Exploring the Performance of Different On-Demand Transit Services Provided by a Fleet of Shared Automated Vehicles: An Agent-Based Model

Automated vehicles used as public transport show a great promise of revolutionizing current transportation systems. Still, there are many questions as to how these systems should be organized and operated in cities to bring the best out of future services. In this study, an agent-based model (ABM) is developed to simulate the on-demand operations of shared automated vehicles (SAVs) in a parallel transit service (PTS) and a tailored time-varying transit service (TVTS). The proposed TVTS system can switch service schemes between a door-to-door service (DDS) and a station-to-station service (SSS) according to what is best for the service providers and the travelers. In addition, the proposed PTS system that allows DDS and SSS to operate simultaneously is simulated. To test the conceptual design of the proposed SAV system, simulation experiments are performed in a hypothetical urban area to show the potential of different SAV schemes. Simulation results suggest that SAV systems together with dynamic ridesharing can significantly reduce average waiting time, the vehicle kilometres travelled and empty SAV trips. Moreover, the proposed optimal vehicle assignment algorithm can significantly reduce the empty vehicle kilometres travelled (VKT) for the pickups for all tested SAV systems up to about 40% and improve the system capacity for transporting the passengers. Comparing the TVTS system, which has inconvenient access in peak hours, with the PTS systems, which always makes available door-to-door transport, we conclude that the latter could achieve a similar system performance as the former in terms of average waiting time, service time and system capacity.


Introduction
It is being said that we are at the dawn of the next mobility revolution with the introduction of automated driving. However, there are aspects of the automated vehicles (AVs) that still need to be understood, for example, there are many legal, regulatory and technical problems that are delaying the deployment of AVs. A eet of shared automated vehicles (SAV), which functions as a centralized taxi service system, will probably bring the most disruptive changes in urban mobility. e real potential of SAVs is that they make the implementation of an entirely new public transportation system possible. at is, SAVs might have the power to fundamentally transform transportation mobility and revolutionize the transport system given the added degrees of freedom of operating shared taxi systems [1][2][3][4][5][6][7][8][9][10].
A eet of SAVs operated in a centralized way in SAV systems could function as an e cient taxi system to provide demand-responsive service for travel demand during a day, especially in urbanized areas. e SAV system could be used to provide station-to-station (stop-to-stop) service (SSS) to transport as many people as possible in busy routes in a demand-responsive fashion. However, the SAV system could also be operated as a door-to-door service (DDS) giving great convenience to travellers as of today's Transport Network Companies such as Uber, Ly , and Didi Chuxing. In this paper, we aim to take into consideration these two ways of operating urban automated transport systems, both in parallel and in sequence, and propose a simulation tool to assess their impact on an urban network.
In addition, SAV systems could facilitate the implementation of dynamic ride-sharing which aims to pool multiple travelers with similar origins, destinations, and departure times in the same vehicle. Dynamic ridesharing has the potential to improve the performance of proposed SAV systems in terms of energy saving, waiting time reduction, VKT reduction, etc. [11][12][13][14][15][16][17]. More importantly, the dynamic ridesharing could enable the SAV system in accommodating more travel demand with the same number of vehicles. e proposed SAV systems offering various service schemes with dynamic ridesharing could eliminate the problems in past attempts to provide demand-responsive transit services [18].
Building upon the on-demand DDS and on-demand SSS, two extra on-demand transit service systems are proposed and simulated. Time-varying transit service (TVTS) that can switch service schemes between DDS and SSS depending on the time of day (peak hours and off-peak hours for example), and the simultaneous operation of DDS and SSS, allowing both of them to operate in parallel (designated as parallel transit service: PTS).
Few studies have explored the operation of variations in the service schemes of SAV systems. As a first attempt to investigate this problem, the paper constructs an agent-based model (ABM) to study different scenarios of operations of different service schemes. With the help of ABM, conceptual design and a preliminary study are presented for different SAV systems as defined above: SSS system, DDS system, TVTS system, and four PTS systems. e ABM describes the SAV system with its details and complexity by modelling the travel requests and vehicle movements, and especially interactions between vehicles and travelers. e model allows us to understand how the system components of the SAV system behave over time and find the potential of SAV systems by studying the most efficient ways of operating them under different service schemes. erefore, the preliminary look at system performance of SAV systems could provide useful information for transport operators when deciding to adopt a SAV system in the future. But it also provides support for future more detailed simulation studies whereby these schemes might be important to test. e remainder of the paper is organized as follows. Section 2 reviews the existing literature related to SAV systems and dynamic ridesharing. Section 3 presents the model specifications. Section 4 gives a detailed description of the experiments that have been run. Section 5 provides an analysis of the simulation results. Conclusions are drawn in Section 6. e final section lists some model limitations and envisions future work.

