SAV OPERATIONS ON A BUS LINE CORRIDOR : TRAVEL DEMAND , SERVICE

Before Shared Automated Vehicles (SAVs) can be widely adopted, they are anticipated to be implemented commercially in confined regions or fixed routes where the benefits of automation can be realized. SAVs would be likely to operate in a traditional transit corridor, replacing conventional transit vehicles, and have frequent interactions with other vehicles as well as pedestrians. This paper micro-simulates SAVs’ operation on a 5 mile-corridor to understand how vehicle size and attributes of SAV-based transit affect traffic, transit passengers, and the system cost. The SUMO (Simulation of Urban MObility) package is employed to model microscopic interactions among SAVs, transit passengers, and traffic. Results show that the use of smaller, but more frequent SAVs leads to reduced passenger waiting times but increased total system travel times. More frequent services of smaller SAVs in general do not significantly affect general traffic due to shorter dwell times. Overall, using smaller SAVs instead of the large 40-seat SAVs can reduce system costs by up to 3.1% while also reducing passenger waiting times, under various demand levels and passenger loading factors. However, the use of 5-seat SAVs does not always have the lowest system costs.


INTRODUCTION 1
Automated vehicles (AVs) and shared mobility will fundamentally change the future traffic 2 pattern, by providing cost, environmental, and safety benefits. Shared AVs (SAVs) offer more 3 potential benefits through a lower-cost on-demand service that can be flexible in both schedule 4 and routes. 5 Currently, SAV tests are being performed all over the world, as people try to envision how SAVs 6 should be operated in both the near and far future (Zhao and Malikopoulos, 2019). Over 40 7 corporations are working on AVs (CBinsight, 2019). Waymo (2017) has tested its AVs in 8 Arizona and Texas, and achieved 4 million self-driven miles by November 2017. Before SAVs 9 can run everywhere, they are anticipated to be implemented commercially in a confined region 10 where full automation benefits can be realized  size of a personal owned AV) to 20 (e.g. fixed-route automated shuttle), or even 40 (e.g. 19 automated bus). Smaller SAVs (like 5-seat sedans) are nimbler and easier to park, can accelerate 20 faster, and may cause less congestion and sightline issues. 5-seat sedans can more easily run 21 flexible routes for point-to-point on-demand services without frequent stops. Riders may 22 experience rerouting in 5-seat sedans, but will experience fewer pick-ups and drop-offs than in 23 larger vehicles. Large SAVs usually run fixed-routes and can be more space-efficient (per person-24 mile traveled) but will have to stop more often at stations. The SAV size that is best for transit 25 corridor operations is not only related to the preferences of riders but is also important to the 26 stakeholders. Riders would like to experience less waiting time and onboard time with fewer 27 stops and rerouting, while SAV operators would like to maximize profit or social welfare. 28 Currently, automated shuttles are operating at a low speed (usually less than 30 miles/h) with 29 limited interactions with traffic modes. It is often seen that these SAVs have their dedicated right 30 of way, or share the right of way with pedestrians . In this way, SAVs have more frequent 31 interactions with pedestrians than with other vehicles. However, with the development of 32 automated technology and the sharing economy, SAVs would be likely to operate in a traditional 33 transit corridor, replacing conventional transit vehicles, and have frequent interaction with other 34 vehicles as well as pedestrians. This work micro-simulates SAVs' operation in a 5 mile-corridor 35 setting to understand how traffic reacts to, and how passengers and system costs are affected by 36 vehicle sizes and performance attributes for SAV-based "transit". Singapore that could serve the travel demand while ensuring a desired level of service. Their 44 results showed that an SAV fleet size of about one-third of the total number of passengers was 1 desired in Singapore. However, a 1 SAV per 9.3 conventional vehicle replacement was shown in 2 Fagnant et al.'s (2015) simulation in the Austin area. Recent fleet sizing decision studies mostly 3 assume that an SAV has a maximum occupancy of 4 passengers, however, the capacity of current 4 SAVs used for the tests is more than 4 passengers (Stocker and Shaheen, 2017) and is expected to 5 be as large as 20 or more passengers. 6 Microsimulation noted in this paper refers to the traffic microsimulation where individual driving 7 behavior is tracked, like detailed car following and lane changing maneuvers. However, many 8 studies are using the term microsimulation to define a simulation where agents' information (e.g., 9 route, speed and mode) are tracked. More often, such kinds of simulations are described by 10 researchers as "mesoscopic", given the underlying traffic models are mesoscopic. Vehicles' 11 performance attributes (e.g., acceleration, deceleration, and headway) are not usually tracked. 12 Vehicles' lane changing is also ignored as vehicles are traveling on a link (or roadway).

