Multi-Agent Renforcement Learning for Dynamic Pricing and Fleet Management in Autonomous Mobility-On-Demand Systems

Abstract

This paper considers the joint fleet management and ride pricing problem faced by a profit-maximizing transportation service provider that operates a fleet of autonomous vehicles to serve the population’s urban mobility needs. Due to intractability issues of solving for the exact optimal real-time control policy, reinforcement learning-based solutions have been shown to perform well [1]. Existing works based on reinforcement learning do not scale well with the network size due to large state and action spaces. We study this issue by applying multi-agent reinforcement learning to develop a real-time policy, where each node acts as an agent. State abstraction allows us to represent the state information in a more compact manner, which reduces the dimension of the state space, while local decisions result in a lower dimension for the action space. We demonstrate the efficacy of the multi-agent reinforcement learning policy on an 8-node transportation network.