Using Reinforcement Learning to Solve a Dynamic Orienteering Problem with Random Rewards Affected by the Battery Status

This paper discusses an orienteering optimization problem where a vehicle using electric batteries must travel from an origin depot to a destination depot while maximizing the total reward collected along its route. The vehicle must cross several consecutive regions, with each region containing different types of charging nodes. A charging node has to be selected in each region, and the reward for visiting each node-in terms of a 'satisfactory' charging process-is a binary random variable that depends upon dynamic factors such as the type of charging node, weather conditions, congestion, battery status, etc. To learn how to efficiently operate in this dynamic environment, a hybrid methodology combining simulation with reinforcement learning is proposed. The reinforcement learning component is able to make informed decisions at each stage, while the simulation component is employed to validate the learning process. The computational experiments show how the proposed methodology is capable of design routing plans that are significantly better than non-informed decisions, thus allowing for an efficient management of the vehicle's battery under such dynamic conditions.