demand responsive transit simulation of wayne county, michigan · kagho, hensle, balac, freedman,...

Kagho, Hensle, Balac, Freedman, Twumasi-Boakye, Broaddus, Fishelson, Axhausen

Demand Responsive Transit Simulation of Wayne County, Michigan 1 2

Grace O. Kagho 3 Institute for Transport Planning and Systems (IVT) 4

ETH Zurich, Stefano-Franscini-Platz 5, 8093, Zurich, CH, Switzerland 5

E-mail: [email protected] 6

7

David Hensle 8 Analyst 9

RSG 10

1515 SW 5th Ave #1030, Portland, OR 97201 11

Email: [email protected] 12

13

Milos Balac 14 Institute for Transport Planning and Systems (IVT) 15



18

Joel Freedman 19 Senior Director 20

RSG 21

1515 SW 5th Ave #1030, Portland, OR 97201 22


24

Richard Twumasi-Boakye 25 Research Scientist 26

Ford Motor Company 27

2101 Village Rd., Dearborn, MI. 48124 28


30

Andrea Broaddus 31 Senior Research Scientist 32

Research & Advanced Engineering, Mobility & Robotics Department 33

Ford Greenfield Labs, Palo Alto, California, 94303 34


ORCID: 0000-0003-3175-5986 36

37

James Fishelson 38 Supervisor, Mobility Research 39

Ford Motor Company 40

2101 Village Rd., Dearborn, MI. 48124 41


43

Kay W Axhausen 44 Institute for Transport Planning and Systems (IVT) 45



48

49

50 Word Count: 6746 + 3 tables = 7496 words 51

mailto:[email protected]


2

1

2

Submitted: August 1, 2020 3


3

ABSTRACT 1

Demand Responsive Transit (DRT) can provide an alternative to private cars and complement existing 2

public transport services. DRT’s potential is enhanced by the advent of automation; such services are often 3

referred to as shared autonomous vehicles (SAVs). However, the successful implementation of DRT 4

services remains a challenge; as both researchers and policy makers can struggle to determine what sorts 5

of places or cities are suitable for it. Research into car-dependent cities with poor transit accessibility are 6

sparse. In this study, we address this problem, investigating the potential of DRT service in Wayne County, 7

USA, whose dominant travel mode is a private car. Using an agent-based approach, we simulate DRT as a 8

new mobility option for this region, thereby providing insights on its impact on operational, user, and 9

system-level performance indicators. We test DRT scenarios for different fleet sizes, vehicle occupancy, 10

and cost policies. The results show that a DRT service in Wayne County has certain potentials, especially 11

to increase the mobility of lower-income individuals. However, we also show that introducing the service 12

may slightly increase the overall VKT. Specific changes in service characteristics, like service area, pricing 13

structure, or preemptive relocation of vehicles, might be needed to fully realize the potential of pooling 14

riders in the proposed DRT service. We hope that this study serves as a starting point for understanding the 15

impacts and potential benefits of DRT in Wayne County and similar low-density and car-dependent urban 16

areas, as well as the service parameters needed for its successful implementation. 17

18

Keywords: Demand Responsive Transit (DRT), Shared Mobility, Agent-based Models, Shared 19

Autonomous Vehicles 20

INTRODUCTION 21

Mobility is important for ensuring people have access to daily needs. Today, the necessity to move, 22

coupled with population growth and economic development in urban areas have led to increased congestion. 23

Clearly, we need more sustainable transport options to move people safely through cities. Fixed-route and 24

timetable-based public transit (PT) present an effective solution in areas with high demand and well-utilized 25

corridors; however, many locations – including cities, report underperforming public transit services, not 26

sufficient for serving all travelers due to low and sparse demand (1). This leaves private cars as the 27

predominant alternative for commuters. 28

While not novel, in recent years, shared mobility has become a viable transportation option. This 29

is partly due to the diffusion of information and communication technologies used in developing systems 30

for requesting trips and making payments within a single software platform. In this paper, we focus on 31

Demand Responsive Transit (DRT). DRTs fall under the broader category of shared mobility and comprise 32

services such as, taxis, paratransit, microtransit, etc. They refer to a type of quasi-public transport that 33

allows vehicles to modify their routes based on service demand (2). Operationally different from fixed route 34

public transit, DRTs permit vehicles to pick-up and drop-off passengers at locations of their choice (3). 35

When carefully designed, DRTs can complement existing public transit. However, there remains the need 36

to investigate how these services will perform amid existing transport modes to improve mobility. Few 37

studies focus on understanding this need, particularly in areas with low PT utilization and significant 38

socioeconomic disparities, such as Wayne County, Michigan. 39

Private cars are the dominant mode of travel in the City of Detroit and Wayne County, with PT 40

accounting for less than 1% of total trips. This is largely because of high auto ownership (~95% of 41

households have access to at least one auto). Additionally, transit access is limited and inconvenient for trip 42

making. This presents a problem for many residents who may struggle to afford high auto insurance rates 43

(4). Wayne County is home to the “big three” U.S. auto makers, Ford, Chevrolet, and General Motors. The 44

local stakeholders have been discussing possibilities for expanding transit and integrating new on-demand 45

mobility options as they anticipate socioeconomic development in the area due to auto industry employment 46

growth and future autonomous vehicle production. Persistent socioeconomic disparities (5) as well as low 47


4

PT ridership make Wayne County a unique location to model DRTs as a possible solution for improving 1

mobility, and will provide valuable insights to transportation agencies and researchers on DRT operations 2

in similar cities. 3

For this reason, we model and simulate a hypothetical DRT service in Wayne County, Michigan. 4

