Analyzing multi-agent training strategies for enhanced EV charging recommendations

The global popularization of Electric Vehicles (EVs) has brought great challenges for charging network. Appropriate charging recommendations therefore plays an important role in balancing charging demand with charging supply systems. In recent years, researchers have attempted to address this issue through multi-agent reinforcement learning (MARL) algorithms. However, adapting a single algorithm to suit multi-agent charging recommendations remains a significant challenge due to the complexity of coordinating multiple agents, handling the high dimensionality of the state-action space, and ensuring stability and convergence. The present work proposes a cooperative multi-agent method utilizing an enhanced Multi-Agent Cooperative Soft Actor-Critic (MAC-SAC) algorithm to tackle the complexity of charging station recommendations. Methodologically, we investigate different multi-agent reinforcement learning training approaches and highlight their respective pros and cons. The results indicate that the approach utilizing shared policy and training networks yields the best results in multi-agent charging recommendation, while other training approaches exhibit certain drawbacks, such as adversely impacting algorithm convergence and stability. Additionally, other approaches require longer training times per iteration and have lower utilization rates of experience. Our research highlights the potential of improving cooperative multi-agent methods in addressing charging recommendation challenges, providing crucial insights for the advancement of EV charging recommendation tasks.