Modeling of Bicycle Sharing Operating System with Dynamic Pricing by Agent Reinforcement Learning

The objective of this research is to encourage users participating in bicycle sharing schemes to return the bicycle to a specific station in return for an incentive. In addition, we aim to verify the business operation model, which differs fundamentally from the conventional operation based on truck allocation. We employed a reinforcement learning technique for directions to user agents to autonomously respond to uncertain events. Reinforcement learning, which is assigned a value, is calculated by function approximation using a hierarchical neural network. In the hierarchical neural network, the network configuration was arbitrarily determined to be fully connected and to contain 3 hidden layers and 16 nodes. Function approximation is considered to be an effective method because there are huge combinations of states and actions. The results indicated that the proposed model reduced the cost by approximately 7% compared to the conventional truck operation. This result suggests the usefulness of autonomous system operation by using reinforcement learning.