Mobility Patterns in Shared, Autonomous, and Connected Urban Transport

3.1 Experimenting with Different DRT Systems

An application of disruptive transportation is the replacement of a fixed route bus service by a partially or non-fixed route service. This can occur when the original service has low patronage and operating costs are relatively high.

In 2013, Public Transport Victoria elected to replace a regional bus service by a service operated by taxis, known as Flexiride [38]. The service has some fixed elements, in that the taxi must always start at the same point at set times during the day. If no one is waiting at that point, nor any bookings have been taken, then the taxi returns to regular service. If there are bookings, the driver decides best how to service those passengers and proceeds. Bookings can only be taken before the taxi leaves the first point.

Data from this new service was used as a base case to experiment with different types of services. A model was built using MATSim, and two scenarios were tested: the existing Flexiride service, along with a fully flexible service, where the constraint on start time and start point were removed [39]. This simulation could then be used to compare the two configurations with respect to the number of successful trips, the average waiting time, and the average excess time. More flexibility, of course, delivers better performance for passengers but higher “costs” in terms of longer travel times or more vehicles needed.

One limitation of this model was that only one day can be simulated, which means that the long-term demand for a flexible service cannot be explored. Another study using the same model analyzed how demand could change over many days based on previous experiences with a flexible service [40]. For example, when deciding to make a trip, the estimated travel time could be used to determine whether the trip is worthwhile. This estimate could be based on personal experiences, for example, how long it usually takes to get from home to the destination, or on global experiences, for example, if others are experiencing delays then you might too.

Another limitation was that only flexible services could be included in the model. This means operation could differ in congested scenarios. However, the use of a multimodal simulation tool such as MATSim means future integration is possible.

The advantage of using simulation in this case is to obtain an idea of how different service configurations will perform with different demand patterns. The ability to build in passenger choice in the future will provide more realistic results.

3.2 Comparing Buses and on-Demand Systems

Public transport operator companies face financial challenges when it comes to providing transit in suburban areas. The main problem is that the demand density is so low that makes it difficult to financially maintain a high frequency public transport and the low frequency is most unattractive to users. Demand-responsive transport (DRT) has promising features to solve this problem. However, there are high financial and operational risks associated with the implementation of a DRT system, which could be foreseen and prevented by the help of simulation. Following preliminary work [41], we use simulation to test our hypothesis that replacing conventional public transport (CPT) with DRT would improve people’s mobility in a small area and in turn solve the problem of underutilized CPT. To this end, the two modes have been compared in term of user performance and cost to the operator.

A limited number of studies have been focused on comparing these modes or finding the demand switch point, which is the highest demand that makes DRT outperform CPT [42, 43, 44, 45]. These studies concluded that DRT outperform CPT in areas with low demand density. Other studies tried to define the demand switch point by analytical methods [46, 47]; however, their studies were focused on DRT operating as feeders.

Two important issues were missing in all these studies: the study of a dynamic ad hoc DRT and considering user preferences. To bridge this gap, this work utilizes an agent-based traffic simulation, MATSim, with an embedded dynamic routing algorithm. This model has the capability of modeling all the passengers as individual agents and offers them ad hoc DRT service to study the impact of replacing CPT with DRT on an individual level. Moreover, the diversity of investigated scenarios helps find the demand switch point.

To test the hypothesis, an extensive comparison was conducted investigating several variations in network shape, demand density and transit frequency. The combination of four transit system (DRT and CPT with three frequencies: 7.5, 15, and 30 min) and five demand density (1, 2, 3, 4, 5, 10 and 15 requests per minute) were investigated in two different network shape: grid and star shape (X) with 16 km² area.

The scenarios were evaluated in terms of the quality of passengers’ mobility and the operator cost. The quality of the mobility has been defined in this work as the passengers’ perceived travel time and presented as Virtual in-vehicle time (VIVT). The importance of people’s perception of travel time in their decision about their mode of transport has been verified previously [48, 49]. It is calculated according to passengers’ walking time, waiting time, in-vehicle ride time, and number of transfers. The cost to the operator was calculated according to the size of the fleet, the operating hours, and the kilometers driven by each vehicle. In this study, the uptake for both services and their ticket price is assumed to be equal.

