, namely, the travel cost. Te population of the two modes are calculated as follows: where θ is the sensitive parameter and μ m w is the minimal cost of mode m between OD pair w. We use user equilibrium (UE) to describe the route choices of the travelers, which can be expressed as follows: c m w is the travel cost of mode m between OD pair w. It is demonstrated that all travelers choose paths involving minimal travel costs, and no one is able to decrease his/her cost by choosing another path. Ten, the network falls into the user equilibrium. In this paper, we consider elastic demand in more realistic circumstances. Te actual demand is computed as follows: where η w is the sensitive coefcient, � d w is the potential demand, and μ w is the minimal travel cost between the OD pair w. Note that μ w is the minimal cost among all travel modes. Tus, it is calculated by the following:

Multimodal Combined Variational Inequality with Trafc
Restriction. Te multimode travel network with trafc restriction can be described by the variational inequality (VI) as follows: where c is the trafc restriction proportion. For example, if c � 0.3, it means 30% of the travelers are forbidden to enter the restriction district. c m,u w,l and c m,r w,l denote the travel cost of the unrestricted and restricted travelers with travel mode m on path l involving minimal cost between OD pair w on the equilibrium, respectively. L N and L R are the set of the paths when the restricted travelers have or have no path to bypass the district, respectively. W r /W represents the OD pairs in which travelers have no path to bypass. Equations (5) and (6) demonstrate the fow conservation of the unrestricted and restricted travelers with the around paths. Equation (7) demonstrates that restricted travelers choose alternate modes without any around the path. f m,r w,l represents the fow of alternatives.
Te VI above follows the logit mode split and user equilibrium, and the Gauss-Seidel decomposition process can be applied to solve it to capture the travel demand of each mode and the fow patterns.

Proposition 1. Te trafc restriction on total demand can be decentralized into the demands of various travel modes in the multimode trafc network.
Proof. From equation (8), we obtain the relationshipcd w � c m∈M d m w � m∈M cd m w by multiple the proportion c simultaneously, which means the description of the trafc restriction on total demand in mathematical modeling can be replaced by that on the various demand of mode m.
By Proposition 1, we know that the impacts of trafc restriction on total demand d w is the same as that on the various demand of mode m. Tis implies that when trafc restriction is implemented to a multimode network, it can be decentralized into various demand travel modes. Ten, we are able to capture demand the fow patterns by solving the subproblem of VI with respect to each mode. Furthermore, the subproblems of VI are routine VI problems, which can readily be solved through an existing approach such as the F-W algorithm, and projection algorithm. Tis makes the procedure ore convenient.

Mode Choice and User Equilibrium in the Presence of Traffic Restrictions and Road Pricing
A handful of studies have been devoted to trafc restrictions, road pricing schemes, alternate travel modes, and multiple car owners [15,17,24]. However, few studies consider those simultaneously. In a trafc system, if travelers fall into trafc restriction, they can choose an alternate mode, pay the price, or own more vehicles to avoid the trafc restriction, which infuences the mode and route choices as well as the network performances. In this paper, we assume there exist solo-driving and carpooling since carpooling is advocated due to its superiorities such as time and energy saving. When implementing trafc restrictions, carpooling travelers and unrestricted solo-driving travelers are allowed to enter the restriction area. Te restricted solo-driving travelers with alternate paths can bypass the restriction district, while those without alternate paths can avoid the trafc restriction by paying the price or owning multiple cars. Te multicar owners are able to use the second car to avoid trafc restrictions. Terefore, we assume that these travelers would not choose carpooling since owning multiple cars plays the same role as carpooling. Te relationship between the travel demands is illustrated in Figure 1. For simplicity, this paper has the assumptions below: (1) Tere exist two travel modes in the network: solodriving and carpooling. And solo-driving travelers have the choice to own A car or multiple cars. Namely, the travel modes include solo-driving with A car denoted as s1, solo-driving with multiple cars denoted as s2, and the carpooling h. (2) Tere is only one platform providing the carpooling service, which implies that the carpooling travelers in the same vehicle pay the same fare.