In this context, we define DRT as a shared fleet of vehicles with operational tolerance for pooling, and with 5

travelers picked-up and dropped-off at their desired locations.. This paper contributes to the state of art by 6

developing a novel computational schema to convert a trip-based travel demand model into inputs for 7

developing a calibrated agent-based model in MATSim, an open-source mobility simulation platform with 8

an integrated DRT module. This required a further step of developing and calibrating a mode choice model 9

to estimate demand for the DRT service. 10

The objective of this paper is to understand the demand potential of DRT for Wayne County based 11

on fleet size, cost and vehicle capacity factors. We ran a set of DRT scenarios with varying levels of these 12

factors as a method of demand estimation and to understand the impact of DRT on operational, user, and 13

system-level performance indicators. For the effectiveness of the designed DRT, we try to answer the 14

following questions: What is the demand for the new service and how will this affect fleet size and vehicle 15

utilization? How does DRT fare affect demand? How do service-design parameters affect user experience 16

in terms of wait time and total trip time due to detour allowances? How will the DRT service impact 17

mobility in Wayne County in terms of system-level vehicle kilometers travelled (VKT)? 18

The remainder of this paper is organized as follows. Section 1 provides background context and 19

review of pertinent literature regarding agent-based modelling and DRT. Section 2 describes the research 20

methodology, demand and supply models, and scenario design for the integrated DRT module. Section 3 21

describes the results, and Section 4 provides a discussion of the research findings. Finally, we present the 22

conclusion of this paper and highlight areas for future work. 23

24

BACKGROUND 25

The past decade has seen an explosion in DRT and related shared automated vehicle (SAV) 26

research and modelling, where vehicles serve passenger demand with both spatial and temporal flexibility 27

as opposed to the fixed routes and schedules of traditional transit (6). An overall review of emerging 28

mobility on demand’s operational concepts is given by Shaheen et al. (7), and Narayanan (8) provides a 29

more focused review of different studies on SAVs. 30

The most general approach comes from “aggregate models,” which use combinations of raw data, 31

assumptions, and equations or assumed relationships, to deterministically estimate system-level 32

performance metrics: costs, time, and more. For example, Greenblatt and Saxena, (8) use existing taxi data 33

as a starting point, adding assumptions about the performance of electric SAVs to estimate the national 34

effects on greenhouse gas emissions. Zachariah et al take a more detailed and hybrid approach, using a 35

network assignment model to develop a time-based trip schedule for all passengers (9). More detailed are 36

“Network Assignment Models,” where traditional travel demand modeling tools (e.g. TransCAD) are used 37

with DRT added as an existing mode (10). They are strong at optimization, such to test different 38

optimization strategies to relocate vehicles when not in use (11,12). However, being macroscopic in scope, 39

these assignment models lack the fidelity to model the actual performance of a DRT system, such as 40

independent pick-ups and drop-offs. 41

The third and most detailed group, “Agent-Based Models” (ABMs) directly model the behavior of 42

SAVs as individual agents in a DRT simulation environment, including at a minimum passenger pick-up, 43

drive time, drop-off, and behavior while empty. Various sharing and relocation algorithmic approaches are 44

often also simulated. These models have often been used to investigate the theoretical question of how 45

many SAVs would be required to achieve the same level of mobility as private vehicles, finding 46


5

replacement ratios ranging from approximately 2-40, (i.e. private vehicles replaced by one SAV) (9). This 1

level of detail makes ABMs the most flexible of the model groups, useful for testing a range of hypothetical 2

situations, such as different service types (13), service areas (14), approaches to ride pooling (15-17), 3

vehicle relocation and staging strategies (18-20), impacts of traffic assignment (21), and cost impacts of 4

different scenarios (22). Note that many researchers have a combined approach, utilizing a network 5

assignment model to first generate and assign demand to yield a spatial origin-destination matrix and 6

roadway travel speed skims, and then employ an ABM to estimate vehicle movements and passenger 7

interactions (11,23,24). Of the ABMs, MATSim is one of the most popular modeling tools, having the 8

benefit of being both open-source and activity based. 9

Getting good estimates of travel demand is an omnipresent challenge; one of the reasons that so 10

many studies focus on New York, Singapore, and Austin is that they have publicly available taxi or TNC 11

data. No studies to date have specifically focused on overall DRT performance in a city like Detroit, which 12

is characterized by low-density, car-dependent urban structure, and relatively low levels of traffic 13

congestion. Similarly, relatively few studies have tried to endogenously model mode choice within an 14

ABM. The network models of Chen et al., Childress et al., and Gucwa (15,19,25) model mode choice 15

exogenously as part of the traditional four-step transportation forecasting process, using multinomial logit 16

models considering monetary costs, wait times, and in-vehicle travel times. Liu et al., (20) extend an 17

existing ABM model of Austin (19) to consider varying per-km costs jointly with mode choice. 18

Other researchers consider mode choice endogenously in their models, as done in this paper. For 19

example, Hörl et al. (26) treats mode choice directly as part of their ABM, with fully specified mode choice 20

utility functions, showing the potential for automated taxis to serve up to 60% of all trips in their 21

hypothetical Sioux City network. Azevedo et al. (27) models the choice between private vehicles, public 22

transit, and SAVs, and argues that factors affecting mode choice could also affect auto purchase decisions, 23

though they do not explicitly include this possibility in the model. Zhang (28) shows that SAVs could take 24

mode share from transit, especially for shorter trips with low in-vehicle travel times. Moreno et al., (29) 25

examines how the shift to DRT could affect overall number of trips and distance traveled. 26

METHODOLOGY 27

28

Overview of MATSim 29

MATSim (30) is an extendable, multi-agent simulation framework implemented in Java. It 30

simulates travelers’ – referred to as agents – activities and trips on a network for an entire day. Agents in 31

MATSim have daily plans, which consist of their activity chains and socioeconomic information. Each 32

activity contains certain information such as activity location, start, and end time. In a typical application, 33

MATSim operates in an iterated loop until it achieves user equilibrium. It uses a co-evolutionary 34

algorithm where each agent optimizes its individual plan until the system converges to a stable state. 35

In this novel application, we combine demand from an existing, calibrated trip-based model for 36

Southeast Michigan with the trip assignment capabilities of MATSim. Since we are modeling individual 37

trips made by agents, rather than agents’ daily activity plans, the MATSim utility scoring mechanism 38

which adjusts activity plans is not used to achieve system equilibrium. Only route choice is optimized to 39

achieve convergence, given fixed origins, destinations, and departure times. We integrate a mode choice 40

model to predict choice of private auto, public transport, walk, bike, and DRT modes. Below we discuss 41

the model network and demand components in more detail. 42

Network Creation 43

All roads inside Wayne County, Michigan were extracted from OpenStreetMap to form the basis 44

of the MATSim network. To reduce computational complexity, the road network was thinned outside of 45


