Dynamic Interaction between Shared Autonomous Vehicles and Public Transit: A Competitive Perspective

by   Baichuan Mo, et al.

The emergence of autonomous vehicles (AVs) is anticipated to influence the public transportation (PT) system. Many possible relationships between AV and PT are proposed depending on the policy and institution, where competition and cooperation are two main categories. This paper focuses on the former in a hypothetical scenario-"if both AV and PT operators were only profit-oriented." We aim to quantitatively evaluate the system performance (e.g. level of service, operators' financial viability, transport efficiency) when AV and PT are profit-oriented competitors with dynamic adjustable supply strategies under certain policy constraints. We assume AV can adjust the fleetsize and PT can adjust the headway. Service fare and bus routes are fixed. The competition process is analyzed through an agent-based simulation platform, which incorporates a proposed heuristic dynamic supply updating algorithm (HDSUA). The first-mile scenario in Singapore Tampines area is selected as the case study, where only bus is considered for PT system. We found that when AV and bus operators are given the flexibility to adjust supply, both of them will re-distribute their supply spatially and temporally, leading to higher profits. In temporal dimension, both AV and bus will concentrate their supplies in morning and evening peak hours, and reduce the supplies in off-peak hours. The competition between AV and PT decreases passengers' travel time but increase their travel cost. The generalized travel cost is still reduced when counting the value of time. The bus supply adjustment can increase the bus average load and reduce total passenger car equivalent (PCE), which is good for transport efficiency and sustainability. But the AV supply adjustment shows the opposite effect. Overall, the competition does not necessarily bring out loss-gain results. A win-win outcome is also possible under certain policy interventions.



page 25


Identifying the Factors that Influence Urban Public Transit Demand

The rise in urbanization throughout the United States (US) in recent yea...

Integrated Agent-based Microsimulation Framework for Examining Impacts of Mobility-oriented Policies

Travel demand management measures/policies are important to sustain posi...

Network Topology and Time Criticality Effects in the Modularised Fleet Mix Problem

In this paper, we explore the interplay between network topology and tim...

Competition and Cooperation of Autonomous Ridepooling Services: Game-Based Simulation of a Broker Concept

Autonomous mobility on demand services have the potential to disrupt the...

Real-time and Large-scale Fleet Allocation of Autonomous Taxis: A Case Study in New York Manhattan Island

Nowadays, autonomous taxis become a highly promising transportation mode...

Virtual Immersive Reality for Stated Preference Travel Behaviour Experiments: A Case Study of Autonomous Vehicles on Urban Roads

Stated preference experiments have been known to suffer from the lack of...

Estimation of Time Delay caused by Point Geometry in Public Transport

Travel time prediction is a well-renowned topic of research. It is prima...
This week in AI

Get the week's most popular data science and artificial intelligence research sent straight to your inbox every Saturday.

1 Introduction

1.1 Background

The emergence of autonomous vehicles (AVs) as a new transportation mode is anticipated to extensively influence the future urban transportation system in different perspectives, including the traffic flow (Talebpour and Mahmassani, 2016), network congestions (Fagnant and Kockelman, 2015), land use (Koh and Wich, 2012) and road safety (Zhang et al., 2016). As an critical component of urban transportation, the public transportation (PT) system cannot avoid the shock of from the AV waves. Given the large uncertainty of future AV application, many possible imaginations of the interactive relationship between AV and PT has been proposed in the academic community (Lazarus et al., 2018). Some researchers are negative to the AV with respect to PT, arguing that AV will be the competitor to traditional PT systems (Levin and Boyles, 2015; Chen and Kockelman, 2016) or even replace the PT systems (Mendes et al., 2017). Others are optimistic to the AV and PT integration, stating they could be complementary in many scenarios (Lu et al., 2017). Various new ”AV-PT” integrated systems are designed and evaluated based on this point of views (Liang et al., 2016; Wen et al., 2018; Salazar et al., 2018).

The interplay between AV and PT depends on policy and operations. Shen et al. (2018) proposed six possible organizational models of the AV-PT systems based on the existing transit governance literature and the emerging experience in regulating the transportation network companies (TNC), among which cooperation and competition are two major relationships. Researchers are able to design new efficient integrated operation strategies for both AV and PT when they are operated together with policy coordination. However, in the free market where AV and PT are operated separately, the competition is unavoidable due to limited demand sources. Therefore, both of the aforementioned AV PT relationships are worth analyzing given that they are both possible for the future mobility system. As for the integration relationship, many studies has conducted quantitatively analysis based on different operation and policy assumptions. For example, Salazar et al. (2018) proposed a tolling scheme for AV and PT intermodal system, which shows significant social benefit improving compared to the situation of AV operation in isolation. Wen et al. (2018) designed a transit-oriented Autonomous Mobility-on-demand (AMoD) system, revealing the trade off between the level of service and the operational cost. However, in terms of the AV-PT competition scenario, the quantitative analysis just began and are still vague (Levin and Boyles, 2015). Some studies provide a basic evidence that the emergence of AV will seize the market share from PT (Levin and Boyles, 2015; Childress et al., 2015). But most of these studies aimed to evaluate the general effect of AV, rather than dedicating on the competition between AV and PT. Also, most of the analysis are survey-based (Levin and Boyles, 2015; Chen and Kockelman, 2016). More comprehensive studies are required for the systematical analysis from different stakeholder’s perspectives. Recently, some simulation-based study shows the powerful analysis in terms of the AV PT interaction. For example, based on the case study in Austin, Liu et al. (2017) argued traditional PT services may not survive once the shared AV services become available. The similar opinion is found in Mendes et al. (2017)’s research, who treated the shared AV services as a cost-efficient alternative to the PT system. Despite some studies have explored the AV and PT competition, most of them focused on the static interaction process. They only evaluated how the system will change when AV is introduced. The behaviors of AV and PT are assumed to be static and no interactive responses are considered. Actually, as the competitors, both AV and PT should have dynamic behavioral responses against each other to maximize their profits or market shares. But this dynamic competition have not been discussed in the literature.

1.2 Motivation and paper objectives

Analyzing the AV and PT competition scenarios is critical to understand the impact of AV application. It can also facilitate the future design of policy and organizational structures for AV and PT operators. This paper aims to quantitatively evaluate system performance (e.g. level of service, operators’ financial viability, transport efficiency) when AV and PT are profit-oriented competitors with dynamic adjustable supply strategies under certain policy constraints. At the general level, ”competition” per se is neither good or bad. The society can take advantage of the market resource allocation by embracing competition. But necessary conditions are often imposed to constrain competitive behaviors, or even mandate certain cooperative behaviors. These can be done by regulation or pricing. Regulations can mandate certain supply behaviors, while pricing can induce certain supply behaviors. We will focus on regulations in this paper and present different degree of policy constraints. At the specific level, we envision four scenarios, each representing a meaningful option of regulation. The structures of the scenarios are shown in Table 1. These scenarios can be organized by the supply policies of PT and AV. We consider whether the transport authority allows PT and AV agency to adjust their supply. ”Fixed” means operators are forbidden to adjust their supply. While ”Adjustable” means operators can adjust their supply to compete. By combining the different policies, we get the scenarios matrix. ”Status quo” corresponds to a fully-regulated organization structure without competition in the market (Shen et al., 2018). The fare, route network, service areas are all designed by the transport authorities. The operators are only responsible for operation, management, and fleet maintenance. ”AV-only” scenario corresponds to TNC-regulated organization structure (Shen et al., 2018), where the AV operation can be licensed or regulated by public authorities, but the services are still provided by the private operators and operated independent from transit services (e.g. in New York City, London and Singapore). ”PT-only” scenario corresponds to ”Scandinavian” organization structure (Costa, 1996; van de Velde, 1999), where the transport authority sets service goals and then contracts out the transit service to private operators. PT operators can decide the operation strategies to maximize the profits while ensuring the service standards. ”AV-PT” scenario is close to the UK-deregulation structure (Wilson, 1991), where PT and AV operators are profit-oriented competitors but also regulated by the public authority.

Scenarios PT supply
Fixed Adjustable
AV supply Fixed Status quo PT-only
Adjustable AV-only AV-PT
Table 1: Regulation Scenarios

1.3 Paper recapitulation and organization

According to Lesh (2013), the first/last mile connection to subway station shows the highest potential for market competition between AV and PT. Thus, we select the first-mile market in Singapore Tampines area as the case study, where only bus is considered for the PT system. The actual road network, travel demand and public transit facilities are also incorporated. In the competition system, AV is operated as AMoD to provide first-mile services with ridesharing. We assume AV and PT are operated by different profit-oriented private companies. Their objective is to maximize their own profits by adjusting the supply under certain constraints. Specifically, we assume AV can only adjust the fleetsize and PT can only adjust the headway. Service fare and bus routes are regulated by the governments thus are fixed. The AV-PT competition system is then evaluated from the perspectives of AV operator, PT operators, passengers and transport authority. The interests of AV/PT operators are financial viability. The passengers’ interests are evaluated by the level of service. And the transport authorities care about transport efficiency and sustainability. The detailed indicators are shown in the following sections.