Mathematical Problems in Engineering
Ten, the travel cost functions of the various type of travelers are expressed as follows: W r represents the OD pairs included in or separated by the restriction district, and in which there is at least one path around the restriction district, while W r /W are the OD pairs included in or separated by the restriction district, and in which there is no path around the restriction district. It implies that the solodriving travelers who do not prefer turning the travel mode have to pay to travel between the OD pairs W r /W, and the solodriving travelers between the OD pairs W r are able to bypass the district. c s1rτ w,l , c s1u w,l , c s2 w,l , and c h w,l denote the cost function of the solo-driving travelers restricted with A car and paying the price, the solo-driving travelers unrestricted with A car, solo-driving travelers with multiple cars, and the carpooling travelers. τ a is the price that the restricted solo-driving travelers pay to use link a in the restriction district. ϕdenotes the amortized cost of owning multiple cars. o is the average occupancy of one car, and σ is the toll that the carpooling travelers pay to the carpooling platform. Δ(o) represents the nonfnancial additional cost, including matching and waiting time of carpooling riders. Yang and Huang assumed the additional cost consisting of the extra cost for the collection and distribution of riders and any other undesirable features as a constant [25]. We consider the cost Δ(o) as the function of carpooling occupancy. It is demonstrated that the average car occupancy is 1.6 in the German district [26]. Terefore, we set the average occupancy of carpooling vehicles as a constant varying from 2 to 5 normally. δ l a is a binary variable, which equals 1 if link a is on path l, and 0 otherwise. L τ represents the set of the paths on which the travelers restricted pay to use the restriction district.
Te solo-driving travelers restricted without alternate paths follow the logit model for the mode split, namely, pay the price and carpooling. Te two modes are expressed as s1rτ and s1rh, respectively. Te demands can be computed as follows: where d w � s1r is the population of solo-driving travelers restricted with A car between OD pair w.
Te solo-driving travelers restricted with A car and without alternate paths choose to pay or carpooling are denoted by s1rτ and s1rh, respectively. Note that the minimal cost of the alternate mode for solo-driving travelers restricted is the minimal cost of carpooling. In other words, μ s1rh w � μ h w and c s1rh w � c h w . Terefore, between the OD pair w in which the solo-driving travelers are restricted have alternate paths, the minimal cost of the solo-driving travelers with A car can be computed by the following: where μ s1u w is the minimal cost of solo-driving travelers unrestricted and μ s1r w is the minimal cost of solo-driving travelers unrestricted having alternate paths. Between the OD pair w in which the solo-driving travelers are restricted and lack of alternate paths, the minimal travel cost of the solo-driving travelers with A car is calculated by the following: � μ s1r w is the minimal cost of solo-driving travelers unrestricted with A car and lacking alternate paths, which can be obtained by the following:  Based on Proposition 2, it is shown that the VI above demonstrates the logit mode split and user equilibrium. In other words, the solution of the VI is the demand and the fow patterns. Ten, we will explore the existence and the uniqueness of the solution. Te solutions might not be unique since we cannot prove the uniqueness of the solution. However, the nonunique solutions are path-based fow patterns. Various solutions probably generate the unique link-based fow pattern and one network performance since they are in the feasible set. Terefore, in terms of the various solutions, we can still calculate the network performances.

Design of the Optimal Scheme
Obviously, elastic demand is more realistic. In the case of elastic demand, on the one hand, trafc restriction curtains the total number of vehicles on the network. On the other hand, the alleviation of congestion deriving from the decrease in travelers might induce more trips. Terefore, there is supposed to be an optimal scheme with the optimal trafc restriction district, proportion, and pricing. Considering a model to generate the rationing scheme, we are able to solve the VI to obtain the demand and fow patterns under the given scheme. If we can receive a comparative optimal scheme through the demands and fows, the optimal scheme is the fnal target. Terefore, bilevel programming is adopted for the design of the optimal scheme, in which the VI is the lower level, and we set an objective as the upper level. In this framework, the demand and fow patterns are captured from the lower level to receive the network performance as well as the optimal district, proportion, and price from the upper level built on the network performance. Since the network performances can be measured by the congestion and social welfare, we use the function of maximizing network congestion and social welfare as the objective of the upper level [13].
Te objective above implies taking the network congestion and social welfare into consideration simultaneously when we try to optimize the system, and the proportion and the price level are supposed to fall into the reasonable range. Te frst part of objective Z is the total overload fow, while the second part is the social welfare. M is a sufciently large and positive coefcient connecting the two parts.