Background Literature
Given that the possible future existence of SAV will bring the most potentially disruptive changes in the urban transport system, the exploration of the implementation of these systems has been a focal point of transportation research in recent years [19][20][21][22][23].
Burns et al. examined the performance of an SAV system and cost to explore the feasibility of a fleet of SAVs to serve the existing travel demand, they found that the SAV systems are compelling due to the shorter waiting time and low operational cost [24]. Spieser et al. examined the problem of fleet sizing of Automated Mobility-on-Demand Systems using the actual transportation data of Singapore and found out that the fleet sizes of SAVs that serve the entire mobility needs between stations in Singapore can be 1/3 of previous passengers' vehicles while keeping an acceptable waiting time [25]. Fagnant et al. investigated the benefits and environmental implications of SAV systems in an Austin-sized city, Texas. eir study results indicate that each SAV can substitute 11 and 9 conventional vehicles in order to serve 3.5% and 1.3% of regional trips respectively. Although approximately 11% extra empty VKT was generated, energy savings and emission reductions may overcome those effects [26,27].
Several studies focus on investigating the potential benefits of SAV systems when considering shared rides. Fagnant et al. investigated the impact of the dynamic ridesharing in SAV systems on vehicle mile travelled (VMT), waiting times and travel costs for SAV users. ey concluded that the dynamic ridesharing could result in a reduction of generated VMT up to 4.2% and the reduction of the waiting time and average total service time up to 4.5 minutes and 0.3 minutes respectively [11]. Zhang et al. focused on the potential impact of SAV fleet size, dynamic ridesharing and clients' preference, and vehicle cruising on urban parking demand by considering 2% of the population as the users of the SAV system in a hypothetical city. eir study indicated that the SAV system could facilitate the reduction of parking demand of about 90%, and the reduction could be further expanded by 1% by considering dynamic ridesharing in the SAV system [28].
Other works concern multi-mode transportation in the analysis of the impact of an SAV system. Martinez in the International Transport Forum investigated the performance of shared automated taxis as a supplement to serve requests for shared buses offering inter-stop service with prebooking. e benefits of shared taxis and bus systems are the reduction of emissions and VKT, peaking at 40% and 30% respectively [29,30]. Zachariah et al. investigated the operation of a fleet of autonomous taxis supplementing the transit train service among fixed taxi stands in New Jersey. Simulation results revealed that shared rides could significantly reduce the VMT. In addition, they found that temporal and spatial demand variations influence the ride-sharing success rate. erefore, the favorable distribution of the SAV fleet based on the demand variations can significantly improve the sharing rate and reduce congestion [31]. SAVs as a feeder service to train stations have been explored from an operational point of view looking at how many vehicles are needed, defining an area of operation, and how to charge the vehicles in case they are electric [32,33].
Although studies on SAV systems are now booming, there are a limited number of research papers that explore the implementation of efficient on-demand SAV systems in terms of the different service schemes in which they could be operated and the potential synergies among those. at is, it is unclear what kind of service schemes the SAV systems should provide in a demand-responsive fashion. is paper attempts to fill that gap through a simulation study in a hypothetical city as a first approach to the problem. An ABM is used to explore the tradeoffs in different SAV systems between the service levels, captured by the waiting time and service time (in-vehicle travel time) and the system efficiency in terms of VKT, system capacity, and served trips.