6

the downtown Detroit area to match the planning network used in the Southeast Michigan Council of 1

Governments (SEMCOG) E-7 trip-based model (31). Routing issues arising from the network thinning 2

procedure were manually corrected. 3

MATSim requires links to be only unidirectional, so, two-way links from OpenStreetMap were 4

duplicated to create two identical links, but with the start and end nodes switched. Link capacities and free-5

flow speeds (when not present in the OpenStreetMap network) were calculated based on the SEMCOG E-6

7 model network. 7

Transit routes were created from General Transit Feed Specification (GTFS) data, where available, 8

or coded manually. Transit operators in the model include the Suburban Mobility Authority for Regional 9

Transportation (SMART) and Detroit Department of Transportation (DDOT) buses servicing downtown 10

and surrounding suburbs, and the QLINE streetcar and Detroit People Mover automated light rail system 11

servicing the central business district. We used the pt2matsim software package (32) to combine GTFS 12

data for these services and to create additional links in cases where routes extended beyond the Wayne 13

County network. This package also created a public transportation vehicle list and schedules. 14

15

Figure 1 MATSim network of Wayne county 16

Travel Demand 17

The SEMCOG E-7 trip-based model was used as the base travel demand for this work. It contains 18

more than 20 million person trips across six counties, 2899 travel analysis zones, 8 trip purposes, and 15 19

trip modes. Transforming the E-7 trips into MATSim “agents” required two steps. First, SEMCOG E-7 20

trip tables were disaggregated into individual trips that could equate to MATSim agents. Secondly, since 21

the SEMCOG model encompasses a much larger area than the modeling area of Wayne County used for 22

this project, the trips/agents were filtered to select only those that traversed our network for inclusion in the 23

DRT simulation. 24


7

For the first step, Production-Attraction (PA) matrices from the E-7 mode choice model were first 1

converted into OD format. The E-7 mode choice model creates Production-Attraction (PA) matrices by 2

income group, trip purpose, and peak or off-peak travel period. Each PA matrix contains a core for each of 3

the 15 available trip modes in the E-7 model. PA matrices were converted into Origin-Destination (OD) 4

format using the PA to OD factors specified in the SEMCOG E-7 model. PA to OD factors are dependent 5

on trip purpose, time of day, and direction. Peak period PA matrices were split into AM and PM whereas 6

the off-peak matrices were split into midday, evening, and night for a total of five time-of-day periods. 7

After conversion to OD format, the shared ride 2 and shared ride 3+ trip tables were divided by 2 8

and 3.5, respectively, to convert from person trips to single vehicle MATSim agents. All cells in the OD 9

matrix were then converted to an integer based on Monte Carlo sampling. Matrices for commercial vehicle 10

trips already in OD format were also integerized and included. These commercial vehicles, which represent 11

light, medium, and heavy trucks are included in the simulation as the background traffic because they 12

contribute to network congestion. 13

For the second step, to select which trips entered or exited the region, we coded and then skimmed 14

unique link identifiers to determine whether a trip crossed the model boundary, and if so, over which link 15

and in which direction. Trips that had the entirety of their journey outside of the modeling area were 16

excluded. Origins and destinations outside the model region were coded to out network’s boundary link. 17

After completion of this process, the resulting demand of about 9 million trips within the modeled 18

area can be seen in Table 1, segmented by trip mode type and time of day. Each of these trips was then 19

input to MATSim as an “agent” making only one trip. 20

Table 1 Trips By Mode and Time of Day 21

Time of Day Period

Trip Mode AM MD PM EV NT All

Commercial Vehicle 108,845 615,406 145,369 38,165 38,032 945,817

Non-Motorized 155,141 308,366 257,522 76,880 - 797,909

Single Vehicle 1,032,462 1,813,018 2,324,005 1,229,201 705,438 7,104,124

Transit 15,225 27,917 22,517 8,107 - 73,766

All 1,311,673 2,764,707 2,749,413 1,352,353 743,470 8,921,616

22

Next, every MATSim agent was assigned a trip start time based on the time-of-day period they 23

originated from. Start times were sampled from the time-of-day distribution of counts provided in the E-7 24

model documentation. Linear interpolation was used to distribute the half-hour count data to one second 25

resolution to avoid bunching on the network due to multiple MATSim agents spawning in the same location 26

at the same time. 27

All MATSim agents were also assigned network links for start and end locations. Trip ends inside 28

the modeled area were allocated to MATSim network links using Monte Carlo simulation based on link 29

length (longer links were more likely to be selected than shorter links). Links corresponding to freeways 30

and freeway ramps were excluded from the sampling subset to avoid agents starting or ending on highways. 31

As noted above, if the agent entered or exited the modeling region, the start or end link corresponded to the 32

appropriate link crossing the boundary. 33

Finally, each agent was assigned attributes for use in a mode choice model. Household income 34

attribute was given based upon the PA matrix the trip originated from. Auto ownership attribute was given 35

based upon a distribution derived from Wayne County Census data for each income category. Commercial 36


