ORL: Reinforcement Learning Benchmarks for Online Stochastic Optimization Problems

Abstract

Reinforcement Learning (RL) has recently achieved state-of-the-art performance in a wide variety of domains: from robotics, to gaming, to traffic control. The domain of Operations Research (OR) is particularly amenable to RL approaches, because many of the canonical problems can be characterized as online stochastic optimization problems where the distribution of data is unknown. While there is a nascent literature at the intersection of RL and OR, there are no commonly accepted benchmarks which can be used to compare proposed approaches rigorously in terms of performance, scale, or generalizability. This paper aims to fill that gap by introducing open source OR+RL benchmarks for three canonical OR problems with a wide range of practical applications: Bin Packing, Newsvendor, and Vehicle Routing. We apply both well-known OR approaches and newer RL algorithms to these problems and analyze results. For each of these problems, we find that RL is competitive with or superior to the OR baselines, pointing the way for future theoretical work and highlighting RL's immediate potential utility in a host of real-world problems.