Model Specifications and Operations
e ABM is intended to simulate the operations of SAVs and their interactions with travelers' real-time requests within a hypothetical city area. We simulate tailored on-demand SAV systems with various service schemes as already described. In this study, the eet operator has no information about the travel requests in advance. In other words, the eet operator has no information about travelers before they request service. A er a traveler requests a vehicle, the eet operator knows the information of the traveler. e eet operator only assigns the idle vehicles to serve the travelers in a real-time fashion, and therefore scheduled assignment in a prebooking fashion is not possible. As shown in Figure 1, the eet operator is responsible for real-time vehicle assignment, dynamic ridesharing, and managing and monitoring information of travel requests and vehicles. In addition, the central operator is designed for route assignments for SAVs. Vehicle assignment means that the eet operator nds idle vehicles to serve real-time travel requests. e route assignment is to nd a route either for en-route pickup vehicles or en-route drop-o vehicles. We distinguish the functions between the eet operator and the central operator, enabling the designed system to keep an expanded capability for multiple operators. e interaction of system components in Figure 1 between SAVs and time-dependent travel requests are illustrated. e eet operator controls the assignment of SAVs to serve realtime travel requests. A er the assignment of SAVs, communications will take place between travel requests and SAVs until travellers arrive at their destination. at is, a er SAVs received the essential information (origin, destination, identi cation) of travel requests, each SAV will communicate with targeted travel requests for pickups and drop-o s. e dynamic ridesharing module in the eet operator aims to group travellers, according to the matching rules. e routing module in the central operator is responsible for the route calculation for real-time vehicle routing. e central operator will transit routing information to the in-service vehicles. e model contents include dynamic generation of time-dependent requests, real-time vehicle assignment, and dynamic ridesharing. To deal with the lack of some essential information, we list the detailed description of model assumptions: when the eet operator failed in nding adequate idle vehicles as input for the Hungarian algorithm or when there is only one request for vehicle assignment in a certain dispatching time interval. A er the SAV assignment, the vehicle will have the essential information about requests (location, requested service, ridesharing status), and communicate with travellers by sending an assignment message. A er that, the traveler waits for the SAV's arrival. erefore, the waiting time can be composed of waiting time for vehicle assignment (due to the unavailability of a SAV) and waiting for the SAV' arrival while it is en-route for picking up the traveller.

Dynamic Generation of Time-Dependent Travel
Requests. Based on the aggregate travel demand, individual travel requests are generated with spatial-temporal characteristics. In this study, the demand generation process can be divided into the following two steps.
(1) Generating a xed number of time-dependent travel requests for each zone over each time interval. Total production of travel requests for each zone is calculated based on an origin-destination (OD) matrix, and then demand production per one-hour interval for each zone is estimated by using the departure time distribution and total demand production per zone for 24 hours. At the beginning of each time interval, a xed number of travel requests are generated, and then the generated travel requests are distributed within this time interval by following a discrete uniform distribution. As a result, all the generated requests for each time interval will be associated with a speci ed time.
(2) Finding a destination zone for each travel request. It is assumed that observations of travel requests in each zone over other tra c analysis zones in the whole study area are known in the OD matrix table. at is to say, the number of requests ending in every other zone is known. Based on these observations of travel requests over tra c analysis zones in the OD matrix table, the destination zone of each travel request will be drawn by using the Monte Carlo simulation process. In the end, each request will have a destination zone. We give a detailed overview of departure time distribution and total travel requests for each zone in the section of detailed travel demand.
As shown in Figure 2, statechart diagram as one of the ve Uni ed Modeling Language diagrams is used to model the dynamic nature of the travellers. e statechart diagram can de ne di erent states of a traveler during its lifetime and these states are changed by events. By using statecharts, traveler behavior can be visually shown. e statechart has states and transitions. Transitions may be triggered by user-de ned conditions (timeouts or rates, agent's arrival, messages received by the statechart, and Boolean conditions). For example, a er the SAV assignment, the vehicle will have the essential information about requests (location, requested service, ridesharing status), and communicate with the clients by sending an assignment message (state transition by receiving a message). A er that, the traveler waits for the SAV's arrival (state (iv) e parking spaces are enough for all the SAVs in each station.
For easier model implementation, we simplify the following model speci cations: (i) SAV speed is prede ned on road segments and updated for peak hours and o -peak hours respectively; (ii) Cancellation of assigned SAV is not allowed; (iii) Travelers will give up a request when the waiting time for being assigned a vehicle exceeds a speci c time threshold; (iv) Travelers' choices between door-to-door service and station-based service are based on a xed willingness to use a certain service, which is an experimental parameter (20%, 40%, 60%, and 80%).

