Iterative Decision-Making in Reinforcement Learning for Ride-Sharing Optimization: A Joint Approach to Dispatching and Rebalancing

This study proposes a reinforcement learning (RL)based decision-making framework to optimize operations in ridesharing (RS) systems. The framework simultaneously addresses two core challenges: passenger-vehicle dispatching and idle vehicle rebalancing under dynamic and spatially imbalanced demand conditions. A Deep Q-Network (DQN) is employed within a centralized architecture, complemented by an iterative Q-valuebased action selection strategy that accounts for inter-vehicle dependencies. To enhance computational efficiency, the framework incorporates feasibility-based action masking. The reward function integrates operational efficiency, service quality, rebalancing effectiveness, and environmental sustainability to support multi-objective learning. Experimental results across four spatial demand scenarios demonstrate that the proposed approach improves system throughput and passenger experience compared to baseline methods. Additional analysis confirms the effectiveness of rebalancing in reducing detour time and increasing request acceptance rates. The framework also shows robust learning convergence and policy generalization in both single- and multivehicle settings, highlighting its scalability and practical applicability to real-world, large-scale RS environments.