RL Agents

This project focuses on generating datasets and training Reinforcement Learning agents on various Atari, Retro, and classic environments using Stable Baselines 3.

Project Structure

The project is organized as follows:

• run_atari_training.sh / run_retro_training.sh / run_classic_training.sh: Main entry points. Shell scripts that define a list of environments and launch training runs for each category.
• sb3_agents/trainer.py: The training implementation used by the run_*.sh scripts (PPO). Supports Atari environments (via ale-py) and Retro games (via stable-retro), with Weights & Biases (WanDB) integration for experiment tracking.
• sb3_agents/inference.py: Script for evaluating trained models and generating datasets. It loads the best model for a given environment, runs inference, and saves the resulting trajectories (states, actions, rewards) into .npz files.
• save/: Contains the best-performing model checkpoints (best_model.zip) organized by environment name (e.g., BreakoutNoFrameskip-v4, SonicTheHedgehog2-Genesis-v0).

Features

• PPO Implementation: Leverages Proximal Policy Optimization from stable-baselines3.
• Environment Support: Compatible with a wide range of:
  • Atari Games: Boxing, Breakout, Pong, etc.
  • Retro Games: SonicTheHedgehog2, SuperMarioBros3, etc.
• Experiment Tracking: Built-in integration with Weights & Biases for real-time monitoring of training metrics.
• Dataset Creation: Capabilities to record and export high-quality agent gameplay data.