The contributions of this paper are threefold. First, the paper considers the dynamic behavioral responses between AV and PT, which fills the research gap in the literature. We assume both AV and PT are able to adjust their operation strategy to improve their profits, then evaluate the system from multiple stakeholders’ perspectives. Second, the AV-PT competition is analyzed through four different policy scenarios, which presents the impact of regulation degree on the market behaviors. Third, the competition process is systematically analyzed through an agent-based simulation platform. The simulation platform is capable of incorporating different system settings including fleet sizes, sharing policies, operation costs, and service prices. A heuristic operator’s supply updating algorithm is also proposed and incorporated in the simulation model.

The remainder of this paper is organized as follows. Section 2 presents a literature review and identifies the research gaps. Section 3 presents design of whole simulation system. Section 4 shows the evaluation indicators of different stakeholders. Section 5 discusses the results and section 6 concludes the paper.

2 Literature Review

AV is expected to reshape the future transportation system from different perspectives. Many advantages of AV, such as absolute compliance, expanded service hours, reduced labor forces and human errors make it an efficient mode in urban tranportation, which enables to reduce the operation cost and serve demand with less fleet sizes (de Cea Ch et al., 2008; Alonso-Mora et al., 2017; Spieser et al., 2014). Given the promising application of AV, many recent studies has examined the impact of AV from different point of views. For example, Azevedo et al. (2016) applied an integrated agent-based micro-simulation model to design and evaluate the impact of AV on people’s travel behavior. It is found the new AV technology can change people’s activity travel patterns, specifically in terms of modal shares, routes and activity destinations. Fagnant and Kockelman (2014) designed an agent-based model to evaluate the impact of shared autonomous vehicle (SAV). They found each SAV can replace around 11 conventional vehicles, but adds up to 10% more travel distance than comparable non-SAV trips. There are also a bunch of empirical studies in Singapore (Marczuk et al., 2015), Lisbon (Martinez and Viegas, 2017), Toronto (Kloostra and Roorda, 2019) and New Jersey (Zhu and Kornhauser, 2017). As the critical component of urban transportation, the PT system may also undergo tremendous changes in the future. However, despite the large number of studies regarding AV’s impact, the number of research focusing on the AV PT interaction is limited. Two major relationships between AV and PT have been discussed in the literature.

  • First, AV and PT can be complementary and integrated. AV and PT can be cooperative in a way that AV is integrated as a part of PT system for social welfare. In line with this assumption, Ruch et al. (2018) used a simulation model to analyze whether AV can substitute the rural PT lines. It was shown that such a service can operate at lower cost and higher service level when the PT lines are short and with less demand. Similarly, Shen et al. (2018) designed an AV-PT integration system with AMoD as an alternative to low-demand bus routes, which showed the integrated system has the potential of enhancing service quality, occupying less road resources and being financially sustainable. Another cooperative scenario is the multi-modality between AV and PT. For example, Yap et al. (2016) conducted a stated preference survey to analyze the use of automated vehicles as egress of train trips. They found AV is a good alternative to the first class train passengers. Liang et al. (2016) used an integer programming model to study AV as a last-mile connection to train trips, which showed using automated taxis can decrease the relocation costs and improve the profit. Vakayil et al. (2017) proposed a hybrid transit system where AMoD serves as the first and last mile connection for metro system. The results show that an integrated system can provide up to a 50% reduction in total vehicle miles traveled. These studies provide sufficient analysis toward the AV-PT integration relationship. Also, the co-operation strategies for AV and PT are also designed (Salazar et al., 2018).

  • Second, AV can be competitive with traditional PT when AV is operated privately for market share and profit. The analysis for this scenario is not sufficient. Based on a multiclass four-step model, Levin and Boyles (2015) found the transit ridership will decrease when AV is introduced. Similarly, Childress et al. (2015) used an activity-based model to evaluate the impact of AV, which revealed the decrease of PT’s mode share. Liu et al. (2017) simulated the impact of shared AV based on the case study of Austin. The results showed conventional PT services may not survive once the shared AV services become available. Mendes et al. (2017) developed an event-based simulation model to compare the shared AV services with the proposed light rail services in New York City. It is found AV is more cost-efficient in providing the same level of service. Basu et al. (2018) applied an activity-driven agent-based simulation approach to test the impact of AV on mass transit and found AMoD will indirectly act as a replacement to mass transit.

Understanding how AV compete with PT is critical for managing future PT services and developing sound intervention on private transportation market. However, among the existing studies addressing AV-PT competition, they only focused on the static interaction process, which looks at what will happen when AV is introduced. However, as competitors, it is highly possible that AV and PT will dynamically adjust the operation strategies in the market. But this interactive behaviors have not been considered in the literature. Also, most studies mainly evaluate the AV’s impact on the mode share of PT, lacking the comprehensive assessment on cost-benefit of different role players including operators, passengers, and transport authorities.

This study applies an agent-based model to simulate the competition scenario of AV and PT considering the dynamic operation behaviors. We also comprehensively evaluates the system from different stakeholders’ perspectives, including the financial conditions, level of services, sustainability and transport efficiency, which aims to fill the research gap in the literarture.

3 System Design

To demonstrate the AV-PT competition system, we select the Singapore Tampiness Planning Area as the study region, with the focus on the first mile trips heading to Tampiness Mass Rapid Transit (MRT) station from surrounding residential blocks. Public transit in Singapore is highly regulated by the Land Transport Authority (LTA), which is responsible for service integration including fares, information, and route design. The LTA tenders out bus services with contract for five years. The bus operators compete for the market to win the contract and enter the market, while the bus routes are still designed by the LTA. In Singapore, walking and bus are currently the dominating modes for the first mile trips. (Mo et al., 2018).

3.1 Basic Assumptions

Based on the characteristics and operational structures of the Singapore transport system, the following assumptions are made for the AV-PT competition system.

  1. System level

    • Travel mode: Before AV entering the market, walking, bus and ride-hailing are the only travel modes available for the first-mile trips. After AV emerges, ride-hailing will be replaced by the AMoD. This assumption is corresponding to the Singapore autonomous vehicle initiative policies (LTA, 2017).

    • Traffic condition: We focus on the mesoscopic simulation, the microscopic features such as road capacity, signal system and driving behaviors are not considered in this study.

    • Demand and supply: The total travel demand and time of day distribution are assumed to be fixed for all simulation days. The supply of AV and PT can be adjusted for competition.

    • Information: The AMoD and PT’s supply information are available for passengers to make modal choice given the widespread information communications technology.

  2. Agent level

    • Ownership: the PT and AMoD are both operated by private enterprises, but also supervised by the government authority.

    • Objective: In the context of market competition, PT and AMoD are able to adjust their supply to improve their profits.

    • Fare: PT and AMoD’s fare structures are regulated by the government and cannot be changed.

    • Operation cost and subsidy: PT and AMoD’s operation cost are proportional to the provided fleetsize and driving distance. There is no operation subsidy for PT based on the current Singapore policy (Kuang and Gopinath, 2018).

    • Constraints: PT has a lower bound and upper bound for the bus headway.

    • Dynamic intraday supply: AMoD and PT’s supply can be varied across time in a specific day.

    • Supply updating frequency: AMoD can update next day’s supply strategy at the end of this day. PT can only update its supply strategy after a sufficient long time (discussed in Section 3.4).

The supply updating for AMoD means determining the vehicle fleet size provided in each time interval during the day. For PT, it means determining the headway for each bus route in different time intervals. The price adjustment is also a powerful approach in market competition. However, the co-updating of price and supply may make the decision process quite complicated, which is beyond the scope in this study. On the other hand, given the high regulation of LTA in Singapore, it is possible that the AMoD are forced to fix their fare structure to avoid the malicious price competition. This can implied from Singapore’s recent effort to regulate the ridehailing pricing LABEL:PTC2019bill. Therefore, we only consider the supply adjustment for each operator. Future research can explore on the co-updating problems with multi-dimensions, incorporating the supply, price and bus routes.

3.2 Data and Study Area

Tampines is a mixed residential and working area located in the east of Singapore. It is centered around the Tampines MRT station with frequent and high-speed train service. The MRT travel demand here is one of the highest among all MRT stations in Singapore (Housing & Development Board, 2015). The total area of the study zone is 6.86 , with 51 bus stops spreaded around (See Figure 1). There are 26 bus routes passing by the MRT station in this area, where three of them are MRT-connected routes, which only serves in the study region as first-mile connection. The rest 23 routes also serves other regions. All these routes are included in the simulation model. Bus service plays an import role in Singapore. Between 7:00 and 9:00 AM on a typical workday, more than 15,000 passengers use the MRT service, including over 8000 taking a bus to access the train station. Tampines is chosen as the case study area because 1) it possesses a significant first-mile demand to the MRT station and 2) it has sufficient bus service supply which allows for competition analysis.

