Aprendizaje profundo de agentes múltiples por refuerzo con estrategia evolutiva para la distribución de vehículos de carga móviles

The dynamic dispatching of mobile charging vehicles (MCVs) has emerged as a promising solution to overcome the limitations of fixed charging infrastructure in electric vehicle (EV) services. However, the stochastic nature of user charging demands and the high operational costs of MCV travel pose significant challenges in balancing supply and demand, particularly under real-time requests and large-scale traffic networks. Existing MCV dispatch strategies are often static, reliant on predefined demand locations, and fail to adapt to the spatiotemporal variations in EV demand and MCV states. To address these challenges, we formulate the MCV dispatching problem as a Markov decision process (MDP) and propose multiagent deep reinforcement learning (MARL)-evolutionary strategy (ES), a novel MARL framework with centralized training and decentralized execution (CTDE), enhanced by integrated action-space ESs. MARL-ES leverages a neural network-based state-value function that incorporates the experiences of MCVs and EV demands, as well as offline historical data. It further employs mutation and segment-based crossover operators to enhance the strategic diversity, improve exploration, and optimize dispatch decisions. We develop a realistic, large-scale city grid simulator to model stochastic EV arrivals, energy constraints, and coordinated MCV interactions. Experimental results show that MARL-ES significantly outperforms static optimization and conventional MARL approaches in total profit, operational cost, and MCV moving distance. The framework also demonstrates robust scalability and adaptability across different fleet sizes and under uncertain demand and energy conditions, offering a practical and adaptive dispatching solution for intelligent mobile charging services (MCSs).