8

vehicle agents and external trips were not given these attributes, and were excluded from the subsequent 1

mode choice model. 2

Mode Choice Model 3

In this work, a discrete mode-choice (DMC) extension of MATSim was used to simulate agents’ 4

mode choice decisions. The extension enables one to utilize traditional discrete mode-choice models within 5

a MATSim simulation. It can consider trip or tour-based constraints and mode availability rules. An 6

estimator assigns utility for each transport mode based on travel costs and other travel characteristics. A 7

selector component defines how the alternative modes are chosen. Detailed description of the extension can 8

be found in (33, 34). In the DMC extension of MATSim, only a randomly sampled portion of the population 9

in any given iteration performs mode choice. For this study, in each iteration, we sample 10% of the agents 10

to perform mode and route choice decisions. 11

Furthermore, we use a multinomial logit model for mode choice. The choice model formulation is 12

given in Equations 1 to 5 while the model parameters are specified in Table 2. The utility 𝑈𝑖, (𝑖 = car, pt, 13 walk, bike, DRT) is calculated per mode, based on agents’ income and auto availability attributes. 𝑘, 14 represents income level, (Low income, Middle-Low income, Middle High & High); while 𝑙 represents auto 15 ownership levels, (No Autos, Some Autos, All Autos), as depicted in Table 2. Choice variables are denoted 16

as 𝑥 while 𝛽 and 𝛼 denote marginal utility parameters and alternative specific constants (ASC) respectively. 17 The mode choice variables include in-vehicle travel time, out-of-vehicle travel time (wait time and 18

access/egress time), and travel cost. The parameters in the model are taken from the SEMCOG mode choice 19

model. Travel times and costs are provided by the MATSim network router based on the trip origin, 20

destination, and departure time. Note that we do not consider the cost of parking in this model. However, 21

parking is plentiful and free in most of Wayne County outside of downtown Detroit. 22

We apply parameters for public transit to the DRT mode (Equation 5) since the SEMCOG model 23

does not consider DRT. The DRT service is available only for internal trips, that is, those trips that originate 24

and end in Wayne County. This is due to the fact that we do not know where external trips in the model 25

start or end. Furthermore, since a door-to-door DRT scheme is applied, there is no access or egress time as 26

for the public transport. 27

𝑈𝑐𝑎𝑟,𝑘,𝑙 = 𝛼𝑐𝑎𝑟,𝑖𝑛𝑐𝑜𝑚𝑒,𝑘 + 𝛼𝑐𝑎𝑟,𝑎𝑢𝑡𝑜𝑠,𝑙 + 𝛽𝑐𝑎𝑟,𝑡𝑟𝑎𝑣𝑒𝑙𝑇𝑖𝑚𝑒 . 𝑥𝑐𝑎𝑟,𝑡𝑟𝑎𝑣𝑒𝑙𝑇𝑖𝑚𝑒 + 28

𝛽𝑐𝑜𝑠𝑡, 𝑖𝑛𝑐𝑜𝑚𝑒,𝑘 . 𝑥𝑐𝑎𝑟,𝑐𝑜𝑠𝑡 (1) 29

𝑈𝑝𝑡,𝑘,𝑙 = 𝛼𝑝𝑡,𝑖𝑛𝑐𝑜𝑚𝑒,𝑘 + 𝛼𝑝𝑡,𝑎𝑢𝑡𝑜𝑠,𝑙 + 𝛽𝑝𝑡,𝑡𝑟𝑎𝑣𝑒𝑙𝑇𝑖𝑚𝑒 . 𝑥𝑝𝑡,𝑡𝑟𝑎𝑣𝑒𝑙𝑇𝑖𝑚𝑒 + 𝛽𝑝𝑡,𝑤𝑎𝑖𝑡𝑇𝑖𝑚𝑒 . 𝑥𝑝𝑡,𝑤𝑎𝑖𝑡𝑇𝑖𝑚𝑒 +30

𝛽𝑝𝑡,𝑎𝑐𝑐𝑒𝑠𝑠/𝑒𝑔𝑟𝑒𝑠𝑠𝑇𝑖𝑚𝑒 . 𝑥𝑝𝑡,𝑎𝑐𝑐𝑒𝑠𝑠/𝑒𝑔𝑟𝑒𝑠𝑠𝑇𝑖𝑚𝑒 + 𝛽𝑝𝑡,𝑐𝑜𝑠𝑡, 𝑖𝑛𝑐𝑜𝑚𝑒,𝑘 . 𝑥𝑝𝑡,𝑐𝑜𝑠𝑡 (2) 31

𝑈𝑤𝑎𝑙𝑘,𝑘,𝑙 = 𝛼𝑤𝑎𝑙𝑘,𝑖𝑛𝑐𝑜𝑚𝑒,𝑘 + 𝛼𝑤𝑎𝑙𝑘,𝑎𝑢𝑡𝑜𝑠,𝑙 + 𝛽𝑤𝑎𝑙𝑘,𝑡𝑟𝑎𝑣𝑒𝑙𝑇𝑖𝑚𝑒 . 𝑥𝑤𝑎𝑙𝑘,𝑡𝑟𝑎𝑣𝑒𝑙𝑇𝑖𝑚𝑒 (3) 32

𝑈𝑏𝑖𝑘𝑒,𝑘,𝑙 = 𝛼𝑏𝑖𝑘𝑒,𝑖𝑛𝑐𝑜𝑚𝑒,𝑘 + 𝛼𝑏𝑖𝑘𝑒,𝑎𝑢𝑡𝑜𝑠,𝑙 + 𝛽𝑏𝑖𝑘𝑒,𝑡𝑟𝑎𝑣𝑒𝑙𝑇𝑖𝑚𝑒 . 𝑥𝑏𝑖𝑘𝑒,𝑡𝑟𝑎𝑣𝑒𝑙𝑇𝑖𝑚𝑒 (4) 33

𝑈𝑑𝑟𝑡,𝑘,𝑙 = 𝛼𝑑𝑟𝑡,𝑖𝑛𝑐𝑜𝑚𝑒,𝑘 + 𝛼𝑑𝑟𝑡,𝑎𝑢𝑡𝑜𝑠,𝑙 + 𝛽𝑑𝑟𝑡,𝑡𝑟𝑎𝑣𝑒𝑙𝑇𝑖𝑚𝑒 . 𝑥𝑑𝑟𝑡,𝑡𝑟𝑎𝑣𝑒𝑙𝑇𝑖𝑚𝑒 + 34

𝛽𝑑𝑟𝑡,𝑤𝑎𝑖𝑡𝑇𝑖𝑚𝑒 . 𝑥𝑑𝑟𝑡,𝑤𝑎𝑖𝑡𝑇𝑖𝑚𝑒 + 𝛽𝑑𝑟𝑡,𝑐𝑜𝑠𝑡, 𝑖𝑛𝑐𝑜𝑚𝑒,𝑘 . 𝑥𝑑𝑟𝑡,𝑐𝑜𝑠𝑡 35

(5) 36

37

38


9

1

2

Table 2 Calibrated Parameters for Mode Choice Model 3

4

5

Simulation of Demand Responsive Transit (DRT) 6

DRT simulations for Wayne County are conducted using a DRT extension (35) of MATSim. The 7

