The Dynamic and Stochastic Electric Vehicle Routing Problem (DS-EVRP) involves routing decisions under both dynamic and stochastic uncertainty. However, most DS-EVRP studies rely on deterministic assumptions or treat these uncertainties separately due to the computational limitations of traditional optimization methods, limiting their applicability in real-world electric vehicle (EV) operations, where uncertainty is inherently more pronounced than in conventional vehicle routing. While reinforcement learning (RL) has emerged as a promising alternative, its application to DS-EVRP remains limited, particularly in settings that jointly consider dynamic and stochastic elements. Moreover, the impact of reward design on learning performance remains underexplored. In our prior work, we compared two RL approaches based on tabular Monte Carlo control with different reward designs against a classical optimization method. The results demonstrated that both RL approaches outperformed the classical method, and that reward design plays a critical role in this performance: the proposed relative performance-based reward consistently outperformed the conventional reward scheme, yielding improved learning effectiveness and solution quality in dynamic and stochastic decision environments. Building on these findings, the present study extends the proposed reward mechanism to larger-scale DS-EVRP instances within a sequential decision-making framework, where tabular RL approaches face significant limitations due to rapid growth of the state-action space. To address this, we transition from tabular RL to function approximation via deep RL and integrate the proposed reward design within actor-critic architectures. Initial experiments indicate that this combination provides a promising direction for achieving robust, adaptive, and scalable routing decisions, directly relevant to real-world EV operations.

Keywords

Dynamic and Stochastic Electric Vehicle Routing Problem (DS-EVRP), Reinforcement Learning, Relative Performance-Based Reward