Real-Time SAV Assignment.
In this model, two assignment methods are designed. e rst vehicle assignment method is to assign the nearest idle vehicles to serve the real-time travel requests according to the rst-come, rst-served (FCFS) principle. We de ne the rst vehicle assignment method as the FCFS vehicle assignment method. e second is an optimal assignment method that assigns a group of idle vehicles to bundled travel requests with the objective of minimizing the total empty travel distance for the pickups.

FCFS Vehicle Assignment Method.
We design a eet operator to assign the idle and nearest SAVs to serve real-time travel requests. e rules of the design are as follows: (i) e eet operator will nd an idle and nearest SAV in the same sub-region as the request departure location based on the FCFS principle; (ii) If there is no available SAV close to the request, the eet operator will nd an idle SAV from the whole study area to serve it; (iii) e eet operator only gives top priority to shared riders. at is, the travelers who will share their rides are sorted from the waiting list, and assigned an idle and nearest SAV as soon as possible.

Optimal Vehicle Assignment Method.
e optimal vehicle algorithm method can assign a group of idle vehicles at means that the eet operators can bundle a certain number of travel requests, each of which is speci ed with a timestamp, assign a group of available vehicles to them with the objective of minimizing the total empty travel distance of the assignment. e size of bundled travel requests varies along the day according to the demand that coincides in the same time interval. e collection of idle vehicles participating in the optimal assignment is found by searching for the nearest vehicles for each travel request in the set . e assignment problem can be formulated as a bipartite matching problem between bundled travel requests and selected idle vehicles in every dispatching time interval. e Hungarian algorithm [34] is used to solve the problem.
Nevertheless, a travel request can be assigned a vehicle by FCFS principle without calling the Hungarian algorithm only ridesharing agent is created, it is responsible for the interaction with an assigned vehicle. Each ridesharing agent records the information of grouped travellers, the OD of the travellers, and the assigned vehicles for grouped travellers. According to the designed rules for dynamic ridesharing, the eet operator dynamically adds and removes ridesharing agents in the simulation process.
3.5. Service Scheme. We have de ned four types of on-demand SAV systems in terms of variations of service schemes as described above: DDS system, SSS system, TVTS system, and PTS system. In all SAV systems, we did not simulate user choices for di erent services based on attributes such as price or travel distance; however, we assume that individual requests have various levels of willingness to use the station-tostation service in the proposed PTS systems. According to the willingness to choose the station-to-station service, the PTS system can be divided into PTS-20%, PTS-40%, PTS-60%, and PTS-80%. is would result from the prices of both services; otherwise, travelers would naturally prefer to use the door-todoor system only because it is more convenient.

Model Application and Implementation
e simulation model was developed from scratch in Anylogic proprietary ABM platform with Java programming language, which is available for research purposes. In this study, SAV systems with di erent service schemes are tested in a hypothetical urban road network.

Urban Road Network.
e road network of a city in the scale of 5 km × 5 km (roughly the size of Del in the Netherlands) is used for testing the operations of di erent SAV systems. e network is taken from the UDES (Urban Dynamics Educational Simulator) model (https://www. researchgate.net/project/ e-Urban-Dynamics-Educational-Simulator-UDES). e road network topology includes 78 links and 77 nodes (see Figure 3). Stations for the drop-o transition by vehicle arrival). e travel request will give up waiting for vehicle assignment when waiting assignment time exceeds a time threshold (state transition by timeout event).

Fleet Size.
e eet size is an experimental parameter in the ABM. We simulate the operations of SAV systems with di erent eet sizes. In addition, in order to illustrate the relations between multiple system characteristics, we estimate a small eet size for keeping an acceptable service quality for SAV systems.