Figure 1: Study Area

The first-mile travel demand is collected from the transit smart card data. The data covers all PT trips in August 2014, with over 175 million travel records. A normal workday in August 2014 is selected as the study date. Since the smart card data in Singapore includes both the tap-in and tap-out information in bus and MRT systems, the accurate date and time of every entry and exit activity are recorded, as well as the boarding and alighting stops/stations, which allows us to extract first-mile trips. The temporal distribution of first-mile travel demand in study date is shown in Figure 2.

Figure 2: Hourly First-mile Demand of Tampines MRT Station during Study Date

3.3 Agent Behaviors

The overall simulation system is composed of three main agents: public bus, AMoD and passenger. The PT system is derived from Shen et al. (2018)’s study, which has been well calibrated and proved to be in line with the real-world situation. The behaviors of each agent are described as following.

3.3.1 Passenger Behavior

Passengers are assumed to enter the system every one minute during the simulation day. To mimic the AV and PT competition process, the simulation need to be run for a long time (e.g. one year). Passengers demand may be changed over time given the change of supply. However, the interaction between demand and supply is a fixed point problem, which deserves a separate study (Wen et al., 2018). In this study, we assume the passenger demand patterns will not vary across different days (i.e. passenger demands for all simulation days are fixed to Figure 2). The Voronoi diagram based on the location of bus stops is used to assign travel demand to each building.

Passengers’ mode choice is modeled based on the result of a mixed logit model

(Shen et al., 2019)

. The trip-specific variables (travel cost, waiting time, on-vehicle time, walking time, etc.) and sociodemographic variables (e.g. income) are considered. This demand model is estimated from an AV preference survey in Singapore, which is well matched with this study. In addition, the mixed logit model allows us to capture the preference heterogeneity among people. Each individual is assigned a set of choice coefficients drawn from the pre-determined distributions. The detailed model coefficients are shown in


When a passenger is generated, we will first calculate the corresponding trip-specific variables based his/her location. Since route planning problem is not the focus of this study, we assume only the nearest bus station is available to the passenger. The corresponding walking time, waiting time, and on-vehicle time are then calculated. Note that the waiting time for AMoD is inversely proportional to the AMoD’s supply, while the expected waiting time for PT is half of the headway. When the current time is not in the route’s operating time, the bus mode will be set as unavailable. Passengers’ incomes are drawn from the distribution derived from Singapore household survey data. The mixed logit model is used to calculate the selection probability of different modes. During the simulation process, people’s mode choice will change with change of AV and PT’s supply. Considering the imperfect information transfer, a lag effect is added to people’s behavior. Denote the

model-derived probability of passenger choosing mode in day is . The actual probability used for simulation is calculated in Eq. 1.


where is the lag factor, which represents how much people’s behavior will be lagged by the previous day. After calculating the choice probability, passengers will be assigned a specific travel mode accordingly.

When walking is chosen as the travel mode, the passengers will walk directly to the MRT station along the footway. When the AV is chosen as the travel mode, the passengers will start to call for the ride repeatedly until the system has successfully booked a car for him/her. Once a vehicle is assigned, he/she then moves to the pick-up location to wait for the AV. As the density of bus stops in Singapore is very high, the pick-up location is set as the closest bus stop from the passenger. To prevent a passenger from waiting too long when no AV is available, a maximum waiting time is set to 10 min in the study, beyond which he/she will forgo the AV request and travel by bus if available, otherwise he/she will travel by walking.

When the bus is chosen, the passenger will walk directly to the bus stop to wait for the next bus. Since we assume the bus agent can update its headway to maximize its profit, it is possible that people may wait for a long time. Therefore, we set a maximum waiting time (30 min) for bus as well. People will switch to AV or walking based on the choice probability when the waiting time is beyond the threshold.

3.3.2 AMoD Behavior

The concept of the mobility-on-demand system was proposed and demonstrated in the 1970s, and was called Dial-a-Ride Transit (DART) (Wilson et al., 1976). In this study, the AMoD service with ridesharing resembles a carpooling system (Galland et al., 2014; Martinez et al., 2015). Only one type of service, namely first mile, is provided. And the service area is restricted to our proposed study zone. This arrangement is consistent with the proposed usage of the AV prototype in Singapore (Chong et al., 2011). The driving behaviors for all AVs are identical, whose routing follows the shortest path without considering the traffic. AVs are assumed to fully comply with the central controller and never reject the service requests from customers. Each AV allows a maximum of four passengers to share the ride. AV’s initial supply is approximated using Singapore Taxi data. Since there is no first mile demand when the MRT systems closed, we adjusted initial AV supply to zero during that time. New AVs are generated from several car clubs (see Figure 1). When the AV is requested to stop the service, it will return to the nearest car club and be removed from the simulation system. AVs will stay idle when there is nobody call for the service.

The operation cost of AMoD agent consists of two parts: fixed cost and variable cost. They are calculated with a resolution of one time interval (i.e. 1 hour for AV). The fixed cost counts for the fee of making one AV available during this hour, such as depreciation cost, parking cost. We use the lowest car renting fare in Singapore (4 $SG/hourveh) to approximate the fixed cost. The variable cost is calculated based on vehicle travel distance and gasoline fare. It is set as 0.12 $SG/km in this study.

The ridesharing only happens under specific conditions. Given the heterogeneous ridesharing preference among population (Nazari et al., 2018), we define a ridesharing-agreement ratio (50% in this study). That is, only half of people are willing to share trips with others. For the short distance first-mile trips, sharing with others is not a good choice given the potential surge in detouring and travel time (Schreieck et al., 2016). Therefore, we assumed all passengers prefer to ride alone if possible, even they agrees to share the ride. Thus, when a passenger calls for a ride, the system first scans all empty AVs. If there are empty AVs, the system assigns the closest available AV to pick up the passenger. Once the AV and the passenger are matched, a notice is sent to the passenger to request a meet-up at the pick-up point. If there are no empty AVs, the system searches for all occupied AVs. Passengers who have agreed to share the ride will be considered to take a sharing ride. An AV is shareable only if 1) it have available seats, and 2) the incremental travel time to on-vehicle passengers due to picking up new passengers cannot exceed a pre-determined threshold. Detailed calculation method can be found in Shen et al. (2018). We share the same parameters setting in the ridesharing module.

The fare structure of AMoD is analogical to the taxi service in Singapore, with a base fare within first km and a distance-based fare beyond km. The difference is that we add a ridesharing discount for AMoD service. The total travel fare for passenger , is formulated in Eq. 2.


where is the base distance, which equals to 1 km in this study; is the base fare within ; is the distance-based fare per km after ; is the direct travel distance of passenger in km; is the detoured actual travel distance in km; is the detour ratio; is the discount degree offered due to the detour.

In terms of competition, the AMoD enterprise could update the hourly supply to improve its profit. Given the complex interaction between agents, it is very difficult to propose an algorithm which ensures every adjustment is optimal for the AMoD enterprise. Also, such algorithm is beyond the scope of this research. In this study, we propose a heuristic supply updating algorithm for AMoD. Instead of finding the optimal adjustment, we assume a fixed step size of each updating. The sign of the updating (i.e. supply increase or decrease) is determined by previous supply and profit data, which aims to make the profit higher. This concept is also in line with the reality, in which information is incomplete and every adjustment is based on previous experience. Detailed description of the algorithm can be found in Section 3.4.

3.3.3 PT behavior

Buses services are schedule-based, which are operated based on the given routes and time tables. The bus routes and initial schedules information are provided by LTA. Once a bus is dispatched, it will follow the route with a constant speed, passing through a sequence of stops. The spatial distribution of bus stops can be found in Figure 1. Upon arrival at a bus stop, each bus dwells for a certain amount of time to pick up passengers. The dwelling time is set as 30 second in this study.

The operation cost of bus agent is purely distance-based, which is 2.71 $SG/km. This calculated from the annually financial reports of two major bus companies in Singapore (SBS Transit, 2017; SMRT Corporation Ltd, 2017), which involves the labor cost, depreciation cost, fuel cost, maintenance costs, etc.. We assume all the operation cost is proportional to the distance traveled. The bus fare structure is set based on the real-world scenario. All passengers are charged a fixed fee of 0.77 $SG per trip.

