// the find
hubbs5/or-gym
Environments for OR and RL Research
A collection of OpenAI Gym-compatible environments for classic operations research problems — knapsack, bin packing, TSP, vehicle routing, multi-echelon supply chain, portfolio optimization. The idea is to give OR researchers a standard RL interface so they can compare RL agents against traditional solvers like integer programming. Backed by a 2020 arXiv paper from Carnegie Mellon and Dow Chemical researchers.
Covers a meaningful breadth of OR problem classes in one installable package — not just toy problems but practically relevant ones like multi-echelon inventory with backlogs and VRP with time windows. The Gym API compatibility means any RL library (RLlib, Stable-Baselines, etc.) works without adaptation. Action masking support in the knapsack environments is a real detail that matters for combinatorial problems where invalid actions dominate. Grounded in actual research papers, so the environment formulations aren't arbitrary.
Dead since October 2023 and pinned to OpenAI Gym, which was deprecated in favor of Gymnasium — any serious new project will hit compatibility issues immediately since the ecosystem has largely moved on. The environment suite is shallow: TSP-v0/v1 share the same codebase with trivial parameter differences, and there are no continuous-action OR environments. Test coverage is minimal — one env_test.py that appears to just instantiate environments without validating reward semantics or edge cases. No support for custom problem instances beyond constructor parameters, which limits use for anything but the exact problem configurations the authors cared about.