A reinforcement learning framework for improving parking decisions in last-mile delivery

This study leverages simulation-optimisation with a Reinforcement Learning (RL) model to analyse the routing behaviour of delivery vehicles (DVs). We conceptualise the system as a stochastic k-armed bandit problem, representing a sequential interaction between a learner (the DV) and its surrounding environment. Each DV is assigned a random number of customers and an initial delivery route. If a loading zone is unavailable, the RL model is used to select a delivery strategy, thereby modifying its route accordingly. The penalty is gauged by the additional trucking and walking time incurred compared to the originally planned route. Our methodology is tested on a simulated network featuring realistic traffic conditions and a fleet of DVs employing four distinct lastmile delivery strategies. The results of our numerical experiments underscore the advantages of providing DVs with an RL-based decision support system for en-route decision-making, yielding benefits to the overall efficiency of the transport network.