To compete with AMoD, the bus enterprise is allowed to adjust its supply strategies (i.e., the headway) to increase the profits as well. The adjusting algorithm is similar to AMoD’s, while the difference is that the adjusting frequency is lower than that of AMoD. To ensure the basic service of PT, we set a upper bound for the adjusted headway. Every step of headway adjustment cannot exceed the upper bound, otherwise passengers may have to wait too long and get annoyed. On the other hand, to avoid the too frequent bus dispatching rate, a headway lower bound is also introduced. Since we calculate the revenue and cost based on the first-mile demand, it is important to ensure the supply adjustment does not affect the market outside the first mile. There are two types of bus routes in this study. One is the MRT-connected buses, which only serve in the study region. The supply adjustment for this route only applies to the first-mile market. Another is the normal bus routes that passes through the study region. If we directly adjust the headway for these routes, it will not only affect the demand for first-mile trips, but also affect other passengers who use the bus outside the study region. Thus, we need to redefine the supply adjustment for these routes. Here, we assume the supply decrease (i.e. increase of headway) for the normal routes is equivalent to that some buses of this route will not stop in the study region. Instead, they directly drive cross the region along the shortest path. For examples, as shown in Figure 3, after supply decrease, some of buses of route 21 will be rerouted to the shorter green path and not stop in the study area. The corresponding decrease in the operation cost is calculated by the decreased travel distance. Note that the decreased operation cost may be small under this settings. In terms of supply increase (i.e. decrease of headway) for the normal bus routes, we assume this is equivalent to adding new MRT-connected bus routes with the same stations in the study region. The corresponding increase of operation cost is calculated as the cost of running this new route. In this way, we are able to isolate the first-mile market, where the supply adjustment for all bus routes will not affect the outside areas.

Figure 3: Rerouting example of route 21

3.4 Simulation platform

We use the AnyLogic 8.1 software to perform the simulation. The simulation is executed for days to make sure the competition process reaches a stable state. There are four parts of input variables/paramaters needed for this model: simulation setting parameters , initial supply strategies of bus , initial supply strategies of AMoD , and the first-mile demand . The detailed notations and values used are given in Table 2. The supply, profit, and supply changing unit are all time-specific for AV, which means for different days and different time intervals they have different values. For bus, these variables are time and route specific, which means they are for different routes and time intervals. The pseudo code for simulation is shown in Algorithm 1.

Categories Parameters/Variables Value
Simulation setting
Duration of simulation () 365 days
Supply changing unit reduced factor () 0.5
Passenger choice lag factor () 0.5
Passenger mode choice parameters () See Table 5
Passenger maximum AV waiting time 10 min
Passenger maximum bus waiting time 30 min
Passenger ridesharing-agreement rate 50%
AV supply time interval () 1 hour
AV supply updating frequency () 1 day
AV fixed operation cost 4 $SG/hourveh
AV variable operation cost 0.12 $SG/km
AV base distance of fare () 1 km
AV base fare () 3.4 $SG
AV distance-based fare () 0.22 $SG/400m
AV detour discount rate of fare () 2
AV supply lower bound () 0
AV supply upper bound ()
AV initial supply changing unit () 10 vehicles
Bus supply time interval () 2 hour
Bus supply updating frequency () 30 days
Bus operation cost 2.71 $SG/km
Bus dwell time 30 sec
Bus fare 0.77 $SG/trip
Bus headway upper bound () 2100 sec
Bus headway lower bound () 210 sec
Bus initial headway changing unit () 3 min
Bus supply Bus headway of route in time interval on day () Intermediate
Bus headway arrangement on day () Intermediate
Bus initial headway arrangement () See section 3.3.3
AV supply Number of AV supplied in time interval on day () Intermediate
AV supply strategies on day () Intermediate
AV initial supply strategies () See section 3.3.2
Demand Demand of first-mile trips See figure 2
AMoD profit in time interval on day () Intermediate

AMoD profit vector on day

AV supply changing unit in time interval for day () Intermediate
AV supply changing unit vector on day () Intermediate
Bus profit of route in time interval on day () Intermediate
Bus profit vector day () Intermediate
Bus headway changing unit of route in time interval for day () Intermediate
Bus headway changing unit vector on day () Intermediate
  • Values ”Intermediate” means the intermediate variables in the model.

Table 2: Notations and Values
1:procedure Simulation(, , , )
2:     initialize , , ,
3:     initialize ,
4:     let day counter
5:     while  do
9:         if  then
11:              initialize
12:         else
15:     return the system evaluation indicators (3)
Algorithm 1 Agent-based Simulation

As shown in Algorithm 1, the overall framework aims to simulate the interaction between AV and PT over time period . One-Day-Sim is a pseudo function of running the simulation for a single day given the supply, demand and , which can be seen as the engine of the simulation model. In this function, each agent will follow the behaviors mentioned in Section 3.3. The profits of AMoD and PT can be obtained after running this function. The profit, supply and supply change unit are then used as the input of function SupplyUpdate (shown in Algorithm 2), deriving the updated supply strategies and new supply change unit for AV and PT. It is worth noting that since we assume the AV supply updating frequency is one day, this function works for AV everyday to get new supply strategies. However, for PT, this function only works every days considering the inflexibility of PT. In reality, the schedule of PT should not be adjusted frequently considering people’s expectation for its reliability. A reasonable changing frequency should be more than 6 months, which allows the transit company to notice the public, and gives people enough time to prepare for the new schedule. However, based on our numerical test, given a specific bus schedule, the system will converge in one or two weeks when only AMoD updates its supply. So, any value of that is greater than two weeks should be able to simulate the long-period updating frequency of bus. Therefore, days are used in this study. The system evaluation indicators, which will be introduced in Section 4, are recorded during the simulation process, and returned at the end of the model.

1:procedure SupplyUpdate(, , , )
2:     let the index of elements in , , and correspond with each other. (i.e. the element of , , and represents the information of same time interval and same route.)
3:     initialize
4:     while  do
6:         let , , be the element in , , and , respectively.
7:         if  then
9:         else
12:         if  violates the upper or lower bound constraints then
13:              set equal to the upper or lower bound
15:         if the difference between and is small enough then:
17:     let and
18:     return ,
Algorithm 2 Supply Updating

In terms of supply updating, we proposed a heuristic algorithm, which applied to both AMoD and PT. We assume the profit state of different time intervals and routes are independent, and it only depends on the corresponding supply. For example, the profit of bus route in time interval on day () only depends on the headway of route in time interval on day (). Then, we can adjust the to improve the corresponding profit as a single-variable optimization problem. So as the profits in other time intervals and routes. Take the AMoD agent in day as an example, if is greater than , we know the last change of supply () led to the profit increase, so we should insist on the same change in next day. Then the corresponding supply change for day is . And the supply for next day (of the same time interval) is . If the change of profit between two supply strategies is small enough (in this study we set the relative difference threshold as 5%), we decrease the size of changing unit by (i.e. ). This setting ensures the final supply could reach a converged stage. The heuristic algorithm is a simplification for the profit maximization problem because the agent’s profit of different time intervals and routes are possibly correlated, especially when the routes have overlapping stations. However, we apply this algorithm for the following reasons. 1) The concept of this algorithm is in line with the reality, where information is incomplete and every adjustment is a trial of a new supply strategy based on previous experience. 2) Capturing the dependency may lead to a complicated optimization problem, which is beyond the scope of this paper. 3) Based on the numerical test, this heuristic algorithm can yield a continually improving profit. So it reaches the purpose of mimicking agents competition for profit improvement, which is enough for this research.

4 Simulation system evaluation

To analyze the AV-PT competition system, we will evaluate the interest of multiple stakeholders. Table 3 identifies four key stakeholders (passengers, AMoD operators, PT operators and transport authority) their evaluation indicators.

For passengers, the main interests are level-of-service and modal choice. The level-of-service indicators include travel cost, total travel time, waiting time and generalized travel cost. The generalized travel cost is calculated as the sum of walking time, waiting time and riding time multiplied by the corresponding value of time (VOT). It is a comprehensive value which incorporates both travel time and travel cost. The VOTs are derived from the estimated choice model in Table 5. In terms of modal choice, the number of passengers choosing walking, bus and AV are recorded, respectively.

For AMoD and PT operators, based on our profit-oriented assumption, their interests are financial viability and supply. The financial viability indicators includes operation cost, revenue, profit and market share, while the supply means the number of AV/bus provided. From the transport authority perspective, the transport efficiency and sustainability are considered. The average load per vehicle is one of the indicator for transport efficiency, which is calculated by the the total passenger travel distance divided by vehicle travel distance. As for sustainability, we consider the vehicle kilometers traveled (VKT) and passenger car equivalent (PCE). The PCE for bus is set as 3.5 in this study (Ahuja, 2007).

These indicators are recorded in the simulation process, which shows the whole changing profile during the competition between AV and PT. The time of day distribution of some indicators are also recorded.

Stakeholder Interests Indicators
Passengers Level of service Travel cost
Total travel time
Waiting time
Generalized travel cost
Modal choice Walk demand
Bus demand
AV demand
AMoD Operators Financial viability Operation cost
Market share
Supply Average number of AV provided per hour
PT Operators Financial viability Operation cost
Market share
Supply Number of bus dispatched per day
Transport Authority Transport efficiency AV average load per vehicle
Bus average load per vehicle
Sustainability AV VKT
Total PCE (AV and PT)
Table 3: Evaluated Indicators of Stakeholders

5 Results and analysis

5.1 Scenarios settings

The simulation setting parameters are shown in Table 2. Although the value of can affect the simulation results, it is beyond the scope of this paper to systematically explore all possible scenarios. For illustration purposes, we only explored the four scenarios with specific policy implications in Table 1. We summarize the characteristics of these four scenarios as below.