The results demonstrate that replacing CPT with DRT results in a significant improvement in people’s mobility. The VIVT is the lowest for DRT users in all scenarios, which shows the superiority of this mode in terms of user performance. Moreover, the percentage of people waiting less than 10 min has been calculated for all scenarios. The average percentage for CPT users is 99% in both networks, while it goes up to 70 and 80% in star shape and grid network respectively.

After verifying that replacing DRT with CPT results in a better situation for users, it is necessary to evaluate their cost. DRT’s cost is less than any frequency of CPT in grid network and the demand switch point can be defined according to the frequency. However, the highest demand for which DRT can outperform CPT (in terms of cost) is almost 7 requests per minute.

As a conclusion, replacing CPT with DRT results in improvement in people’s mobility in small areas with low demand, mostly without any extra cost to the operator. This means that this new system can solve the problem of uneconomical transit services for the operators and provide a better mobility option for inhabitants. A high quality, door-to-door public transport service also improves the social equity of a suburb for its habitants and increase the standard of living. The fact that the operator cost depends on the demand in DRT operation demonstrate its flexibility, which is very useful in adapting the supply to demand in different time periods. This work demonstrated the power of simulation tools to evaluate the performance of a new system and compare it to old ones on a limited time scenario. However, it is expected that by developing more complex models, it is also possible to study the DRT performance on a longer run (for instance over a week or a month).

3.3 Motivating Behavioral Change

Human mobility and travel decision-making are complex behaviors that are affected not only by physical conditions, e.g., space-time limits, but also by sociopsychological factors. Especially for such collaborative behaviors with multiple participants as ridesharing, decisions are made in many aspects. Trajectory analysis has indicated that many people could do ridesharing according to their space-time concurrence [50]. But despite the environmental and economic benefits, there is still a low rate in participation of ridesharing [51, 52, 53]. The contradiction is partially explained by some surveys showing that the willingness to share rides with social contacts (a first or second degree socially connected person, hereafter called “friend”) is significantly higher than with strangers [52, 53]. Therefore, friendship can be a good drive for ridesharing. While the low willingness to share rides with strangers signifies that many of the existing rides are in fact invalid for a certain person, ridesharing exclusively with friends and declining offers from strangers nearby may lead to higher detour cost and even less opportunities to get a ride within given space-time budgets.

To prove the benefit of ridesharing with friends, two null hypotheses against the objective are to be rejected. The first one is that sharing rides only with friends significantly increases detour cost. The second is that the number of successfully matched rides is significantly lower with friends than with anyone. Detour cost and matching rate are influenced by social similarity and spatial distribution of friends. Social similarity contributes to the willingness to share rides with and detour tolerance for a person, while spatial distribution decides detour cost.

Agent-based microsimulation provides a way to cope with the complex behavior process. An agent-based transport simulation is set up using NetLogo assuming that everyone in the study area has a car. The simulation measures the reduction of the total amount of cars by changing the behaviors of private car owners. Based on a regular gridline road network with artificially generated origin/destination points of trips, the model systematically tests different parameters to reach the general conclusions applicable to different urban contexts. The parameters include social network structure (average degree of friendship), spatial distribution of friends (spatially clustered vs. random), and varied tolerance and willingness for different social levels (i.e., direct friends, indirect friends, and strangers). Two populations with small world social network structures are simulated, with 2000 and 5000 agents respectively. There are three matching patterns: (1) any driver and passenger must be direct friends, (2) any driver and passenger must be either direct or indirect friends, and (3) no one has to be friends.

The detour costs of the cross categories of social network structures and matching patterns are calculated for statistical analysis. The analysis contributes three major findings. First, detour cost with friends is not necessarily significantly higher than collaboration with anyone. Especially when friendships are spatially clustered, sharing with friends saves more. Second, even if not excluding strangers, giving priority to friends drastically increases successful matching between friends. Finally, a successful matching rate is not positively associated with the size of choice set. It also depends on the spatial distribution of friendship.