13
Mesoscopic simulations provide valuable results in regional fleet sizing decisions, mechanisms of 14 ridesharing or even dynamic ridesharing, and traffic patterns under dynamic traffic assignment, 15 but Detailed car following and lane changing models raise the time burden in microscopic 38 simulations, compared with mesoscopic and macroscopic simulations. This leads to a confined 39 simulation area with a low share of realistic travel demand in current microsimulation studies 40 . Vehicle data at the trajectory level can only be obtained through micro-41 simulation, which provides more accurate vehicle movement and energy information. transit structure and demand characteristics. An agent-based supply-side simulation was built to 2 assess the performance of the proposed service with different fleet sizes and ridesharing 3 preferences in Singapore's 12 km2 area during morning peak hours from 7 am to 9 am. Authors 4 showed that the integrated system has the potential of enhancing service quality, occupying fewer 5 road resources, being financially sustainable, and utilizing bus services more efficiently. Wen solving the upper-level problem using a nonlinear programming solver and solving the lower-17 level problem using an iterative agent-based assignment-simulation approach. Two bus types and 18 15 train types were simulated, and four modes are involved: walk, transit, SAVs, and SAVs + 19 transit. Results indicate significant traveler benefits, in terms of improved average traveler 20 waiting times compared to the initial transit network design. 21 Overall, there has been extensive research, no matter the kind of scope, dedicated to 22 understanding the future travel pattern with automation technology. The focus is on the desired 23 vehicle fleet size to meet travel demand, considering the simple link model and car-following 24 model without lane changing. Some consider traffic assignment, which involves the travel 25 behavior of user equilibrium and the integrated system of SAV and transit, but traffic modelling 26 is still simplified. Microsimulation has also been investigated, but vehicles' performance 27 attributes and interactions between SAVs and traffic at the microscopic level have not been 28 studied enough in relation to vehicle size decision. Therefore, this work leverages the Simulation 29 of Urban MObility (SUMO) software (Krajzewicz et al., 2012) to simulate the relationships 30 among vehicles sizes, service frequencies, and travel demand, by considering SAVs serving a 31 transit corridor, to provide insight on how vehicle sizes would impact transit passenger, traffic, 32 and system costs. 33

SUMO simulation 35
SUMO software is a powerful tool used to simulate multimodal transportation, as it has 36 advantages in micro-simulating interactions among different modes. For example, it can simulate 37 the accurate process of transit access and egress, as well as riders getting on and off the transit. 38 Such detailed manipulations are achieved through TraCI (Traffic Control Interface), a toolkit in 39 SUMO that allows users to retrieve real-time values of simulated objects and to manipulate their 40 behavior "on-line" through Python scripts. 41 SUMO simulation starts with the input of travel demand and network information. Network 42 information includes all the roadways, links, junctions with signals, and transit platforms. The 43 simulation network setup in this study is a 5-mile, 2-lane, straight corridor with traffic signals ( Figure 1). This corridor has a lane width of 3.5 m and a speed limit of 30 mph based on 1 recommended designing practice from the American Public Transportation Association (Barr et  2 al., 2010). SAV stations (or bus stops) are evenly placed (about every quarter mile) along the 3 corridor (Walker, 2012). Each bus stop is ten meters long. SAVs and conventional vehicles 4 (background flow) can travel in both lanes. As the stations are curbside, when SAVs are serving 5 passengers at stations, they will obstruct the vehicles behind. Scenarios that have parking bays for 6 SAVs are also tested. 7 SAVs are inserted into the corridor to serve riders with a fixed schedule (i.e. a fixed frequency) 8 and run a fixed route to the end of the corridor, with stops at the stations. After an SAV completes 9 its journey, it goes back to the starting point of the corridor, assuming that it does not have wait 10 time at the depot. A Poisson distributed background flow of conventional cars with a mean 11 departure rate is assumed. The base case scenario will test an arrival rate of 0.7 vehicles per 12 second (approximately 1260 vphpl), to reflect a common flow rate on a 30-mph corridor (Barr et  13 al., 2010).With a high dispatching rate (frequency) of SAVs, "bus bunching" can be observed, as 14 SAVs can overtake previously dispatched SAVs. 15 Travel demand is determined by SAV riders/passengers, who walk on the road, wait for SAVs 16 and ride in SAVs. Other active modes are ignored here because they would not affect the 17 operations of SAVs or conventional vehicles, although there may be taxis, which will stop and 18 pick-up/drop-off passengers, and scooters/bicycles, which potentially slows down the traffic.