  • Status quo: this scenario assumes both AV and PT cannot adjust their supply, i.e. they always operate as the initial state. The initial state is described and justified in Section 3 based on the situation in the study area.

  • PT-only: this scenario assumes only PT can adjust its supply (headway) every month (i.e. days). The supply of AV is fixed. The initial state for this scenario (and all other scenarios) is same to that of Status quo.

  • AV-only: this scenario assumes only AV can adjust its supply (fleetsize) every day (i.e. day). The supply of PT is fixed.

  • AV-PT: this scenario assumes both AV and PT can adjust their supply. For AV the frequency is every day and for bus is every month.

In following subsections, we will show how the indicators of different stakeholders change during the whole competition process. The following paragraphs are organized by different stakeholders’ perspectives.

5.2 PT Perspective

The revenue, operation cost and profit of the PT operator are shown in Figure 4. One point represents the average value of the corresponding month (same for all following graphs). Since the bus will not change supply in AV-only and Status quo scenarios, the curve for these two scenarios are relatively stable. The patterns for PT-only and AV-PT scenarios are similar. The profit of bus first increases and then becomes stable during the simulation, which proves the effectiveness of proposed supply updating algorithm. The final profit of PT-only scenario is higher than all others’, which is reasonable because there is no AV competition in that situation. The final profit of AV-only scenario is lower than the status quo, which shows the competition of AV can chase the benefit from PT. Looking at the revenue and operation cost of PT, we find they both decreases over the simulation period, where the operation cost shows a sharper reduction. This implies the strategy for bus in competition is to reduce the amount of services so as to reduce the operation cost. In terms of bus supply and market share (see Figure 5), we find the similar phenomenon. The number of bus dispatched per day (i.e. bus supply) keep decreasing and PT’s market share shows the similar diminishing trend. Since bus will not change supply in PT-only and status quo scenarios, the supply curve for these two situations are flat lines. Comparing different scenarios, we find bus will decrease to the lowest service frequency with AV’s competition. Correspondingly, the market share for AV-PT scenario is also the lowest. The operation cost of AV-PT and PT-only scenarios are similar though their final supplies (Figure (a)a) are different. This is because most of the bus routes are not MRT-connected routes and only crossing the study area. These routes can only be rerouted (see Section 3.3.3), thus the corresponding decrease of operation cost is very limited. Another interesting phenomenon is the overlay effect of supply adjustment of both bus and AV. Take the bus market share as an example. The curve of PT-only scenario represents how much unprofitable market share the PT agency gives up (impact from bus supply adjustment). The AV-only curve shows how much market share the bus is grabbed by AV (impact from AV supply adjustment). The curve of AV-PT scenario thus shows the overlay effect of two, which is nearly the sum of the above two reductions.

(a) PT Revenue
(b) PT Operation Cost
(c) PT Profit
Figure 4: PT Finance over the Simulation Process
(a) PT Supply
(b) PT Market Share
Figure 5: PT Supply and Market Share over the Simulation Process

In addition to the competition process, we also analyze the before-and-after supply distribution. Figure 6 shows the temporal and spatial bus supply distribution before and after the simulation. Note that AV-only and status quo scenarios do not have variable bus supply so they are not plotted. The bus supply of the first month are same for PT-only and AV-PT scenarios so only one curve are shown. From the temporal distribution, we found after the supply adjustment, the number of dispatched bus for most time periods except for morning peak (6:00 - 10:00) are reduced. For PT-only scenario, the bus supply of evening peak also does not decrease. These suggest the PT agency will concentrate supplies to morning peak and evening peak, which have more demand and are more profitable. In terms of the spatial distribution, some routes (e.g. 291, 282) are distributed with more bus services, while some routes (e.g. 2912, 2921) are adjusted to a low service rate. The geographical information of these routes are shown in Figure 7. It is found increased-supply routes are usually short and cost-efficient, which cross the residential areas and goes direct to the MRT station. The routes with decreased supply are usually long and have high detouring rate. Great reduction in operation cost can be achieved by decreasing the supply of these routes. Therefore, the spatial distribution implies some coordination within the bus routes (even if our algorithm does not consider the inter-routes coordination). The PT agency will decrease the supply for high-cost routes, thus transfer the demand to the low-cost routes, which makes the overall operation more profitable.

(a) PT Supply Temporal Distribution
(b) PT Supply Spatial Distribution
Figure 6: PT Before-and-after Supply distribution
(a) Supply Increase Routes (AV-PT)
(b) Supply Decrease Routes (AV-PT)
(c) Supply Increase Routes (PT-only)
(d) Supply Decrease Routes (PT-only)
Figure 7: Spacial Distribution of PT Supply Change

5.3 AMoD Perspective

Figure 8 shows the AMoD revenue, operation cost and profit over the simulation period. Since there is big changes of AMoD’s supply happening in the first several days, the curves for the first month are also shown with the red line, which corresponds to the upper x-axis with red color. Note that the first month curve for the AV-only and AV-PT scenrios are same because the bus supplies stay the same, and AV does not change for another two scenarios. So only one curve are shown for the first month. From the figures we find the AMoD’s revenue, operation cost and profit all increase and then become stable during the simulation periods, which suggests the AMoD will provide more services to improve its profit in the competition. The trend of AMoD’s supply and market share also validate this. As shown in Figure 9, the AMoD’s supply and market share keep increasing during the first several months, and then become stable. The major change of AMoD’s supply from the status quo happened in the first month. This is because the initial AV supply is approximated by the taxi service, which is not commonly used as the first-mile mode in Singapore (Mo et al., 2018).

Comparing different scenarios, we find the AV-PT scenario can reach the highest AV profit and market share, though the AV supply for this scenario is only the second highest (AV-only is the highest), which implies the bus supply adjustment can not only improve the profit of bus, but also benefit the AV. This is because the initial bus is over-supplied for the first-mile market. When bus is allowed to adjust its supply, the abandoned unprofitable demand of bus is then served by AV. This observation corresponds to the previous research of AV-PT integration system (Shen et al., 2018). Therefore, we may infer that, when the PT is oversupplied, despite that we assume AV and PT are competed with each other, they will still show the cooperative attributes to some extent, which results in higher profit for both of them. However, we must notice that the cooperative attributes are only shown in the bus supply change dimension. In terms of AV supply change dimension, as we discussed above, it is only unilaterally beneficial to AV. Another interesting thing can be found in AMoD’s operation cost, where the value of AV-PT scenario and AV-only scenario are very similar despise they have different supplies. This can be explained by the VKT figures in Section 5.5. Though the AV-PT scenario has less AV supply, it can produce more VKT (more distance-based cost) than the AV-only scenario, which suggests the AV utilization rate in AV-PT scenario is higher. We can also observe the overlay effect of supply adjustment on AV. But different from the effect on bus, the two dimensions both benefit to AV. Thus the AV-PT scenario has the higher profit and market share than both AV-only and PT-only scenarios.

Similar to our analysis for PT, we also plot the time of day distribution of AMoD supply before and after the simulation. As shown in Figure 10, the converged supply patterns of AV-PT and AV-only scenarios are similar. After 12 months of adjustment, the supply has been re-distributed across time. More supply will be provided in the morning and evening peak hours, which is considered more profitable. And the supply in off-peak hours (e.g. 11:00 - 13:00) stays the same or becomes even lower. This temporal changes is corresponding to the demand distribution in Figure 2.

(a) AMoD Revenue
(b) AMoD Operation Cost
(c) AMoD Profit
Figure 8: AMoD Finance over the Simulation Process
(a) AMoD Supply
(b) AMoD Market Share
Figure 9: AMoD Supply and Market Share over the Simulation Process
Figure 10: Time of Day Distribution of AV Supply

5.4 Passenger Perspective

As for passengers perspective, we are interested in the level of service and their mode choice. The level of service is shown in Figure 11. During the simulation process, passengers’ travel cost shows different patterns for different scenarios. It will decrease for PT-only scenario and increase for AV-PT and AV-only scenario, though the magnitude of change is relatively small (around 0.03 $SG per people). This trend is reasonable because the travel cost for AV is more expensive. When AV starts to compete and serve more demand, the average travel cost will increase. When there is only bus adjusts its supply, a great number of people will change to walk and only a few people change to AV (as shown in Figure 12), the average travel cost will decrease. In terms of the total travel time, we find all scenarios show the decreasing trends. This indicates for PT-only scenario, both passengers’ travel time and travel cost will decrease after the adjustment of bus supply, which implies absolute benefits to social welfare. As for the AV-PT and AV-only scenarios, we cannot directly conclude the actual effect on passengers since the change of travel time and cost are inconsistent. Therefore, to capture the effect of both travel time and travel cost, we calculated the generalized travel cost, and assumed the level of service can be reflected by the generalized cost curve. As shown in Figure (d)d, passenger’s generalized travel cost decreases for all scenarios, where the AV-PT scenario shows the largest reduction. This implies the supply adjustment of bus and AV can not only benefit the operators, but also increase the level of service for passengers. The major contribution is via the decrease in travel time. Another important indicator for passenger is the waiting time. As shown in Figure (c)c, for PT-only scenario, passenger’s waiting time keeps increasing, which makes sense because of the reduction of bus supply. For AV-only scenario, passenger’s waiting time keeps decreasing, which can be explained by the addition of AV supply. Moreover, in terms of AV-PT scenario, we can observe the overlay effect of the above two. The gap between first month and status quo is caused by the AV supply increasing in the first month. Start from the first month, the passenger’s waiting time gradually increases, which shows the effect of bus supply reduction. But due to the effect from AV side, the increase in waiting time is not as sharp as the PT-only scenario. Finally, the converged waiting time is still less than the status quo.