3.4 Using Simulations to Test Optimization Strategies

In order for a DRT service to operate efficiently, an important aspect to consider is the optimization of the vehicle routes. By carefully planning the vehicle routes, the operating cost can be minimized while still guaranteeing a certain level of service quality. Thus, the main use of the optimization is to determine the “sweet spot” of the trade-off between service quality and cost, which eventually leads to an economically justifiable DRT service framework.

The routing problem encountered in a demand-responsive transport is categorized as a Pick-up and Drop-off Problem with Time Windows (PDPTW), where each customer has a load to be picked up and dropped off within the nominated time windows. Occasionally, late arrivals (outside of the time windows) are allowed with some penalties being accrued. The PDPTW is typically considered to be a static problem with customer requests being fixed in advance. A number of solvers have been developed for the PDPTW, including exact solution methods [54, 55, 56, 57, 58] and heuristic algorithms [59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70]. When the scale of the problems increased and more layers of complexity are added into the problem, most exact solution methods are computationally expensive and, thus, implementation usually relies on heuristic algorithms [71].

The inevitable challenge when implementing most kinds of heuristic algorithms is the fine-tuning of their parameters. Arguably, the performance of these algorithms is heavily dependent on the selection of these parameters, and a set of parameters can be good for one scenario but not the others. One of the benefits of using a simulation platform is the ability to fine-tune the selection of these parameters prior implementation. In most cases, the methods used to fine-tune these parameters require iterations of specific conditions, which are impossible to perform in practice.

To illustrate this point, an example will be drawn from the algorithm discussed in [72]. The considered algorithm is a two-layered neighborhood search. At the bottom layer, the neighborhood algorithm is used to optimize the route of a vehicle given a certain set of trip requests to be served. At the top layer, the passenger allocation is optimized using another neighborhood search algorithm, requiring multiple call of the lower layer optimization.

In the proposed algorithm, there are two parameters at each layer, defining the neighborhood size and the number of iterations. At each iteration, the algorithm first randomly chooses and evaluates the costs of a number of solutions that are considered as the “neighbors” of the solution at the current iteration. Then, the best neighbor is defined as the solution for the next iteration (this is done even though the best neighbor is worse than the solution of current iteration, which is useful to avoid being trapped at a local extremum) and the process is repeated multiple times according to the predefined number of iterations. The detail of the algorithm is outlined in [72].

The optimization result when using this algorithm heavily depends on these parameters. With higher neighborhood size and number of iterations, it is expected that the algorithm would asymptotically approach the best possible result, at the cost of an increasingly heavy computational burden. Therefore, it is of the operator’s interest to select a set of parameters that provides an excellent optimization result with a minimal computational time.

In order to do this, a pilot study can be carried out in a simple simulation platform. As an example, the pilot study is carried out by considering a single vehicle case serving a total of 30 trip requests in a certain predefined network. Figure 1 shows, firstly, the optimization result for various parameters values (note that the neighborhood size is obtained by multiplying the neighborhood size factor and the number of stops in the route, i.e., 60) is shown and, secondly, a similar plot for the total wall time is produced. By superimposing the contour maps of both graphs, it is possible to determine a pair of values for these parameters that produces the best result while still minimizing the wall time of the optimization (in this example, via visual observation). Thus, the pilot study enables a convenient way to determine the optimal parameter values, which otherwise is impossible to obtain in practice.

Open image in new window

The benefit of using a simulation platform for routing optimization is not limited to the illustrated example. Another important aspect is to decide which algorithms to use. This is especially important in a dynamic vehicle routing problem, where heuristic is the prevalent approach. For instance, in the dynamic case, an optimal solution might become relatively of poor quality when new ad hoc requests emerge. Therefore, many suggest exploiting some known stochastic information about the demand [73, 74, 75, 76]. A comparison study can be easily carried out by using a simulation platform. Therefore, this again shows the importance of a simulation platform when deciding the most suitable optimization algorithm in a DRT service.