19
Passengers are uniformly generated at random along the corridor, arriving with a uniform 20 distribution in a 3-hr peak time period. Since the distance between two consecutive stations is 1/4 21 miles, passengers who have origins and destinations between two consecutive stations probably 22 give up taking transit. Therefore, only those passengers who have a trip distance longer than 1/3 23 miles are randomly generated along the corridor. Riders (bus line users) walk to the nearest 24 station, get on the next available SAV and get off at the station closest to his/her destination. An 25 SAV that has not reached its capacity will stop at a station where new riders are waiting or where 26 current riders want to alight. Further, SAVs wait for passengers running towards the bus, if the 27 bus is stopped at a station and the running passenger can catch the SAV in 10 seconds. After a 28 rider gets off the SAV, he or she will walk to their destination location and then disappear from 29 the simulator. including the edge length, as well as the length and position of the bus stop. Travel demand is 5 then processed to determine the start point of the journey, departure station, arrival station and 6 destination point of the itinerary. After that, simulation in SUMO starts and the itinerary of riders 7 is processed by the simulator at time 0. Beginning at the first step of simulation, background flow 8 comes out on the corridor and every few minutes an SAV departs from the start point of the 9 corridor. During every step of the simulation, which is every 0.5 seconds, the status of riders and 10 SAVs are simulated and tracked. Riders' status is tracked so that SAVs can react based on riders' 11 information (e.g. whether riders are still walking to the bus station, waiting at the bus station, or 12 already on board). The status and locations of SAVs are tracked, to determine whether the SAVs 13 need to stop at the next station. Basically, the simulation checks whether riders need to get off at 14 the next station. If no one is getting off, it then checks whether passengers are waiting at the next 15 station and if there are available seats on board. When the SAV is parking at the bus stop, it keeps 16 checking whether a rider can catch the SAV in 10 seconds and will then let those people get on 17 the SAV. After the SAV leaves the station, it sets stops for all new riders' destination stations. 18 SAVs and riders' status and locations are checked every timestep until the simulation reaches the 19 time horizon. The output includes vehicles' and riders' travel time and waiting time, and riders' 20 walking distance and riding distance. 21

Simulation parameters
Stocker and Shaheen (2017) have envisioned four types of potential SAV and service models: 1 micro-vehicles (1 or 2 passengers), small vehicles (3-7 passengers) mid-sized vehicles (7-20 2 passengers) and large vehicles (20+ passengers). However, one can imagine that the future world 3 may also have a large capacity SAVan automated bus to serve a heavy demand transit corridor, 4 but such a corridor probably does not allow micro-vehicles to travel. Therefore, for the simplicity 5 of vehicle sizes, four types of vehicles are simulated, from a normal sedan size of 5 seats (no 6 driver due to full automation) to an automated bus of 40 seats. 7 As shown in Table 1, background flow uses the SUMO default value for "passenger" vehicle 10 types (Krajzewicz et al., 2012). Although there would be differences in the lane-changing model 11 between conventional vehicles, this study focuses on longitudinal effects instead of lateral effects. 12 Therefore, the lane-changing model is assumed to be LC2013, the default from SUMO (Erdmann, 13 2015). The LC 2013 model also provides flexibility in setting strategic, cooperative, tactical and 14 regulatory lane changes (Erdmann, 2015). 15 Automated bus or shuttle tests may proceed with caution at early implementation stages due to 16 the unreliable and unstable camera recognition and slow data processing, however, in the future, 17 automated buses will probably have faster speeds than human-driven vehicles (Litman, 2017 3.3 seconds/passenger using a contactless card or 4 seconds/passenger using the magnetic strip 5 when the only front door is used for boarding and both doors are used for alighting. The current 6 tested automated shuttles have one door for both boarding and alighting, but the door is wider 7 than the mid-sized bus. Here 3.5 seconds (considering 0.5-second simulation timestep) is used for 8 the average boarding and alighting time, although there could be variations due to the vehicle 9 design and payment method. 10