The passenger mode choice for the first-mile trips are shown in Figure 12. In general, the demand for AV will increase, while the demand for bus decreases. The demand for walking varies across different scenarios. For AV-PT and PT-only scenario, more people will choose to walk due to the bus supply reduction. While for the AV-only scenario, less people will walk because of the increased supply of AV. In addition to the overall variation among three modes, we are also interested in the attributes of people who change their behaviors. For the sake of illustration, we only analyze the AV-PT scenario. Two attributes of passengers are explored, the household income and the distance to MRT station. We aim to answer what types of people will change their mode choice (from bus to AV/walking). People who originally chooses bus are set as the baseline group. People who change their travel modes from bus to AV or walking are set as the experiment groups. Our purpose is to judge whether there is statistically significant difference between the baseline groups and the experiment groups in terms of their household income and distance to MRT station. The Kolmogorov-Smirnov (KS) test for 2 samples is applied to this problem. Table 4 summarizes the comparison results. For household income, we found people who change their mode choice from bus to walk do not show significant difference from baseline group. However, as for people now using AV, the household income shows a significant difference. The average income for this group is $SG 5222.8, while the baseline group is $SG 4780.9. This implies higher income people tend to change their choices from bus to AV after bus decreasing its supply. But there is no evidence showing low income people are forced to walk, which implies the competition between AV and PT does not deteriorate the condition of the poverty. In terms of the distance to MRT station, both experiment groups shows significant difference from the baseline group. By comparing the average value, we found people who live near MRT stations tend to convert to walk, while people who live far from MRT stations tend to convert to AV. This implies the interaction of AV and PT tends to polarize people’s travel mode choices structure. On the other hand, it motivates more people to walk when the distance is not so far, which can be seen as a good phenomenon for public health and sustainability.

(a) Travel Cost
(b) Total Travel Time
(c) Waiting Time
(d) Generalized Travel Cost
Figure 11: Level of Service during Simulation
(a) AV Demand
(b) Bus Demand
(c) Walking Demand
Figure 12: Passenger Mode Choice over the Simulation Process
Attributes Groups Mean (Std.) KS Statistics p-value
Household Income ($SG) Baseline 4780.9 (3669.7) N.A. N.A.
Experiment (Bus to AV) 5222.8 (3734.4) 0.061 0.000*
Experiment (Bus to Walk) 4773.0 (3635.6) 0.007 0.963
Distance to MRT Station (m) Baseline 1065.0 (334.2) N.A. N.A.
Experiment (Bus to AV) 1130.2 (313.7) 0.087 0.000*
Experiment (Bus to Walk) 924.5 (311.6) 0.177 0.000*
  • Baseline groups are set as reference, so they do not have KS statistics and p-value.

  • The larger the KS Statistics, the greater the difference between the two groups.

  • *: Significant at 99% confidence level.

Table 4: KS Test Results for AV-PT Scenario

5.5 Transport Authority Perspective

The interest for transport authority is the transport efficiency and the sustainability, which are shown in Figure 13 and Figure 14, respectively. In Figure 13, we found the AV average load slightly increase for PT-only scenario. However, for the AV-PT and AV-only scenario, the AV load decrease a lot in the first month (the gap between the status quo and the first month) and then becomes stable. This suggests that to maximize the profit, ride-sharing behavior is not encouraged in this model. This may be because in the first-mile scenarios, travel distance are usually short. Hardly can ride-sharing happens or generates higher profits for the operator. On the other hand, considering the price structures of AMoD we proposed, serving two passengers with two vehicles separately may yield more profit. Therefore, we know after the competition, though the AMoD operator earns more money, the AV’s transport efficiency are decreased. Looking at the bus average load curve, different from AV, the bus average load increases for the AV-PT and PT-only scenarios, which means after competition, the bus are operated more efficient, with not only higher profit, but also higher average load.

In terms of sustainability, we measured the VKT and PCE for bus and AV. As shown in Figure 14, in general, the VKT of AV are increased, while the VKT of bus are decreased, which is corresponding to the supply change. As for the total PCE, since the bus is dominate in magnitude, the shape of total PCE is similar to bus VKT. The system total PCE decreases for AV-PT and PT-only scenario, which means the bus policy relaxation can reduce the system traffic flow and result in more sustainability. When bus is not allowed to change its supply, the total PCE will increase due to the increase in AV VKT.

(a) AV Average Load
(b) Bus Average Load
Figure 13: Transport Efficiency over the Simulation Process
(a) AV VKT
(b) Bus VKT
(c) Total PCE (Bus+AV)
Figure 14: Transport Sustainability over the Simulation Process

5.6 Summary

To better illustrate the indicators change for different stakeholders, we summarize the corresponding diagram in Figure 15. The ”triggers” are highlighted by red color, and the arrows represent the causality. The increase and decrease of different indicators are shown by ”+” and ”-”, respectively, compared to the status quo. We found the AV-only scenario is beneficial to AV (profit+) and passengers (generalized travel cost-), but is harmful for PT (profit-) and transport authorities (PCE+). PT-only and AV-PT scenarios are both beneficial to all stakeholders, but AV-PT scenario shows higher degree of benefits.

(a) AV-only
(b) PT-only
(c) AV-PT
Figure 15: Summary of Indicators Change (Red: Triggers, Purple: Increase, Green: Decrease)

6 Conclusion and Discussion

This paper proposed, simulated and evaluated the interaction between AV and PT from a competition perspective, assuming both AV and PT are profit-oriented and aim to improve their own profit by supply adjustment. The first-mile market in Singapore Tampines is selected the case study. Four scenarios with specific policy implications are evaluated. The main findings are summarized below.

  • Overall, allowing profit-oriented bus and AMoD services to compete by adjusting supply can improve profits for both competitors while still benefiting the public and the transportation authority. Such competition forces bus services to trim inefficient routes, and allow AVs to fill in gaps in bus coverage.

  • During the competition, both AV and bus will re-distribute their supply spatially and temporally. For bus, the supply of routes with long distance and high detouring rate are decreased. The increased-supply routes are usually short, which cross the residential areas and go direct to the subway station. In temporal dimension, both AV and bus will concentrate their supplies in morning and evening peak hours, and reduce the supplies in off-peak hours.

  • The competition between AV and PT can decrease passengers’ travel time but increase their travel cost. The generalized travel cost are still reduced when counting the value of time.

  • The competition can polarize peoples travel mode choices structure. Bus demand decreases while the AV and walk demands increase. People who live near MRT stations change to walk. While people who live far from MRT stations change to use AV. Higher income people tend to change their choices from bus to AV. But there is no evidence showing low income people are forced to walk.

  • The bus supply adjustment can increase the bus average load and decrease total (PCE), which is good for transport efficiency and sustainability. But the AV supply adjustment shows the opposite effect.

  • Comparing different scenarios, we found AV-only scenario is beneficial to AV and passengers, but is harmful for PT and transport authorities. PT-only and AV-PT scenarios are both beneficial all stakeholders, but AV-PT scenario shows higher degree of benefits.

6.1 Policy implications

The comparison of four scenarios indicates that competition does not necessarily bring out loss-gain results. A win-win outcome is also possible under certain policy interventions. This can help authorities to design future AV-PT marketing policies when these two modes are operated by private companies. Since the PT supply adjustment can benefit to the transport efficiency and sustainability, government should allow PT operators to optimize its supply strategies. However, as passengers’ waiting time and travel cost will increase due to PT’s supply adjustment, government need to further discuss how much degree of freedom should be given to PT operators so as to avoid harming the interests of passengers. One possible solution is to set specific constraints or operation goals to private PT operators while allowing them to optimize supply, such as maximal headway/waiting time, minimal ridership, and minimal passenger satisfaction scores.