Algorithm
Bilevel programming is a mixed integer programming problem. Taking the convenience of implementation into consideration, we assume that the restriction district is connected. Some studies have explored a similar problem, such as how to optimize the connected area and the relevant variables. For instance, Zhang and Yang proposed an approach to design the cordon-based congestion pricing scheme to capture the optimal connected area and the congestion pricing level [10]. Shi et al. adopted the genetic algorithm to solve the optimal trafc restriction problem to obtain the optimal restriction district and the proportion [13]. Chen et al. used the surrogate assistant (SA) model to solve the bilevel congestion pricing problem and compared the efects of various SA models via a real case. It is known that the surrogate assistant model can help us reduce the expenses of the problems. Tus, in this paper, we propose an algorithm combined with a genetic algorithm and surrogate assistant model to capture the optimal restriction district, the proportion, and the linkbased toll pattern. We train the SA model with the evaluation function and the variables, then apply the SA model to the genetic algorithm. Te specifc procedures are as follows: Step 0. Generate the various initial solutions involving the trafc restriction districts, the proportions, and the toll levels.
Step 1. Construct the surrogate assistant function with the districts, the proportions, and the toll levels as the variable and the objectives as the results. In this paper, we use the quadratic polynomial function as the surrogate assistant model, which is written as follows: 1≤ i y ≤j y ≤ n ζ n−1+i y +i y x i y x j y . (16) Where ξ i y is the parameters needed to be estimated, and n is the number of the variables.
Step 2. Use a genetic algorithm to solve the bilevel programming. Note that (i): Te trafc restriction district is connected. In the code of the solution process, we use the binary variable ρ q in the gene to denotes the determination of the restriction district. ρ q � 1 means the node q is included in the restriction district and ρ q � 1 otherwise. Te proportion and the toll are also coded and connected to the gene consisting of ρ q . (ii): Te evaluation function in the process of the genetic algorithm is replaced by the surrogate assistant function built on the specifc results of the district, the proportion, and the toll 6 Mathematical Problems in Engineering pattern. Te parameters of the SA model can be calibrated through the results.