Scenario design 11
This paper tests a base case scenario that has riders' demand varying from 100 riders per hour to 12 600 riders per hour. Background flow is set to 1260 vehicles per hour per lane, with no parking 13 bay at the station and no traffic signals. Based on the demand and capacity, the headway of SAVs 14 is set for each level of demand such that the total demand can be met considering the total 15 available seats of dispatched SAVs and an average loading factor of SAVs, recognizing that there 16 is waiting time for riders at the station. Table 3 shows the headway of SAVs when assuming that 17 SAVs are always full (load factor = 1), while the base case assumes a load factor of 0.7 for all 18 types of SAVs. For example, when assuming SAVs are always full, 0.5 min headway of SAV 19 (120 SAVs/hr) is required to meet the demand of 120 × vehicle size × load factor = 120 × 5 × 1 = 20 600 persons/hr. In this case, a high load factor indicates a low frequency. 21 Other than the base case scenarios, a few scenarios have been generated for comparison, as 23 shown in Table 4. These scenarios include varying the background flow (varying from 0.4 to 0.8 24 vehicles per second by Poisson distribution), adding station bay and traffic signals, varying the 25 SAV headways (via various assumed load factor), and different value of travel times (VOTTs).

26
Since VOTT is considered for both background travelers as well as SAV riders, background 27 travelers' VOTT is tested, with SAV riders' value of walking, riding, and waiting time changing 28 accordingly. Background flow scenarios aim to test how the SAV fleet and the system perform 29 under different congestion conditions led by the background vehicles. This could reflect the 30 optimal SAV size when the corridor is under different levels of services. Load factor scenarios 31 are used to investigate operators' decisions in the frequency of dispatching SAVs to serve various 32 demand levels considering the total system cost. The station bay scenario will be able to present 33 the case when there is a potential to obtain the right of way for SAVs parking at stations without 34 interrupting the background flow. Last, the traffic signal scenario will show the case when the 1 background flow cannot flow freely, which is more likely to happen in a real transit corridor.

Evaluation metrics 13
The average vehicle travel time, average person/passenger waiting time, total system travel time 14 and total system cost will be evaluated for each scenario. The average vehicle travel time 15 (background vehicles) can evaluate how congested the traffic is, while the average person waiting 16 time shows the efficiency of the transit system. The total system travel time, considering travel 17 times of background vehicles and SAVs, can show the overall system performance and the total 18 system cost will evaluate the total cost to serve travel demand under different types of vehicles 19 and levels of service. The total system cost considers the travel time cost and operating cost of 20 SAVs and background vehicles. Table 5 shows the details components of the total system cost.