DRT extension has a dispatching algorithm managing the movement of DRT fleet and travelers’ requests. 8

This process, like in real life, is dynamic and depends on the state of the simulation system. An agent 9

choosing the DRT service, submits a trip request and waits for a vehicle. The dispatcher then assigns a 10

vehicle with the smallest detour time loss, to handle the request. Detour time loss is a measure of detouring 11

due to adding additional passengers to a vehicle, which consists of pick-up detour time, drop-off detour 12

time, and stop duration for pickup and drop off. It is added to the total travel time of agents with shared 13

rides. If there are more trip requests than vehicles in the system, it is possible that some agents cannot be 14

served within pre-defined wait time. In that case these agents’ trips will be labelled as "rejected", and then 15

removed from the micro-simulation of the current iteration. When a DRT vehicle drops off its last passenger 16

and there are no other trip requests at that particular time, the DRT vehicle stays at the location where the 17

last passenger has been dropped off. 18

The DRT module requires some configuration settings which define the vehicle fleet and 19

operational parameters. Vehicles are generated based on fined fleet size, maximum vehicle capacity, and 20

start and end locations. We set a maximum customer wait time and detour time of 15 minutes each and a 21

Trip Mode Low

Income

Middle-

Low

Income

Middle-

High &

High

Income

No

Autos

Some

Autos

(0


10

stop duration of 105 seconds (1.75 minutes). Finally, we set the start time for the simulation at 12 AM and 1

end time at 4:00 PM. A total of approximately 5.4 million trips are simulated over this period. 2

3

DRT Service Level Scenarios 4

The objective of this paper is to understand the demand potential of DRT for Wayne County based 5

on fleet size, cost and vehicle capacity factors. Therefore we defined 16 DRT scenarios consisting of four 6

different vehicle fleet sizes (100, 250, 500 and 1000), two different vehicle capacities (4 and 7 occupants) 7

and two different fare rates ($2 and $4). Vehicle capacity had no influence on the results, hence analysis 8

are presented for 4 occupancy vehicles. 9

10

RESULTS 11

In this section, we compare the different scenarios and analyze the impact of different fleet sizes, 12

vehicle sizes, and fares on, DRT demand, operational performance indicators, and system performance. 13

First we looked at three operational performance indicators, fleet size and fare level, vehicle kilometers 14

travelled (VKT), and vehicle occupancy. This provides insights on DRT utilization and an understanding 15

of minimum requirements for a feasible DRT service in Wayne County. Next we looked at how the DRT 16

service parameters (wait time, detour time) affect two measures of customer experience; waiting time and 17

affordability. Finally, we look at the system impact of DRT in Wayne County by comparing the baseline 18

scenario to the DRT scenarios. 19

Fleet size and fare level 20

The first scenario runs, were sensitivity tests for the minimum fleet size required to serve demand 21

at each fare level. Results are shown in Figure 2, with demand at the $2 and $4 levels shown with lines, and 22

served/rejected trips shown with bars. In MATSim, the DRT demand is the total number of agents 23

requesting the DRT service, whether served or not; it is the sum of actual DRT rides and rejections. Total 24

rejected rides are quite high with a small fleet, but are reduced substantially as more vehicles are added to 25

the fleet. 26

As can be seen in the figure, increasing fleet size (i.e. supply) results in a minimal change in the 27

total demand, observable as a downward trendline. However, a 100% increase in the price from $2 to $4 28

strongly reduces the demand by about 50%. At the $2 fare level, DRT trips keep increasing with fleet size, 29

with the rate of change in rides per fleet size increase as 50%, 35%, and 18% respectively. This is not the 30

case at the $4 fare level. Between 100 and 250 fleets, there is an increase of 26%, as the fleet size doubles, 31

this increase drops to 13% and finally a second doubling from 500 to 1000 fleet leads to a minimal increase 32

of 7%. The DRT trip rejections follow a similar trend. 33

To explain the almost stable and slightly decreasing demand, one must look at how mode choice is 34

simulated in our adapted MATSim model, and the possible impact of DRT vehicles on congestion in the 35

network. Besides the maximum waiting time and maximum detour time, other parameters of the service 36

reliability (such as trip rejections) are not taken into account in the mode choice model. Consequently, if in 37

a certain area, the demand is high and a lot of people are not served, in the next iteration, agents’ decision 38

to use DRT service will not be influenced by rejection effect of the previous iteration. This would explain 39

the reason why there is almost the same number of requests even for a fleet size of 100 whose rejection rate 40

is about 80%. 41

The decreasing trend of the demand, on the contrary, possibly reveals the impact of additional DRT 42

vehicles on the network. Even though the maximum fleet size of 1000 contributes to less than 0.1% of the 43

vehicles on the network, the actual usage of the fleet may have impacted congestion in certain high demand 44

areas. Thus, increase in travel time is fed back to the mode choice model, and influences agents’ decisions 45


11

not to choose DRT service. Another explanation for the decrease in demand could be due to stochastic 1

nature of the mode choice model. This suggests a need to apply different random seeds in future work to 2

better understand these phenomena. 3

4

Figure 2: DRT demand (Requests, Rides, Rejections) 5

Vehicle kilometers travelled 6

Average total distance per DRT vehicle decreases as fleet size increases. At the $2 fare level, it 7

ranged from 483 km per DRT vehicle for 100 fleet, to 188 km for 1000 fleet; similarly, 435km and 98km 8

for a $4 fare. Even though there are more vehicles serving more trips, the overall average VKT per vehicle 9

reduces as fleet size increases. 10

Also, as fleet size increases, the ratio of passenger-less (i.e. empty) to passenger VKT decreases 11

(the ratios are 0.32, 0.26, 0.19, 0.15, and 0.33, 0.25, 0.19, 0.15 for $2 and $4 fare respectively). From Figure 12

