Adaptive Decision-Making Policies for Stochastic Electric Vehicle Routing with Time Windows Using Genetic Programming

Full abstract

Read the full abstract

Intelligent transportation systems increasingly rely on electric vehicle fleets operating under uncertain conditions, where routing decisions must adapt to changes in customer demand, service duration, and travel time while satisfying battery, charging, capacity, and time-window constraints. This work extends a genetic programming (GP)-based route generation scheme (RGS) to the stochastic electric vehicle routing problem with time windows (EVRPTW) and positions it as a downstream decision layer for sensor-enabled fleet management: progressively revealed operational conditions update the route state and are translated into online routing actions without repeated optimization. The experiments represent this information flow through controlled stochastic perturbations, enabling reproducible evaluation independently of a particular sensing configuration. The GP representation is extended with stochastic and global-state descriptors, and two capacity-aware RGS variants are introduced. Five RGS variants are evaluated across multiple stochastic scenarios for fleet size, energy consumption, and total tardiness. Limited parallelism performs best for fleet-size minimization, sequential construction for energy minimization, and parallel construction for tardiness minimization. Across the evaluated uncertainty scenarios, the identity of the best-performing RGS family for each objective remained largely unchanged, although absolute performance and dispersion varied non-monotonically. Structural analysis further showed that descriptors related to demand, energy, timing, and fleet state recurred more consistently in the analyzed policies than explicit variability descriptors. With an average online route-construction time of approximately 10 ms per 100-customer instance, the framework provides a computationally lightweight mechanism for converting sensor- and system-derived state updates into adaptive electric-fleet decisions under uncertainty.