21
When riding in an SAV, riders are assumed to perceive a VOTT that is half that of those driving 22 vehicles. However, when they are walking to and from SAV stop locations or waiting for an SAV 23 to arrive, their VOTT is assumed to be double that of a driver (Liu et al., 2017 In this section, the analysis of three scenarios are presented: the base case scenario, adding SAV 4 station bays, and adding traffic signals. Each scenario performed five runs and the average value 5 is shown in the results. Figure 3 shows the results of the base case scenario. For each SAV size, 6 the average background vehicle travel time increases with increasing passenger demand (see 7 Figure 3a). This is because SAVs stop more frequently, affecting background traffic, to 8 accommodate greater passenger demand. In this base case scenario, smaller SAVs in general do 9 not significantly affect background traffic, except for the combination of the 5-seat SAVs and 10 highest passenger demand of 600 persons/hr when the SAV flow is substantial. While smaller 11 SAVs mean more frequent services to serve the same number of passengers, dwell times at each 12 stop and thus the durations of affecting traffic would be shorter. Figure 3b shows an expected 13 result: SAVs with more frequent services reduce persons/passenger waiting times at all passenger 14 demand levels. Waiting time discrepancies between different SAV sizes shinks under a higher 15 demand level, due to a higher frequency of all types of vehicles. For each SAV size, the total 16 system travel time, considering both background vehicles and SAVs, increases with greater 17 passenger demand (in Figure 3c). This is consistent with the increase in the background vehicle 18 travel time when passenger demand is greater. However, the effect of SAV sizes on total system 19 travel time is evident as smaller SAVs are associated with higher total system travel times, 20 particularly with high passenger demand. It is noted that the flow of SAVs is quite substantial 21 with high passenger demand. While using smaller SAVs reduces passenger waiting time, it can 22 increase total system travel time. Thus, the smallest size of SAVs is not necessarily the optimum 23 size. Figure 3d illustrates the results of total system costs where the 40-seat SAVs are 24 outperformed by smaller SAVs at all passenger demand levels. Compared to the 40-seat SAVs, 25 using smaller SAVs would reduce the system cost by up to 1.9%. While the 5-seat and 10-seat 26 SAVs have lower total costs with the demand of 300 persons/hr or less, the 10-seat and 20-seat 27 SAVs have lower total costs with the demand of 400 persons/hr or more. SAVs, the average vehicle travel time lightly increases with greater passenger demand, which is 4 less than 15 seconds. This slight increase would be attributed to the noticeable flow of 5-seat 5 SAVs (i.e. headway of 0.35 minutes when passenger demand is 600 persons per hr). Like the base 6 case scenario, smaller SAVs reduce passenger waiting time, but increase total system travel time. 7 Compared to the base case scenario, the total system travel time in the SAV station bay scenario 8 is lower (up to approximately 4% lower when the passenger demand is 600 persons/hr). This 9 demonstrates the benefits of providing station bays. Figure 4d suggests using smaller-size SAVs 10 instead of 40-seat SAVs can reduce the system cost by up to 1.4%, which is consistent with the 11 base case scenario. 10-seat SAVs tend to be more cost-efficient under different levels of travel 12 demand. Regarding the improvement in total system travel time compared to the base case 13 scenario, little reduction in cost due to station bay is observed when travel demand is low, and the 14 total system cost falls up to 1.8% in the 600-transit-users-per-hour scenario during the 3-hour 15 morning peak. It is worth noticing that such cost savings are under the situation when a bay is 16 built for each of the stations in this 5-mile corridor, which will add construction costs and 17 potentially right-of-way costs. In a real network, there is a need for cost-benefit analysis across a 18 long evaluation time horizon and for a more specific area (e.g., with congested intersections, or 19 higher SAV ridership, leading to longer stopping times at stations).  all SAV sizes under various travel demands is substantially higher compared to the previous 4 scenarios due to delays at traffic signals. The trend of the average vehicle travel time remains the 5 same, but the increasing trend for 5-seat SAVs is more obvious. When passenger demand 6 increases from 100 persons/hr to 600 persons/hr, using 5-seat SAVs increases the average vehicle 7 travel time by nearly 10%. This contributes to the higher system costs of 5-seat SAVs when 8 compared to 40-seat SAVs with the demand of 600 persons/hr. Other than that, 40-seat SAVs 9 tend to have higher system costs compared to smaller SAVs. It is noted that when the demand is 10 300 persons or less, the 5-seat SAVs still perform well in terms of system costs, and 10-seat 11 SAVs are the most cost-efficient for a demand less than 500 persons/hr. case scenario. Figure 6 shows total system costs with background flow rates of 720, 900, 1080, 4 and 1440 vphpl, which have a similar pattern compared to the results of a background flow rate of 5 1260 vphpl in Figure 3d. That is, total system costs increase almost linearly with increasing 6 passenger demand and the system with 40-seat SAVs has higher costs. While the system with 5-7 seat SAVs performs well when passenger demand is low, its performance decreases compared to 8 the mid-size SAVs (10 and 20 seats) when passenger demand is larger than 300 persons/hr, owing 9 to the higher frequency of 5-seat SAVs. These results suggest the adoption of smaller-size SAVs, 10 rather than large 40-seat SAVs, would reduce the system costs, by up to 2.7%, in both low and 11 high passenger and background traffic demand levels. Of course, a higher background flow will 12 lead to higher system costs, due to a higher flow of traffic and thus more vehicle-mile costs. total system costs when the background flow is 720 vphpl. Figure 7a shows the total system cost 3 when the headway is half of the value shown in Table 3, under each SAV size and each level of 4 travel demand. The total system cost of the 5-seat SAV system is greater than that of the 40-seat 5 SAV system when the demand is 600 persons/hr. When varying the headway of service, the 6 systems with 10-seat and 20-seat SAVs consistently have lower costs compared to those with the 7 40-seat SAVs, favoring 10-seat SAVs at low demand and 20-seat SAVs at high demand. The 8 total cost is generally stable across these scenarios, but the benefits of using smaller-size SAVs 9 instead of 40-seat SAVs tend to be greater with increasing loading factor (decreasing frequency).