Dynamic Ride-Sharing.
e SAV can facilitate the implementation of dynamic ride-sharing. Dynamic ridesharing aims to pool multiple travelers with similar temporal and spatial characteristics.
In this model, we design a set of rules for the implementation of the dynamic ridesharing. Travelers who have common OD zones are allowed to share a SAV. Note that the grouped travelers with common OD zone may have di erent departing and arriving speci c locations within each zone. e travel requests can be served at a service station or at their doorstep.
From the service scheme point of view, we design a set of rules for dynamic ridesharing.
(i) If both of the shared rides need to be served at a station, the assigned SAV will pick them up at the origin station, and then drop them o at the destination station. (ii) If both of the shared rides need to be served in a doorto-door fashion, the assigned SAV will rst pick up the passenger who is closer to it and then pick up another one. Based on the trip distance of the passengers, the SAV will rst drop o the passenger who has a shorter trip distance, and then it will drop o the second passenger at its speci c destination of the same zone. If the assigned SAV has the same estimated travel distance from the two passengers in two di erent locations, the SAV will rst pick up the passenger who sent the request earlier, and then pick up the second passenger at his or her doorstep. A er reaching the rst passenger's destination, the second passenger will be dropped o . (iii) If one of the shared rides needs to be served at a station and the other one is to be served at the doorstep, the SAV will rst pick up the passenger who is closer to it and then pick up the other passenger. Based on the trip distance of the passengers, the SAV will rst drop o the passenger who has a shorter trip distance, and then it will drop o the second passenger at its destination (designated station or speci c location) of the same zone. If the assigned SAV has the same estimated travel distance from the two passengers in two di erent locations, the SAV will rst pick up the passenger who sent the request earlier, and then pick up the second passenger at his or her doorstep. A er reaching the rst passenger's destination, the second passenger will be dropped o .
In this ABM, a ridesharing agent type is introduced to delegate the grouped travel requests. at is, once the the maximum number of travelers in a shared car is two. e time interval being used for the assignment is 5 seconds.

Analysis of the Impact of Vehicle Assignment Methods.
To look at how the optimal vehicle assignment method impacts the performance of di erent SAVs systems, 70 scenarios for di erent SAV systems with variations of eet size are simulated (see in Table 2). e simulation results in Figure 5 indicate that the optimal vehicle assignment algorithm can reduce empty VKT. e eet operator can optimally assign idle vehicles to serve the travelers while minimizing the total empty travel distance for the pickups.
e degree of reduction of empty VKT greatly depends on the eet size. In Figure 5(a), the optimal assignment can reduce the empty VKT for all the SAV systems in about 40%, while there is almost all of the same empty VKT for both assignment methods with a 4000-SAV eet size in Figure 5(e). and pickup service in SAV systems are uniformly distributed among the tra c analysis zones (TAZs) in the whole study area. e scale is graphically de ned in the agent simulation environment as: one pixel corresponds to ten meters. e SAVs shortest paths are computed using the Dijkstra algorithm.

Detailed Travel Demand.
e SAV systems will serve a total demand of 110 000 trips in a full day. Figure 4 depicts the departure time distribution of the demand and the total production of travel requests for each zone that are used as input in the simulation model as explained in Section 3.
To mimic the commuting patterns, OD matrices with different assumed observations are used: one in the rst half of the day and the other for the rest of the day. e destination zones are found by using the Monte Carlo simulation process. erefore, heterogeneous observations in the trip table enable the simulation to generate di erent results.
Travel demand is not only generated and attracted in the centroid of each TAZ but speci c points inside the zones are used, in order to simulate the operation of di erent service schemes. at means that travellers would walk from/to the station when using the station-based service or waiting for their pickup at their places of residence if there is a door-todoor service. Table 1 shows basic input parameters for the SAV simulation. e vehicle speed is predetermined in all SAV systems in peak hours and opeak hours respectively. Based on the research conducted by Wang et al. [35] in terms of speeds during the di erent times of the day, the reduction of the speed in peak hours range between 10% and 30%. erefore, we assume that the speed of the SAV is 20% lower than that in o -peak hours. In this ABM, we assume the SAV speed in o -peak hours is 36 km/h. e energy e ciency of di erent electrical vehicles roughly ranges from 1 kWh per 7.16 km to 1 kWh per 4.82 km (https:// pushevs.com/electric-car-range-e ciency-nedc/). erefore, for energy consumption, we adopt a rate of electricity consumption of 1 kWh per 7 kilometers that is reasonable for a two-seat, light-weight vehicle. We assume that travelers will give up requesting a SAV when the waiting time for a vehicle assignment exceeds 5 minutes. In this paper, we assume that   in Tables 3 and 4 indicate that the average peak-hour waiting time in all systems with dynamic ridesharing ranges from 8.68 minutes to 13.17 minutes when we adopt the 2000-SAV fleet size. For smaller fleet sizes, the service quality would be lower. Furthermore, there is little difference in the average waiting time in the four-PTS system when the fleet size is reduced from 3500 to 2000 as shown in Figure 8. erefore, we could analyze the SAV systems' performance starting from the estimated 2000-SAV fleet size.