Numerical Examples
Example 1. We take the network in [12] as the example shown in Figure 2. Te free-fow travel time t 0 a and the capacity C a are listed in Table 1. Tere are four OD pairs (1, 2), (1, 3), (4, 2), and (4, 3), and the potential travel demands are 1200, 900, 900, and 1200, respectively. Te other parameters are as follows: ϕ � 15, λ � 2, β � 0.6, η w � 0.04, θ � 0.5, M � 50, τ a,min � 0, τ a,max � 10. We adopt the BPR function to measure the link travel time, which is expressed as the following: where t 0 a is the free-fow travel time on link a. α 1 , α 2 are the parameters. Normally, we let α 1 � 0.15, α 2 � 4. Te link fow v a and the path fow f m w,l satisfes the following: v a � w m l f m w,l · δ l a , w ∈ W, m ∈ M, l ∈ L. Figure 3 illustrates the change of the objective against various average carpooling vehicle occupancy. It can be seen that the tendency of the objective change fuctuates   7  10  1000  10  10  1000  13  20  1000  16  10  1000  2  10  1000  5  10  1000  8  10  1000  11  10  1000  14  10  1000  17  10  1000  3  10  1000  6  10  1000  9  10  1000  12  10  1000  15  10  1000  18 30 1000   Mathematical Problems in Engineering instead of being linear with the occupancy. Furthermore, in terms of some occupancies, the objective value with the scheme is lower than that without the scheme. Figure 4 illustrates the changes in the social welfare and the overloaded fow with the occupancy are nonlinear neither. Tis is because the increasing average occupancy value means less vehicles, which releases congestion and thereby brings more trips. After the mode split and the trafc assignment, we have the results in Figures 3 and 4. Terefore, this scheme is benefcial to the network but not always accounting for the topology, the parameter of the network, and the demand patterns. Note that in Figure 4, when the overloaded fow is zero, the social welfare is on the top when the occupancy is 4.4, which means the network is not congested and the population travel most. It proves the rationality of the objective function. Namely, in some situations, this scheme is able to capture the maximal social welfare and uncongested roads. Figure 5 depicts the change of the objective with the various payments of the carpooling travelers to the carpooling platform, and Figure 6 portrays the change the social welfare and the overloaded fow against the payment. Figure 5 illustrates that with the increase of the payment, the network performances with the scheme and without the scheme change alike so that the gap of the objectives between the cases with and without the scheme changes slightly. Te reason is that the change of payment alters the travel demand and the congestion state of the network. Te network performance maintains approximately due to the mode split and trafc assignment. Tus,      No  17251  13246  6  3278  501  472017  446967  1  19136  10163  11  7816  139  465637  458687  2  21308  11275  9  9620  475  462314  438564  3  18863  13361  8  5753  482  471917  447817   8 Mathematical Problems in Engineering payment is not the key to improving the network performance under this scheme. Te decision-makers could consider other factors for system optimization, e.g., the proft of the platform. Figure 6 shows that with the increase of the payment, the social welfare and the overloaded fow decrease with the scheme, implying that higher payment is not benefcial for carpooling encouragement.
In reality, plate-number-based odd-even rationing and connected district are considered for the convenience of the policy implementation. In addition, some places must be included in the restriction district due to some reasons, e.g., special events happening in the places. Terefore, we set three scenarios: (i) the theoretical optimal scheme, (ii) the optimal scheme including node 9, and (iii) the proportion is 50%, namely, plate-number-based odd-even rationing. We still use BPR function for computing the link travel time. In Figure 7, the three scenarios illustrate the network with the optimal schemes, respectively, in which the restricted districts consist of the black links. From Table 2, we know that the carpooling demand is proportional to the total demand, which means the travel demand can be released by carpooling. However, the carpooling demand is negatively relevant to social welfare. It might derive from the sensitive parameter η w . Tus, carpooling is not always benefcial for social welfare as the topology of the network and other policies have impacts on that. From Table 2, it can be seen that the relationship between the carpooling demand and the overloaded fow is not linear. When the carpooling demand is on the top, the overloaded fow is not minimal. Tis is because the various schemes bring diferent actual demands, which lead to the various travel mode demand and the assignment. In some OD pairs, the demands increase, and  the paths become more congested. It implies that the carpooling demand cannot determine the network congestion. Furthermore, the demand for travelers owning more vehicles varies slightly, implying that those travelers are hardly infuenced by the schemes. Table 3 shows the toll level with the theoretical optimal scheme, and the restriction proportion is 79%.

Conclusions and Future Directions
Road pricing and trafc restriction have superiorities and drawbacks for encouraging carpooling. In this paper, we propose a scheme that allows carpooling travelers and solodriving travelers willing to pay the price to enter the trafc restriction district. In terms of this scheme, we describe the behaviors of travelers, design an optimal scheme, and explore the impacts of this scheme on the network. To be specifc, frstly, we use a VI to describe the travelers' mode and route choice following the logit mode split and UE principle. Ten taking the VI as the lower level, we use bilevel programming with the objective combined with social welfare and overloaded fows as the upper level. At last, we propose an algorithm consisting of a genetic algorithm and surrogate assistant model to determine the optimal scheme involving optimal connected restriction district, proportion, and the pricing, and apply the model and the algorithm to a simple network example as well as a sioux falls example with three scenarios: theoretical scenario, the scenario including a special node, and the scenario with plate-number-based odd-even rationing. From the results, we know that this scheme can be applied to reality. In our examples, this scheme is benefcial to the network performance, but not always. It implies this scheme could be an alternative for decision-makers with the purpose of improving the trafc system. On account of the topology of the network, the parameters, and the travel demand pattern, the higher average occupancy and carpooling demand are not always benefcial to the network with elastic demand case. Terefore, it is necessary to consider the individual case when encouraging carpooling to improve the network performance. In addition, the impacts of trafc policy on travelers owning multiple cars might be slight since the number of travelers are not too much. Te decision-makers may decrease the weight of the travelers to some extent when making trafc policies.
Tis study ofers a model and an algorithm to determine a trafc policy. Nevertheless, this model assumes the fxed carpooling average occupancy. For realistic circumstances, unfxed and nonaverage occupancy is supposed to be considered in this scheme. Besides, more travel modes, e.g., transit, need to be involved in the study in future directions.