A reinforcement learning-based metaheuristic algorithm for on-demand ride-pooling

In this paper, we propose a ride-pooling algorithm which selects the parameters of a heuristic dispatching procedure using reinforcement learning. This meta-heuristic approach is applicable for the selection of arbitrary parameters of a centralized dispatching algorithm and is aimed to find sequences of decisions to optimize target function over the simulation period, to capture dynamics of demand and supply in the ride-pooling system. A reinforcement learning environment is implemented in SUMO traffic simulation software. The experimental study compares the results of the metaheuristic with the strategies based on the exhaustive search and shows that the proposed approach is able to find a near-optimal solution with advantage of 2%-10% compared to baselines.