MERCI: Multi-agent reinforcement learning for enhancing on-demand Electric taxi operation in terms of Rebalancing, Charging, and Informing Orders

Abstract

The development of intelligent transportation systems is being driven by the increasing electrification and the Internet of Things. On-demand electric taxis (OETs) are seen as a potential way to meet personalized travel needs and improve transport efficiency. While research is being done to create a multi-agent reinforcement learning (MARL)-based framework to achieve intelligent operation, there are still challenges to be addressed, such as the balance between exploration and exploitation, and the non-stationary issue. This study proposes an ensemble MARL framework to manage the daily operations of OETs, such as rebalancing, charging and informing orders. To address the non-stationary issue caused by the dynamic nature of operations, a demand awareness augmented architecture is proposed to use order information to make better decisions. Experiments using real-world data in Shenzhen show the emergence of intelligence of the proposed framework during operation and its superiority over traditional greedy methods. Additionally, ablation studies demonstrate that the proposed framework outperforms basic MARL architectures.