AV’s supply adjustment can decrease passengers’ travel time, but also decrease the transport efficiency and sustainability. The policy to AV operators should focus on limiting its negative impact to the overall system. Government can directly regulate AV’s operation by limiting the number of licenses, operation time and service areas. The regulation policy need careful design, which should combine the supply adjustment of PT. For example, government can limit AV to serve at low PT frequency areas and time periods, which makes AV more like a complementary mode to PT, especially when PT decrease its supply in some low-profit areas. Besides direct regulation, more soft policy, such as subsidy and incentive to AVs serving target passengers, can also be implemented.

At individual level, the increase of walking mode share implies that some people will be sacrificed due to AV-PT competition. Authorities may compensate people who suffer from higher travel cost or longer travel time, for example, giving them discounts of using AV and PT, and providing other feeder modes (e.g. bike sharing service, E-scooter sharing). Transport authorities can also charge AV and PT operators with certain tax to cover the expense on providing the alternative travel modes.

6.2 Limitations and future research

The paper can be improved in the following aspects. At the data level, the results presented in this paper are based on the assumptions and parameters in Singapore; thus they may not be conclusive for other countries and regions. This is a typical problem in many simulation-based works (Liu et al., 2017; Loeb and Kockelman, 2019). However, this paper is not intended to do a specific forecast, but provides a general analysis framework and presents examples from a case study. Future research can systematically explore the impact of initial settings on the context of different cities, scenarios and assumptions. On the other hand, As Singapore’s demographics and travel and land use patterns are comparable to other big cities in Asia (e.g. Hong Kong, Tokyo). Regardless of setting, the trends and relative trade-offs between different stakeholders are expected to remain same for these cities and regions. Operators from other regions can also extrapolate and obtain inspiration from these results using local data.

At the methods level, several technical improvements can be done in the future. First, the heuristic supply updating algorithm may not converge to the optimal points, which suggests the supply adjustment for AV and PT is not in the most effective way. Future research can apply more advanced algorithms (e.g. reinforcement learning) to replace it. Also, the supply updating can be implemented in a multi-agents learning process, which allows more flexibility for operators’ competing behaviors. Second, we assume the total demand of passengers is fixed during the competition process, which is not realistic in the real world. Future research can introduce a supply-demand interaction module

(Wen et al., 2018) to relax this assumption.

At the case study level, the paper only considers the first-mile scenario in Singapore Tampines area, which may not be general enough compared with the whole city scenario. However, since the framework needs to simulate the agents behaviors for a long time (i.e. 1 year), it would be computationally hard to incorporate the whole urban network. Future research could explore the proposed AV-PT competition framework in larger urban network and incorporate more travel modes by implementing a more efficient simulation platform.

Beyond this paper, future research can also be done in broader areas. The first is to evaluate the impact of pricing, either from operators perspective or from authorities perspective. Price is a powerful tool in market competition. Future research can develop a multi-variables optimization algorithm which allows operators to adjust both fare and supply. This is an extension of current framework that relaxes the fare constraints. From authorities’ perspectives, we can estimate impact of different incentives and subsidies structures on the system. Operators and passengers will response to different pricing policy, leading to different system performance. We can study how government can better design the game mechanism so as to improve service quality and social welfare. The second is to evaluate the impact of technology and technical competence. Generally, AV is more flexible than PT in terms of supply strategies adjustment. We can study the impact of different technical competence of PT on its competitiveness. This can help PT agencies to better understand the trade-off between service stability and flexibility. The third is the comparative study of different AV PT organizational structures. Given many possible relationships in the future (Shen et al., 2018), it is necessary to identify which one is better based on the interests of different stakeholders. Despite there are many previous studies on AV and PT interaction, these studies are based on various contexts and backgrounds. Future research can conduct the comparative studies based on a general (representative) and uniform setting (i.e. demand and land use patterns, infrastructures, and socio-demographics).


This research is supported by National Natural Science Foundation of China (71901164), Natural Science Foundation of Shanghai (19ZR1460700) and the Fundamental Research Funds for the Central Universities (22120180569). The research is also supported by the National Research Foundation, the Prime Minister’s Office of Singapore under the CREATE programme, and the SMART’s Future Urban Mobility IRG.