Analysis of the Impact of Dynamic
Ridesharing. SAV systems allow travellers to share their rides according to the designed rules. In this analysis, we analyse the impact of dynamic ridesharing in the SAV system. Compared with a nonridesharing system in Tables 3 and 4, SAV systems with ridesharing significantly reduce at least 50% of the average waiting time, 6.0% of VKT and 4.7% of total SAV trips. e dynamic ridesharing could improve the performance of all proposed systems. e DDS system reaches a peak of approximately 34.4% of shared rides, while the SSS system has the lowest percentage of shared rides (around 15.6%). Four-PTS systems have slightly high percentages of shared rides from 18.7% to 25.9%. e shape of the polyline depicting the results of the optimal assignment displayed in Figure 5 is similar to that representing the results of the FCFS assignment. It is evident that the trend of the generated empty VKT over different SAV systems for both vehicle assignment methods is similar to each other.
at means that although the optimal assignment method can reduce the generation of empty VKT, the difference of generated empty VKT across SAV systems remains the same to some extent.
Considering the number of drop-outs (unsatisfied trips), it is possible to see the simulation results in Figures 6 and 7 for the total number of dropouts with both vehicle assignment methods and for all tested systems. Results indicate that the optimal vehicle assignment can enable the SAV systems to transport considerably more travelers. is can be explained because of a reduction in the waiting time due to the higher efficiency of the optimal vehicle assignment method.

Analysis of Fleet Size Variations.
We provide a performance analysis of the SAV system for different fleet sizes. In addition, a small fleet size for the base scenario to keep an acceptable level of service quality is determined to analyze the other characteristics of different SAV systems. Simulation results (in-vehicle travel time) in case of the 2000-SAV eet size. TVTS system has a similar performance in terms of average waiting time and service time with the PTS-20% and PTS-40% system. We can infer that the SAV systems, e.g., PTS-20% and PTS-40% system, that allows two service schemes to operate in parallel with a degree of restricted access to the door-todoor service could provide a similar system performance than the TVTS system which only o ers station-based service in peak hours. When a total eet size of 3500 SAVs is adopted ( Figure  8(d)), the average waiting time in the PTS systems with 80% willingness to request station-based services could achieve a similar value with that of TVTS system with approximately 22.8% of average service time. is means that the system performance in terms of average waiting time and average service time achieved by the sequential operational rules in the TVTS system can be obtained by the proposed parallel modes of service schemes in the PTS-80% system.

Analysis of VKT and Energy Consumption.
e DDS system has more VKT and energy consumption than other SAV systems as can be seen in Figures 9(c) and 9(d). Except for the DDS system, other proposed systems converge to the same amount both in VKT and energy consumption respectively, when eet size approaches 4000 vehicles. Both VKT and energy consumption experience is a growing trend in four PTS systems with an increase in eet size from 2000 to 2500, while the TVTS system has a high level of energy consumption and VKT. Nevertheless, with the continued growth of eet size to 4000, the TVTS system decreases the energy consumption and VKT to a relatively low level compared to the energy consumption on the PTS systems. TVTS system could operate a relatively large eet size to provide quality service while consuming less energy. Figure 9(a) showing the number of total SAV trips indicates that total SAV trips rise rst, then fall for each system with an increment of eet sizes for each SAV system. One of the possible explanations is that with the increase of the SAV eet size, fewer travel requests drop out of the SAV system. erefore, the SAV system satis es many more trips that result in the increase in the total number of SAV trips. On the other hand, the gradually increased eet size will potentially reduce the empty SAV trips for pickup. e decline of empty (unoccupied) SAV trips for en-route pickups appears to reduce the total SAV trips. As a result, the total SAV trips rise rst and decline for each SAV system. e peak number of total SAV trips is about 131342 trips in the TVTS system, while the DDS, PTS-20% and PTS-40% systems only reach about 118000 trips with the 2000-SAV eet size. Results in Figure 9(b) indicate that the empty trips with a 2000-SAV eet size for each SAV system occupy 30-40% of the total trips served. e percentage of empty trips in the SSS system has a minimum of 28.1% of the total served 97973 trips with 2000-SAV eet size, while the DDS reaches a peak of 42.0% with a total of 83480 trips. e percentage of extra empty trips in the TVTS system is the second-largest percentage (40.6%). With a total eet size of 2000 SAVs, SAV systems seem to generate a higher percentage of empty trips. e high Especially, the TVTS system is about the same as the PTS-60% for the percentage of shared rides with a 2000-SAV eet size, reaching 20.1% of total serviced trips. e PTS systems and TVTS system, providing two service schemes, can achieve a relatively high sharing rate of trips. Although the simulation results for dynamic ridesharing may not give conclusive evidence under designed matching rules to group travelers, the preliminary investigations of the impact of dynamic ridesharing on di erent SAV systems provide useful insights into the deployment of di erent SAV systems. Figure 8  smaller eet size can roughly keep 75% of the travelers waiting 10 minutes or less. With the shi s of SAV eet size to 3500, DDS system still has a peak of 14.1% requests whose waiting time is larger than 10 minutes. Results indicate that PTS-20% still maintains a high percentage of travelers whose waiting time exceeds 10 minutes with a 3000-SAV eet size (21%). erefore, we can infer that the PTS system with a low willingness to choose the station service will lead to a long wait.

