Demand-Responsive Transport Dynamic Scheduling Optimization Based on Multi-agent Reinforcement Learning Under Mixed Demand

Demand-Responsive Transport (DRT) is an innovative mode of public transportation that focuses on individual passenger needs by offering customized transportation solutions. Most prior researches rely on historical passenger flow to generate static schemes and lack the optimization of optional dynamic demands from the perspectives of passengers and transportation agencies simultaneously. Therefore, this paper addresses the dynamic scheduling optimization problem of DRT under mixed demand, minimizing overall system costs and ensuring equitable passenger waiting times. We initially construct a dual-objective optimization model for DRT dynamic scheduling to solve this. Subsequently, we propose the Action-Refinement Multi-Agent Dueling Double Deep Q-Network (AR-MAD3QN) algorithm to tackle the challenge of simultaneous route optimization for a fleet of vehicles considering static and optional dynamic passenger demands under dynamic road conditions. Additionally, the action-refinement module improves the network structure of MAD3QN, preventing the generation of invalid and unstable actions and improving training efficiency. Experiments are conducted on the Sioux Falls network, with the AR-MAD3QN algorithm compared against baseline algorithms in different settings. The results show that our AR-MAD3QN algorithm exhibits superior optimization with faster and more stable convergence.