Proposal of Bicycle Sharing Operation System by Multi-agent Reinforcement Learning Using Transfer Learning

In this research, we propose a new autonomous bicycle sharing management system by local residents using the MARL (multi-agent reinforcement learning) model that adopts DQN (Deep Q-Network) with four stations as one group. In addition, we will define similar environments by assigning demand-based labels to stations in order to adapt to changes in the environment, such as the addition of more stations, and to confirm the effectiveness of efficient transfer learning. As a result of the experiment, the proposed model allowed multiple reinforcement learning agents to learn cooperative behavior and avoid a situation in which the number of remaining bicycles reaches zero. Furthermore, the performance of the model with and without transfer learning was compared, and the learning speed was higher when transfer learning was used, indicating the effectiveness of the model and the possibility of efficient service operation.