Colosseum is a pioneering Python package that creates a bridge between theory and practice in tabular reinforcement learning with an eye on the non-tabular setting.

If you are new to the package, extensive explanations and tutorials are available at project page.

Core capabilities

• The computation of three theoretical measures of hardness for any given MDP.
• Empirical study of the properties of hardness measures.
• Principled benchmarking for tabular algorithms with rigorous hyperparameters optimization.
• Non-tabular versions of the tabular benchmark for which tabular hardness measures can be computed.
• Extensive visualizations for MDPs and analysis tools for the agents' performances.