This library provides a simple, reproducible framework for experimenting with Reinforcement Learning (RL) in Python. It targets researchers and students needing a straightforward tool to test RL algorithms, offering ease of use and clear experiment tracking.

How It Works

The framework is built around core RL components: Agents, Markov Decision Processes (MDPs), and Experiments. It allows users to define custom MDPs or use provided ones (like GridWorld) and implement various agents (e.g., Q-Learning, Random). A key feature is the run_agents_on_mdp function, which orchestrates agent-environment interaction, collects results, and generates plots. The "easy reproduction of results" feature saves experiment configurations to files, enabling quick re-runs and verification of results.

Quick Start & Requirements

• Install with: pip install simple_rl
• Requires: numpy, matplotlib. pygame for visuals.
• Supports OpenAI Gym environments.
• Documentation: https://simple-rl.readthedocs.io/en/latest/
• Examples: https://github.com/david-abel/simple_rl/tree/master/examples

Highlighted Details

• Built-in reproducibility feature (reproduce_from_exp_file) for tracking and re-running experiments.
• Includes implementations for Value Iteration and Monte Carlo Tree Search (MCTS) (MCTS is in development).
• Supports basic agents like Q-Learning, RMax, and Random agents.
• Offers basic visualization capabilities via Matplotlib and Pygame.

Maintenance & Community

• Primarily maintained by David Abel.
• Open to contributions; specific areas for improvement are listed in the README.

Licensing & Compatibility

• MIT License.
• Permissive for commercial use and integration with closed-source projects.

Limitations & Caveats

The reproducibility feature currently only supports standard MDPs and does not extend to OO-MDPs, POMDPs, or Markov Games. Deep RL algorithms (like DQN) and advanced planning methods (like RTDP) are listed as "In Development."