  • A. S. Ahuja (2007) Development of passenger car equivalents for freeway merging sections. Cited by: §4.
  • J. Alonso-Mora, S. Samaranayake, A. Wallar, E. Frazzoli, and D. Rus (2017) On-demand high-capacity ride-sharing via dynamic trip-vehicle assignment. Proceedings of the National Academy of Sciences 114 (3), pp. 462–467. Cited by: §2.
  • C. L. Azevedo, K. Marczuk, S. Raveau, H. Soh, M. Adnan, K. Basak, H. Loganathan, N. Deshmunkh, D. Lee, E. Frazzoli, et al. (2016) Microsimulation of demand and supply of autonomous mobility on demand. Transportation Research Record: Journal of the Transportation Research Board (2564), pp. 21–30. Cited by: §2.
  • R. Basu, A. Araldo, A. P. Akkinepally, B. H. Nahmias Biran, K. Basak, R. Seshadri, N. Deshmukh, N. Kumar, C. L. Azevedo, and M. Ben-Akiva (2018) Automated mobility-on-demand vs. mass transit: a multi-modal activity-driven agent-based simulation approach. Transportation Research Record, pp. 0361198118758630. Cited by: item 2.
  • T. D. Chen and K. M. Kockelman (2016) Management of a shared autonomous electric vehicle fleet: implications of pricing schemes. Transportation Research Record 2572 (1), pp. 37–46. Cited by: §1.1, §1.1.
  • S. Childress, B. Nichols, B. Charlton, and S. Coe (2015) Using an activity-based model to explore the potential impacts of automated vehicles. Transportation Research Record: Journal of the Transportation Research Board (2493), pp. 99–106. Cited by: §1.1, item 2.
  • Z. Chong, B. Qin, T. Bandyopadhyay, T. Wongpiromsarn, E. Rankin, M. Ang, E. Frazzoli, D. Rus, D. Hsu, and K. Low (2011) Autonomous personal vehicle for the first-and last-mile transportation services. In Cybernetics and Intelligent Systems (CIS), 2011 IEEE 5th International Conference on, pp. 253–260. Cited by: §3.3.2.
  • Á. Costa (1996) The organisation of urban public transport systems in western european metropolitan areas. Transportation Research Part A: Policy and Practice 30 (5), pp. 349–359. Cited by: §1.2.
  • J. de Cea Ch, R. H. Malbran, et al. (2008) Demand responsive urban public transport system design: methodology and application. Transportation Research Part A: Policy and Practice 42 (7), pp. 951–972. Cited by: §2.
  • D. J. Fagnant and K. M. Kockelman (2014) The travel and environmental implications of shared autonomous vehicles, using agent-based model scenarios. Transportation Research Part C: Emerging Technologies 40, pp. 1–13. Cited by: §2.
  • D. J. Fagnant and K. Kockelman (2015) Preparing a nation for autonomous vehicles: opportunities, barriers and policy recommendations. Transportation Research Part A: Policy and Practice 77, pp. 167–181. Cited by: §1.1.
  • S. Galland, L. Knapen, N. Gaud, D. Janssens, O. Lamotte, A. Koukam, G. Wets, et al. (2014) Multi-agent simulation of individual mobility behavior in carpooling. Transportation Research Part C: Emerging Technologies 45, pp. 83–98. Cited by: §3.3.2.
  • Housing & Development Board (2015) External Links: Link Cited by: §3.2.
  • B. Kloostra and M. J. Roorda (2019) Fully autonomous vehicles: analyzing transportation network performance and operating scenarios in the greater toronto area, canada. Transportation Planning and Technology 42 (2), pp. 99–112. Cited by: §2.
  • L. P. Koh and S. A. Wich (2012) Dawn of drone ecology: low-cost autonomous aerial vehicles for conservation. Tropical Conservation Science 5 (2), pp. 121–132. Cited by: §1.1.
  • L. C. Kuang and M. Gopinath (2018) Lessons from bus operations. Cited by: item 2d.
  • J. Lazarus, S. Shaheen, S. E. Young, D. Fagnant, T. Voege, W. Baumgardner, J. Fishelson, and J. S. Lott (2018) Shared automated mobility and public transport. In Road Vehicle Automation 4, pp. 141–161. Cited by: §1.1.
  • M. C. Lesh (2013) Innovative concepts in first-last mile connections to public transportation. In Urban Public Transportation Systems 2013, pp. 63–74. Cited by: §1.3.
  • M. W. Levin and S. D. Boyles (2015) Effects of autonomous vehicle ownership on trip, mode, and route choice. Transportation Research Record: Journal of the Transportation Research Board (2493), pp. 29–38. Cited by: §1.1, §1.1, item 2.
  • X. Liang, G. H. de Almeida Correia, and B. Van Arem (2016) Optimizing the service area and trip selection of an electric automated taxi system used for the last mile of train trips. Transportation Research Part E: Logistics and Transportation Review 93, pp. 115–129. Cited by: §1.1, item 1.
  • J. Liu, K. M. Kockelman, P. M. Boesch, and F. Ciari (2017) Tracking a system of shared autonomous vehicles across the austin, texas network using agent-based simulation. Transportation 44 (6), pp. 1261–1278. Cited by: §1.1, item 2, §6.2.
  • B. Loeb and K. M. Kockelman (2019) Fleet performance and cost evaluation of a shared autonomous electric vehicle (saev) fleet: a case study for austin, texas. Transportation Research Part A: Policy and Practice 121, pp. 374–385. Cited by: §6.2.
  • LTA (2017) Land Transport Authority. External Links: Link Cited by: item 1a.
  • Z. Lu, R. Du, E. Dunham-Jones, H. Park, and J. Crittenden (2017) Data-enabled public preferences inform integration of autonomous vehicles with transit-oriented development in atlanta. Cities 63, pp. 118–127. Cited by: §1.1.
  • K. A. Marczuk, H. S. S. Hong, C. M. L. Azevedo, M. Adnan, S. D. Pendleton, E. Frazzoli, et al. (2015) Autonomous mobility on demand in simmobility: case study of the central business district in singapore. In Cybernetics and Intelligent Systems (CIS) and IEEE Conference on Robotics, Automation and Mechatronics (RAM), 2015 IEEE 7th International Conference on, pp. 167–172. Cited by: §2.
  • L. M. Martinez, G. H. Correia, and J. M. Viegas (2015) An agent-based simulation model to assess the impacts of introducing a shared-taxi system: an application to lisbon (portugal). Journal of Advanced Transportation 49 (3), pp. 475–495. Cited by: §3.3.2.
  • L. M. Martinez and J. M. Viegas (2017) Assessing the impacts of deploying a shared self-driving urban mobility system: an agent-based model applied to the city of lisbon, portugal. International Journal of Transportation Science and Technology 6 (1), pp. 13–27. Cited by: §2.
  • L. M. Mendes, M. R. Bennàssar, and J. Y. Chow (2017) Comparison of light rail streetcar against shared autonomous vehicle fleet for brooklyn–queens connector in new york city. Transportation Research Record 2650 (1), pp. 142–151. Cited by: §1.1, §1.1, item 2.
  • B. Mo, Y. Shen, and J. Zhao (2018) Impact of built environment on first-and last-mile travel mode choice. Transportation research record, pp. 0361198118788423. Cited by: §3, §5.3.
  • F. Nazari, M. Noruzoliaee, and A. K. Mohammadian (2018) Shared versus private mobility: modeling public interest in autonomous vehicles accounting for latent attitudes. Transportation Research Part C: Emerging Technologies 97, pp. 456–477. Cited by: §3.3.2.
  • C. Ruch, L. Sieber, S. Hörl, K. W. Axhausen, and E. Frazzoli (2018) Autonomous mobility-on-demand providing superior public transportation in rural areas. Pergamon. Cited by: item 1.
  • M. Salazar, F. Rossi, M. Schiffer, C. H. Onder, and M. Pavone (2018) On the interaction between autonomous mobility-on-demand and public transportation systems. In 2018 21st International Conference on Intelligent Transportation Systems (ITSC), pp. 2262–2269. Cited by: §1.1, §1.1, item 1.
  • SBS Transit (2017) SBST annual report 2017. Report SBS Transit in Singapore. Cited by: §3.3.3.
  • M. Schreieck, H. Safetli, S. A. Siddiqui, C. Pflügler, M. Wiesche, and H. Krcmar (2016) A matching algorithm for dynamic ridesharing. Transportation Research Procedia 19, pp. 272 – 285. Note: Transforming Urban Mobility. mobil.TUM 2016. International Scientific Conference on Mobility and Transport. Conference Proceedings External Links: ISSN 2352-1465, Document, Link Cited by: §3.3.2.
  • Y. Shen, B. Mo, X. Zhang, and J. Zhao (2019) Built environment and autonomous vehicle mode choice: a first-mile scenario in sinagpore. Technical report Cited by: Appendix A, §3.3.1.
  • Y. Shen, H. Zhang, and J. Zhao (2018) Integrating shared autonomous vehicle in public transportation system: a supply-side simulation of the first-mile service in singapore. Transportation Research Part A: Policy and Practice 113, pp. 125–136. Cited by: §1.1, §1.2, item 1, §3.3.2, §3.3, §5.3, §6.2.
  • SMRT Corporation Ltd (2017) SMRT corporation ltd annual report 2016. Report SMRT Corporation Ltd in Singapore. Cited by: §3.3.3.
  • K. Spieser, K. Treleaven, R. Zhang, E. Frazzoli, D. Morton, and M. Pavone (2014) Toward a systematic approach to the design and evaluation of automated mobility-on-demand systems: a case study in singapore. In Road vehicle automation, pp. 229–245. Cited by: §2.
  • A. Talebpour and H. S. Mahmassani (2016) Influence of connected and autonomous vehicles on traffic flow stability and throughput. Transportation Research Part C: Emerging Technologies 71, pp. 143–163. Cited by: §1.1.
  • A. Vakayil, W. Gruel, and S. Samaranayake (2017) Integrating shared-vehicle mobility-on-demand systems with public transit. Technical report Cited by: item 1.
  • D. M. van de Velde (1999)

    Organisational forms and entrepreneurship in public transport: classifying organisational forms

    Transport policy 6 (3), pp. 147–157. Cited by: §1.2.
  • J. Wen, Y. X. Chen, N. Nassir, and J. Zhao (2018) Transit-oriented autonomous vehicle operation with integrated demand-supply interaction. Transportation Research Part C: Emerging Technologies 97, pp. 216–234. Cited by: §1.1, §1.1, §3.3.1, §6.2.
  • N. H. Wilson, R. W. Weissberg, and J. Hauser (1976) Advanced dial-a-ride algorithms research project. Technical report Cited by: §3.3.2.
  • N. H. Wilson (1991) Organizational options for public transportation in the us. Transportation Planning and Technology 15 (2-4), pp. 405–414. Cited by: §1.2.
  • M. D. Yap, G. Correia, and B. Van Arem (2016) Preferences of travellers for using automated vehicles as last mile public transport of multimodal train trips. Transportation Research Part A: Policy and Practice 94, pp. 1–16. Cited by: item 1.
  • R. Zhang, F. Rossi, and M. Pavone (2016) Routing autonomous vehicles in congested transportation networks: structural properties and coordination algorithms. arXiv preprint arXiv:1603.00939. Cited by: §1.1.
  • S. Zhu and A. Kornhauser (2017) The interplay between fleet size, level-of-service and empty vehicle repositioning strategies in large-scale, shared-ride autonomous taxi mobility-on-demand scenarios. In Transportation Research Board 96th Annual Meeting, Cited by: §2.

Appendix A Mixed Logit Model for Passenger Mode Choice

The mixed logit model is conducted based on an AV preference survey in Singapore Shen et al. (2019). The deterministic part of utility function for individual choosing alternative in choice situation is:


where is the vector of trip specific attributes of mode for individual in situation ; is the vector of sociodemographic variables of individual ; is the alternative spefic constant to estimate the inherent preference of individual on mode ; and are the corresponding coefficients to be estimated.

The results for the mixed logit model is shown in Table 5

, where for each parameter, we estimate both the mean and the standard deviation.

Variables Parameters value t-test
Alternative Specific Constant
Walk Mean fixed at 0
Std. fixed at 0
Bus Mean -0.569 -2.11 **
Std. 0.818 1.89 *
On-demand AV Mean -0.568 -2.56 **
Std. 0.758 3.72 ***
Generalized travel cost
Walk: Walking time (min) Mean -0.363 -28.20 ***
Std. 0.171 22.26 ***
Bus: Travel cost ($SG) Mean -1.14 -8.86 ***
Std. 0.436 0.16
Bus: In-vehicle time (min) Mean -0.212 -12.10 ***
Std. 0.174 9.16 ***
Bus: Waiting time (min) Mean -0.271 -10.27 ***
Std. 0.223 5.31 ***
Bus: Walking time to bus stop (min) Mean -0.214 -10.61 ***
Std. 0.140 4.14 ***
On-demand AV: Travel cost ($SG) Mean -0.984 -18.56 ***
Std. 0.465 13.54 ***
On-demand AV: In-vehicle time (min) Mean -0.195 -11.0 ***
Std. 0.0288 1.16
On-demand AV: Waiting time (min) Mean -0.222 -8.50 ***
Std. 0.0310 0.58
Sociodemographic variables
On-demand AV: household income less
than $SG 4,000 per month
Mean -0.497 -2.81 ***
Std. 0.300 0.62
Statistical summary
Number of individuals 1,242
Number of observations 8,689
Number of random draws 5,000
Initial log-likelihood at zero -10832.448
Final log-likelihood -6581.302
Adjusted McFadden 0.390
  • *: ; **: ; ***: . Std.: standard deviation.

Table 5: Estimation Results of Mixed Logit Model