Scikit-decide: An Open-Source AI Framework for Reinforcement Learning, Planning, and Scheduling

Summary

Scikit-decide is an open-source AI framework for reinforcement learning, automated planning, and scheduling. Initiated at Airbus AI Research, it also received contributions through the ANITI and TUPLES projects and from ANU. The framework lets users describe a decision-making problem once and match it with compatible solvers, including cases where planning or scheduling problems can be addressed with reinforcement-learning solvers using graph neural networks. Its supported domains include Gymnasium environments, PDDL through unified-planning and plado, RDDL through pyrddl-gym, flight planning, scheduling, and toy environments. The solver catalog covers algorithms from Ray RLlib and Stable-Baselines3, action-masking adaptations, graph-based reinforcement learning with PyTorch Geometric, autoregressive approaches for parametric actions, planning and RDDL solvers, and a broad set of search and partially observable decision-making methods. It also integrates scheduling tools and optimization libraries such as OR-Tools, Gurobi, MiniZinc, and Optuna for hyperparameter studies. The project is designed to wrap existing state-of-the-art domains and solvers, remains community-extensible, and is released under the MIT license. Some integrations, including parts of the Ray RLlib migration, are still described as works in progress.