Control strategies for transport network optimisation

Abstract

Bus Bunching is one of the problems that impinges the perception of public bus transport services and happens when buses with a predetermined headway end up arriving at the same bus stop together. This happens due to unpredictability in passenger demand and other external factors such as traffic and road accidents. This bunching then leads to an increase in passenger waiting times, since the buses end up passing less frequently. Moreover, unbalanced occupancy rates arise since the passengers waiting at a stop will board the first bus in the bunched group, leaving the rest relatively empty, hence resulting in an inefficient use of resources. As a result, especially in a society where traffic is always on the increase, it is imperative that this issue is tackled so that the bus service is perceived as reliable and dependable in order to encourage the public to make use of it. The problem is widely studied and most of the earlier works use an optimisation or rule‐based approach. This study proposes the use of Reinforcement Learning to tackle the problem. The resulting trained policy would choose the action to be taken by each bus at specific control points, which in our case are bus stops on the route. Most studies usually make use of just one control action, but we propose to use both Bus Holding and Stop‐Skipping. The algorithms will choose whether to hold the bus at a stop, skip the stop, or else proceed normally whenever a bus arrives at a bus stop. The Bus Bunching phenomenon was simulated by replicating an existing benchmark scenario using the SUMO traffic simulator on which the algorithms, namely TRPO and PPO, were trained. Moreover, we incorporate traffic into the simulator in order to provide a more realistic scenario since this is usually omitted from similar studies. We also train the algorithms on scenarios where the buses are already bunched such that the policy also learns how to recover from a bunched scenario. The performance of the policies were tested and evaluated using the headway standard deviation, average waiting time, and occupancy dispersion metrics. They were compared to a No Control policy and a Rule‐Based Control policy. This study is one of the first of its kind since the results show that the trained policy is able to prevent bus bunching from occurring and also restore a bunched network back to the ideal headways.