Rebalancing Autonomous Vehicles using Deep Reinforcement Learning

The shared autonomous mobility-on-demand (AMoD) system is a promising business model in the coming future which provides a more efficient and affordable urban travel mode. However, to maintain the efficient operation of AMoD and address the demand and supply mismatching, a good rebalancing strategy is required. This paper proposes a reinforcement learning-based rebalancing strategy to minimize passengers’ waiting in a shared AMoD system. The state is defined as the nearby supply and demand information of a vehicle. The action is defined as moving to a nearby area with eight different directions or staying idle. A 4.6 4.4 km2 region in Cambridge, Massachusetts, is used as the case study. We trained and tested the rebalancing strategy in two different demand patterns: random and first-mile. Results show the proposed method can reduce passenger’s waiting time by 7% for random demand patterns and 10% for first-mile demand patterns.