4 one can clearly see that with larger fleets, the percentage of empty vehicles driving around reduces, as 13

more vehicles spread throughout the network. This effect is the same for both fares except for the magnitude 14

of demand. We will also see this trend reflected in the reduced wait times in Figure 6. 15

Furthermore, compared to an average trip distance of 10km for car travel, the majority of the 16

passengers are using DRT for relatively shorter trips with average trip distances between 5 and 7km. 17

18

Vehicle occupancy 19

Figure 3 shows little actual ridesharing in any of the scenarios tested. This is also clearly seen in 20

Figure 4 which shows that most of the DRT trips throughout the 16hr day period are single occupancy trips. 21

This may be due to the fact that the model area includes a large suburban portion of Wayne County 22

characterized by low density development, which makes it challenging to pool rides. The highest average 23

occupancy per vehicle achieved in the simulations (1.12) was less than the smallest vehicle capacity tested 24

(4 passengers). As a consequence, vehicle capacity had no visible effect on the results. 25


12

It is interesting to note that even with the high percentage of idle vehicles throughout the day for 1

fleet size of 500 and 1000, these vehicles still reject agents’ requests, likely because of the pre-defined 2

service parameters, maximum wait time and detour time, and the consequence of modelling a large region. 3

Since there is no rebalancing between trips, the vehicles are sparsely spread in all of Wayne County, as they 4

pick up and drop off passengers to their destination. So even if vehicles are not in use, rejections can still 5

happen, as the requests may be too far away from where the vehicle is sitting and are unable to meet the 6

detour constraints for in-vehicle passengers and the wait time constraint of the requesting passengers. There 7

is need to optimize the pooling, by testing different detour factors and wait times in future scenarios as well 8

as limiting the DRT service to specific areas with high demand. 9

10

11

12

Figure 3: Vehicle Distance traveled by Occupancy 13

14

15

16


13

1

2 Figure 4: Vehicle Occupancy by time of day (a: $2 fare, b: $4 fare, note: stay means idle time) 3

4


14

Passenger rides and wait time 1

We have already seen more agents making requests for the DRT service than the service can meet. 2

Figure 5 shows the hourly distribution of the DRT passenger rides. The passenger rides which are the 3

accepted requests, are higher in the morning peak period between 6am and 8am. Afterwards, there is a 4

slight decrease that spreads throughout the whole day for smaller fleet sizes, but peaks again in the afternoon 5

for larger fleet sizes. This is because with more vehicles, more demand is met at peak times. In Figure 4, 6

smaller fleets have no idle vehicles and are unable to meet the demands at peak times for about 45% of 7

these passengers who are either going home or going to work. 8

As shown in Figure 6, average wait time reduces as fleet size increases on average at [10, 8, 6, 5] 9

minutes for the fleet sizes respectively. For all fleet sizes, high wait times are experienced in the morning 10

peak period when there are more rejections. The fare policy does not have a significant effect on the wait 11

time, but the wait time for the $4 fare scenario is slightly lower than the $2 fare scenario. 12

13

14

15

16

Figure 5: Passenger rides per time of day 17


15

1

Figure 6: Wait time per time of day 2

3

Affordability 4

Figure 7 shows the distribution of DRT riders by income class at each fare level with share of PT 5

users that switched to DRT from the baseline scenario. The PT fare is set at $2 while DRT fares tested are 6

$2 and $4. Income classes of DRT users are similar to that of public transport riders regardless of the fare 7

policy (INC1, INC2, and INC3 represent low, middle-low, and high-middle high income earners 8

respectively). This shows that DRT fare is relatively affordable for all income segments, which makes it 9

suitable for improving everyone’s mobility. 10


16

1

2

Figure 7: Percentage of income levels of DRT users and PT users (a: $2 fare for PT and 3

DRT, b: $2 for PT and $4 for DRT) 4


17

System VKT impact 1

Private cars are dominant in Wayne County and contribute to almost 80% of the mode share. An 2

interesting question is how the addition of vehicles that drive almost all day, would impact total VKT. To 3

calculate this, all agents that used DRT are extracted from the baseline scenario, their share of mode 4

replaced are compared, and the difference in VKT generated in the transport network is computed. This is 5

summarized in Table 3 showing; the share of replaced mode in the baseline scenario, the difference and 6

percentage change in VKT between replaced modes from baseline scenario and DRT scenarios, and the 7

difference in system-wide VKT between baseline scenario and DRT scenarios. 8

Looking at the share of replaced modes, the DRT trips come mostly from car trips. Walk, public 9

transit and bike only contribute about 15% to the modes replaced by DRT trips. However, these agents have 10

increased VKT in the transport system by switching to DRT. The VKT change depending on fleet size has 11

the highest increase at 66% for 100 fleet and the lowest at 27% for 1000 fleet. 12

Nevertheless, the percentage change of VKT on a system-wide level is small. The highest across 13

the different scenario is about 0.07%. Further optimization on fleet size, service parameters, as well as 14

limiting the service area, should minimize and possibly have a positive impact on the system-wide VKT. 15

16

Table 3: DRT Impact on Mode Replacement and VKT 17

Fare_Fleet

Size

Share of Replaced Modes Difference in VKT for Replaced Modes Net Change for

System-wide VKT

(%)

Bike

(%)

Car

(%)

PT

(%)

Walk

(%)

DRT

VKT

(km)

Baseline

VKT

(km)

VKT

difference

(km)

VKT

relative

change

(%)

$2_100 3.40 82.41 1.65 12.53 48309 27775 18280 65.81 0.031

$2_250 3.50 83.53 2.35 10.61 105541 70597 34944 49.50 0.060

$2_500 3.24 85.18 2.54 9.04 161701 121128 40573 33.50 0.069

$2_1000 3.10 86.06 2.57 8.26 188340 148660 39680 26.69 0.068

$4_100 2.97 83.76 2.25 11.03 43475 26128 17347 66.39 0.030

$4_250 2.65 86.13 2.38 8.85 79228 53863 25365 47.09 0.043

$4_500 2.49 86.98 2.53 8.00 92702 69002 23700 34.35 0.041

$4_1000 2.46 87.33 2.52 7.69 98218 76569 21649 28.27 0.037

Note: Overall VKT for baseline scenario is 58409642 km. This is used in computing the difference in system

wide VKT (VKT difference per fleet/Overall VKT)