10
For example, cost reductions are between 0.7% and 2.3% when the loading factor is 0.6, and 11 between 1.7% and 3.1% when the loading factor is 0.8. persons/hr, because the AVO is robust to the travel demand, probably due to the fixed 5 relationship between SAV dispatching headway and the travel demand. With a higher frequency 6 (low loading factor), AVO trends to increase, but large-size vehicles witness a large increment 7 compared with small-size vehicles. When load factor increases from 0.5 to 0.9, AVO of 5-seat 8 SAVs slightly goes up from 0.9 to 1.6, but AVO of large size SAVs raises from 6.9 to 12.7. 9 However, the percentages in AVO are stable for all sizes of vehicles, from 17% to 31%, when the 10 load factor climbs up from 0.5 to 0.9. The AVO percentages also align with statistics from current 11 studies (FTA, 2016). riding and waiting time also changed based on the assumption in Table 5. The total system cost 8 increases linearly with the increase of VOTT, due to the linear function in calculating the cost of 9 background drivers and SAV riders. With a higher VOTT, the discrepancies in total system cost 10 between each level of demand also increase. This can be explained by the added cost when more 11 riders perceive higher VOTT. Although the results are straightforward, it should be noted that 12 VOTT also impacts the other mode choice of road users, not only personal drivers and SAV 13 riders. Heterogeneity in road users' VOTT could also exist. However, this is beyond the 1 discussion of this study, but could lead to more practical results. This study investigated the performance of an SAV-based bus transit corridor, where different 5 sizes of SAVs replace conventional bus transit vehicles. SUMO was used to simulate microscopic 6 interactions between SAVs and background traffic and between SAVs and transit passengers, 7 under various background traffic conditions, SAV sizes and associated characteristics, passenger 8 demand levels, and loading factors. Different configurations of the 5-mile bus transit corridor 9 were considered, including non-signalized corridor, signalized corridor, and corridor with SAV 10 stations in bays. Detailed bus behaviors were incorporated, including waiting for approaching 11 riders, and skipping stops when the vehicle capacity is reached. 12 Simulation results show that the use of smaller, but more frequent SAVs leads to reductions in 13 passenger waiting times but increases in total system travel times. It is found that more frequent 14 services of smaller SAVs in general do not significantly affect background traffic given their 15 shorter dwell times at stations. There are few exceptions, such as in traffic signal scenarios with 16 5-seat SAVs and high passenger demand, where the substantial flow of 5-seat SAVs negatively 17 affects background vehicle travel times. Results highlight that the systems with 10-seat or 20-seat 18 SAVs have lower costs than those with 40-seat SAVs, consistently across various scenarios. 19 While the system with 5-seat SAVs has relatively low costs at low passenger demand, its 20 requirement of high SAV frequencies at high passenger demand can increase system costs 21 substantially. Indeed, the cost of the 5-seat SAV system can exceed that of the 40-seat SAV 22 system in high passenger demand scenarios when there are traffic signals, or the loading factor is 23 low. Overall, using smaller SAVs instead of the large 40-seat SAVs can reduce system costs by 24 up to 3.1% while improving transit passenger experience with reduced waiting times. Although 25 conventional bus transit scenarios, usually with larger vehicles are not simulated in this study, 26 their system costs and passenger waiting times would be higher than the 40-seat SAV scenarios. 27 Thus, replacing conventional bus transit vehicles with SAVs of smaller sizes would offer greater 28 reductions in system costs and passenger waiting times. Results also suggest that the smallest 29 SAVs are not always the optimum solutions, right-sized SAVs and associated frequencies should 30 be considered based on passenger demand, network configuration, and loading factors. 31 However, limitations of the micro-simulation in this study still exist. The relationship between 32 headway and demand is assumed to be fixed, based on the assumed load factor of SAVs.

33
Optimization techniques could be utilized to find the best headway as well as vehicle size for the 34 most cost-efficient system, but these techniques would not be easy to integrate. On the other 35 hand, considering the complex behavior of buses waiting for approaching riders, and skipping 36 stops when the seats are full is much easier to integrate. It is also not clear whether future 40-seat 37 SAVs would be able to provide standing area, in which case the capacity of the vehicle would be 38 more than 40 seats. This study assumes that the capacity of a 40-seat SAV is 40 riders, but 40-39 seat SAVs have the potential to be favored by the operators if standing is allowed on board. 40 It should be acknowledged that in the micro-simulation, background traffic is simulated with 41 typical driving behaviors. Future research could explore the impacts of SAV-based transit when 42 background traffic is also partly or fully automated by considering different AV penetration rates.

43
For a high frequency bus corridor, platooning of SAV-based transit vehicles should also be 44