Analysis of Waiting Time and Service Time. Simulation results in
Simulation results in Figure 7 indicate that the number of travelers who give up waiting for a SAV assignment has a signi cant descending trend with the increase of SAV ee sizes in seven SAV systems. Except for the system with DDS, a eet of 3500 SAVs can accommodate almost all of the requests. e eet size that accommodates the total 110000 requests in the DDS is approximately 4500 SAVs. erefore, e SAV system only with door-to-door service needs many more SAVs to handle the high demand. A large number of vehicles in the eet, has the potential to reduce vehicle utilization. In fact, the simulation results in Figure 11 reveals that the DDS system has the lowest number of the served trip per SAV in all scenarios.
In addition, the numbers of served trips per SAV in Figure 11 are from 41.7 trips to approximately 49.0 trips in the SAV system with a 2000-SAV eet size. We nd out that percentage of empty vehicle trips in DDS and TVTS has the potential to cause heavy tra c congestion.

Analysis of System Capacity and Drop-Out Requests.
With the 2000-SAV eet (Tables 3 and 4), the peak number of drop-outs is 24554 trips corresponding to 22.3% of the total number of requests (110000) in the DDS system, while in the SSS system this number goes down to 12027 drop-outs, only accounting for 10.9% of the 110000 requests. e dropout rate in TVTS system approximates that of PTS-60% system with a 2000-SAV eet size, reaching 15% of the total number of requests (110000). e PTS-80% system has the lowest number of dropouts. It is evident that the PTS system with a relatively high percentage of willingness to choose stationbased service would be able to accomplish the performance of the TVTS system. Results in Figure 10 indicate that the number of trips whose waiting time exceeds 10 minutes is between 29493 trips and 16624 trips, going down from 35.3% to 16.9% of system capacity (total number of served trips) with a 2000-SAV eet size. e percentage of trips whose waiting time exceeds 10 minutes is about 25% in both PTS-40% and TVTS system, which are slightly larger than that of PTS-60% and PTS-80%. Both the TVTS system and PTS system with a relatively 1715-SAV eet size within a network in the scale of 12 miles × 24 miles. at is, each SAV can approximately serve 32.8 trips. e served trips per SAV are relatively lower than ours. One reason is that the road network in Fagnant and Kockelman's study is relatively larger than that of this study. Another reason is that a relatively large number of vehicles the PTS-60% and PTS-80% systems present about the same number of served trips as the TVTS system at about 47 trips per SAV. We compared the number of served trips per SAV with Fagnant and Kockelman (2016)'s study. Fagnant and Kockelman (2016)'s study indicates that the SAV system considering ridesharing can serve 56324 person-trips with

Analysis of Empty Trips.
Unlike human-driven vehicles parking at the destinations, SAVs could have an unoccupied journey to pick up the next request (no pro-active rebalancing in anticipation of future demand are considered in this study). erefore, additional empty trips will be generated to satisfy the next trip. In this study, the additional empty trips by the vehicle movement between di erent zone stations are calculated. ese empty trips have the potential to in uence tra c congestion to a great extent. erefore, it is of importance to know the number of empty trips by SAVs. As shown in Figure 12, dynamic ridesharing can signi cantly reduce the generation of empty trips. PTS systems have   11: e number of served trips per SAV for di erent SAV systems.