18


CONCLUSION 1

This section summarizes some of the most important findings and presents potential improvements of the 2

approach presented here to model a DRT service. 3

The results show that while the potential demand for DRT is high in Wayne County, it is 4

challenging to serve the demand due to the large geographic size and low population density of the modeled 5

area. While it is possible to serve a larger proportion of trips with a relatively larger fleet, this results in 6

underutilization of the vehicles, as vehicles are unable to meet wait time and detour time constraints for 7

picking up passengers. We find, based on the scenarios tested, low potential for ridesharing in a DRT service 8

at the scale of Wayne County. Further analysis could test smaller service areas with higher population 9

density which may have higher potential. 10

We found that the net increase in VKT due to DRT vehicles is low, even with the largest fleet size 11

tested. We note that the current model does not consider two or more people from the same household 12

traveling together, so the results would tend to under-estimate vehicle occupancy as well as the VKT. We 13

also note that the simulation period ends by 4 PM, so certain types of discretionary travel, such as dining 14

out, often undertaken by multiple persons from the same household, are under-represented in the simulation. 15

While average wait times for DRT increase with respect to demand, on average the wait times are 16

reasonable even at the highest demand period modeled. We also find that pricing DRT similar to transit 17

results in greater mobility for lower income travelers, and that pricing DRT at higher rates may substantially 18

reduce demand. 19

We identified several possible further improvements of the methodology based on the limitations 20

of the study. The demand generated from the SEMCOG model is trip based. Therefore, the activity-chains 21

of individuals are not preserved. This is a limiting factor as certain constraints that tours impose are not 22

captured. A possible way to overcome this limitation is to utilize an activity-based model, to generate full 23

daily plans (i.e. ActivitySim) or other approaches used to generate mobility demand for MATSim (36). The 24

mode choice parameters for DRT are not estimated based on empirical data. While in the context of the 25

Wayne County, it is reasonable to adapt the parameters for public transport for DRT, it would be valuable 26

to conduct a stated preference mode-choice survey where DRT service is explicitly captured. Information 27

on rejected agents in previous iterations is not used in the following iterations when agents make mode-28

choice decision. It would be interesting to capture this measure of reliability of the service in the mode-29

choice model. How individuals value reliability and how it affects their choice set, is however an open 30

research question, and one worth investigating. 31

Based on the findings of the current study, for future scenarios, we may want to test different service 32

parameters as well as rebalancing strategies. The DRT service in this paper covers the whole study region. 33

Therefore, it is hard to meet pickup constraints, as the service is spread thin within the region. Idle vehicles 34

are unable to service requests that are too far away from where the vehicle is currently waiting. Using the 35

spatial and temporal information on the demand in the whole region it would be possible to optimize the 36

service area in order to increase the share of pooled rides, and potentially create a service that is able to 37

reduce the VKT in the region. In order to further optimize the DRT service, and to make the final judgement 38

of its potentials in the region, additional sensitivity tests are required. Those would include different cost 39

structures, fleet sizes, acceptable waiting and detour times. Currently the DRT service does not anticipate 40

potential demand and does not perform preemptive relocations in order to better serve the demand. To 41

overcome this limitation, different relocation policies could be implemented in order to analyze their 42

potential improvements of the service. 43

In summary, there is potential for the DRT service to drive demand and be efficiently utilized. 44

Presently the results show reasonable demand for the service, low empty distance, and that the average 45


19

VKT per vehicle lowers with increasing fleet sizes. However, there is still the need to optimize the DRT 1

service parameter in order to maximize the efficiency of the system and improve ride sharing. 2

3

ACKNOWLEDGEMENT 4

The authors would like to acknowledge Ford Motor Company for funding this research, Justin Culp (RSG) 5

for contributions to network development, and Southeast Michigan Council of Governments (particularly 6

Jilan Chen and Alex Bourgeau) for making the SEMCOG model and data available to the project team. 7

8

AUTHOR CONTRIBUTION 9

The authors confirm contribution to the paper as follows: study conception and design: all authors; data 10

preparation: Grace O. Kagho, David Hensle, Milos Balac, Joel Freedman; analysis and interpretation of 11

results: all authors; draft manuscript preparation: all authors. All authors reviewed the results and 12

approved the final version of the manuscript. 13


REFERENCES

1. Inturri, G., Le Pira, M., Giuffrida, N., Ignaccolo, M., Pluchino, A., Rapisarda, A., & D'Angelo, R. (2019). Multi-agent simulation for planning and designing new shared mobility services. Research

in Transportation Economics, 73, 34-44.

2. Logan, P. (2007). Best practice demand-responsive transport (DRT) policy. Road & Transport Research: A Journal of Australian and New Zealand Research and Practice, 16(2), 50.

3. Brake, J., Nelson, J. D., & Wright, S. (2004). Demand responsive transport: towards the emergence of a new market segment. Journal of Transport Geography, 12(4), 323-337.

4. Cooney, P., Phillips, E., & Rivera, J. (March 2019). Auto insurance and economic mobility in Michigan: A cycle of poverty. M Poverty Solutions–University of Michigan. Retrieved from:

https://poverty.umich.edu/files/2019/03/PovertySolutions-CarInsurance-PolicyBrief-r1.pdf 5. Lee, J. (2014). Urban built environment and travel behavior: Understanding gender and socio-

economic disparities in accessibility and mobility of urban transportation. Michigan State

University.

6. TRB, “Shared Mobility and the Transformation of Public Transit,” Transportation Research Board

of the National Academies of Science, Washington, DC, 188, 2016. doi: 10.17226/23578.

7. Shaheen, S., Cohen, A., Yelchuru, B., Sarkhili, S., & Hamilton, B. A. (2017). Mobility on demand

operational concept report (No. FHWA-JPO-18-611). United States. Department of

Transportation. Intelligent Transportation Systems Joint Program Office.

