Joint taxi dispatch and station pricing optimization: A game-theoretic multi-agent reinforcement learning approach via alternating training

Electric taxis play dual roles as both mobility service providers in the transportation network and dispatchable loads within the power grid. Achieving synergistic optimization between dispatching profitability and charging replenishment has become a key challenge affecting the overall efficiency of these spatiotemporal coupled systems. To address this issue, this paper proposes a game-theoretic multi-agent reinforcement learning (MARL) framework via alternating training for joint taxi dispatching and station pricing optimization (JDPO). At the vehicle layer, a dispatching policy is developed based on a single-step policy optimization (SSPO) algorithm, which exploits taxi homogeneity to bypass long-horizon value estimation. At the station layer, we introduce a Transformer-enhanced minimax multi-agent double-delayed deep deterministic policy gradient (M3TD3) algorithm to achieve robust charging and pricing optimization offered by charging stations. Optimal power flow (OPF) and locational marginal prices (LMPs) are incorporated to couple station pricing with power grid constraints. The agents in the two layers share global environmental states and optimize their respective dispatching and pricing policies through an alternating training mechanism. Experimental results demonstrate that the proposed method exhibits superior performance in terms of service success rate, passengers waiting time, and charging station revenue.