Conclusions
is paper developed an agent-based simulation model to assess the potential of on-demand SAV systems with various service schemes. With the help of the developed ABM, we understand what the performance of SAV systems with di erent service schemes is, and how the associated factors (variation of eet size, dynamic ridesharing, di erent vehicle assignment methods) in uence the service quality of SAV systems. Our study shows that the promotion of ridesharing can signi cantly improve the performance of the proposed SAV systems in terms of reducing the average waiting time, VKT and empty trips. Moreover, compared to the FCFS vehicle assignment method, the optimal assignment can reduce the generation of empty VKT for all tested systems and enable the SAV systems to transport considerably more travellers.
Although the DDS system brings great convenience of doorstep service for real-time requests; it is evident that DDS generates almost 13% of extra VKT than that of the PTS system with a eet size of 2000 SAVs. In addition, the DDS system generates approximately 42% additional empty trips. e percentage of dropout requests takes up 22.0% of the total 110000 person-trips. at is, the DDS system cannot transport as many more travelers as the other SAV systems do. Compared to the DDS system, the TVTS system and PTS systems can reduce at least 14.6% and 14.8% of the average waiting time respectively. e empty trips in the TVTS and PTS systems with dynamic ridesharing account for 41.0% and 33.3% of total served trips respectively. e TVTS and PTS system provides a signi cant gain in terms of transport capacity, waiting time and additional trips by empty SAVs. In other words, the SAV systems that include two di erent on-demand services have the most signi cant improvements in system performance.
DDS system ranks the highest in total energy consumption and VKT. Compared to the VKT in the DDS system, the TVTS system and PTS system can reduce at least 7.6% and 14.0% of the VKT with 2000-SAV eet size. On the other hand, the DDD system transports a relatively small amount of travel requests and reduces vehicle utilization that is the average number of served trips per day per vehicle. Based on the analysis of the proposed SAV systems, TVTS and PTS systems are a promising alternative to be implemented to satisfy the intracity transportation needs. In both systems, a SAV can serve many more trips per day with relatively less waiting time. e PTS systems with a relatively high percentage of choosing station-to-station service show a high level of service that could transport many more requests with less waiting time and empty trips. Although the TVTS system could generate many more VKT and consume much more energy, this system still has a relatively small waiting time and fewer dropouts with providing doorstep convenience. In the future deployment of SAV systems, the station-based service combined with the door-to-door service parallely in time and space, is of importance, since blended service could make the system operate at a relatively high degree of service quality without the inconvenient access. the greatest reduction, reaching a peak of 23% in PTS-60% system; however, there is a large number of additional empty (unoccupied) trip in all SAV systems. DDS and TVTS system with dynamic ridesharing generate many more empty trips at around 40.5% of total served trips. e PTS systems with dynamic ridesharing generate relatively fewer empty trips than that of the TVTS system.
Simulation results in Figure 13 indicate the generation of empty trips with dynamic ridesharing is sensitive to the eet size. As the eet size increases, the percentage of empty trips experiences a downward trend. e percentage of empty in all SAV systems drops below 5% when the eet size is 4500. In addition, it depicts that the SSS system, TVTS system, and PTS-80% system have low numbers of empty trips by SAV, compared with other systems.

Model Limitations and Future Work
In this model, we did not take into account realistic traffic dynamics. A traffic flow model can be implemented in the model framework to capture the traffic dynamics. Moreover, we did not design the optimal ridesharing rules for travelers and limited the maximum number of grouped travelers to two. It is acceptable to use ultra-compact vehicles with designed two seats. ese low capacity vehicles could keep travelers in privacy and comfort. We will study the formation of vehicle platooning between vehicles with low seat capacity in future research. e small, ultra-compact vehicles could operate together in a platooning fashion to improve traffic capacity and eventually save energy consumption. e fact that we use a synthetic network can introduce some limitations in the study however we also believe that by having created realistic trip requests and realistic vehicle movements the small network allows to compare well the different scenarios for the assumptions that were taken. is is a first study of exploring many different possibilities of operating the system and in the continuation of this research, we will expand the size of the network to be analysed.
Data Availability e data of travel demand and road network used to support the findings of this study are available from the corresponding author upon request.

Conflicts of Interest
e authors declare that there is no conflicts of interest regarding the publication of this paper.