8. Narayanan, S., Chaniotakis, E., & Antoniou, C. (2020). Shared autonomous vehicle services: A

comprehensive review. Transportation Research Part C: Emerging Technologies, 111, 255-293.

9. Zachariah, J., Gao, J., Kornhauser, A., & Mufti, T. (2014). Uncongested mobility for all: A proposal

for an area wide autonomous taxi system in New Jersey (No. 14-2373).

10. Yang, H., & Wong, S. C. (1998). A network model of urban taxi services. Transportation Research

Part B: Methodological, 32(4), 235-246.

11. Spieser, K., Treleaven, K., Zhang, R., Frazzoli, E., Morton, D., & Pavone, M. (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). Springer, Cham.

12. Liang, X., de Almeida Correia, G. H., & Van Arem, B. (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, 115-129.

13. International Transport Forum, “Shared Mobility: Innovation for Liveable Cities,” Text, May 2016.

Accessed: May 15, 2017. [Online]. Available: https://www.itf-oecd.org/shared-mobility-innovation-

liveable-cities.

14. Fagnant, D. J. (2014). The future of fully automated vehicles: opportunities for vehicle-and ride-

sharing, with cost and emissions savings (Doctoral dissertation).

15. Childress, S., Nichols, B., Charlton, B., & Coe, S. (2015). Using an activity-based model to explore

the potential impacts of automated vehicles. Transportation Research Record, 2493(1), 99-106.

16. Lokhandwala, M., & Cai, H. (2018). Dynamic ride sharing using traditional taxis and shared

autonomous taxis: A case study of NYC. Transportation Research Part C: Emerging

Technologies, 97, 45-60.

17. Hyland, M., & Mahmassani, H. S. (2020). Operational benefits and challenges of shared-ride

automated mobility-on-demand services. Transportation Research Part A: Policy and Practice, 134,

251-270.

18. Di Febbraro, A., Sacco, N., & Saeednia, M. (2012). One-way carsharing: Solving the relocation

problem. Transportation research record, 2319(1), 113-120.

https://poverty.umich.edu/files/2019/03/PovertySolutions-CarInsurance-PolicyBrief-r1.pdfhttps://www.itf-oecd.org/shared-mobility-innovation-liveable-citieshttps://www.itf-oecd.org/shared-mobility-innovation-liveable-cities


21

19. Chen, T. D., Kockelman, K. M., & Hanna, J. P. (2016). Operations of a shared, autonomous, electric

vehicle fleet: Implications of vehicle & charging infrastructure decisions. Transportation Research

Part A: Policy and Practice, 94, 243-254.

20. Liu, J., Kockelman, K. M., Boesch, P. M., & Ciari, F. (2017). Tracking a system of shared

autonomous vehicles across the Austin, Texas network using agent-based

simulation. Transportation, 44(6), 1261-1278.

21. Alam, M. J., & Habib, M. A. (2018). Investigation of the impacts of shared autonomous vehicle

operation in halifax, canada using a dynamic traffic microsimulation model. Procedia computer

science, 130, 496-503.

22. Loeb, B., & Kockelman, K. M. (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, 374-385.

23. Brownell, C. K. (2013). Shared autonomous taxi networks: An analysis of transportation demand in

NJ and a 21st century solution for congestion. Princeton University, April.

24. Martinez, L. M., Correia, G. H., & Viegas, J. M. (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), 475-495.

25. Gucwa, M. (2014, July). Mobility and energy impacts of automated cars. In Proceedings of the

automated vehicles symposium, San Francisco.

26. Hörl, S., Erath, A., & Axhausen, K. W. (2016). Simulation of autonomous taxis in a multi-modal

traffic scenario with dynamic demand. Arbeitsberichte Verkehrs-und Raumplanung, 1184.

27. Azevedo, C. L., Marczuk, K., Raveau, S., Soh, H., Adnan, M., Basak, K., ... & Ben-Akiva, M. (2016).

Microsimulation of demand and supply of autonomous mobility on demand. Transportation

Research Record, 2564(1), 21-30.

28. Zhang, W., Guhathakurta, S., Fang, J., & Zhang, G. (2015). Exploring the impact of shared

autonomous vehicles on urban parking demand: An agent-based simulation approach. Sustainable

Cities and Society, 19, 34-45.

29. Moreno, A. T., Michalski, A., Llorca, C., & Moeckel, R. (2018). Shared autonomous vehicles effect

on vehicle-km traveled and average trip duration. Journal of Advanced Transportation, 2018.

30. Horni, A., Nagel, K., & Axhausen, K.W. The Multi-Agent Transport Simulation MATSim. London:

Ubiquity Press, 2016.

31. Cambridge Systematics, Inc. & AECOM. (2019). SEMCOG E7 travel model improvement and

update. Technical report prepared for Southeast Michigan Council of Governments.

32. Poletti, F., Bösch, P. M., Ciari, F., & Axhausen, K. W. (2017). Public transit route mapping for large-

scale multimodal networks. ISPRS International Journal of Geo-Information, 6(9), 268.

33. Hörl, S., Balac, M., & Axhausen, K. W. (2018). A first look at bridging discrete choice modeling

and agent-based microsimulation in MATSim. Procedia computer science, 130, 900-907.

34. Hörl, S., Balać, M., & Axhausen, K. W. (2019). Pairing discrete mode choice models and agent-

based transport simulation with MATSim. In 2019 TRB Annual Meeting Online (pp. 19-02409).

Transportation Research Board.

35. Maciejewski, M., Bischoff, J., Hörl, S., & Nagel, K. (2017). Towards a testbed for dynamic vehicle

routing algorithms. In International Conference on Practical Applications of Agents and Multi-Agent

Systems (pp. 69-79). Springer, Cham.

36. Hörl, S., & Balac, M. (2020). Reproducible scenarios for agent-based transport simulation: A case

study for Paris and Île-de-France. In Arbeitsberichte Verkehrs-und Raumplanung, 1499; IVT: ETH

Zurich, Zurich, Switzerland, 2020.


22

demand responsive transit simulation of wayne county, michigan · kagho, hensle, balac, freedman,...

Documents