PWhiddy/PokemonRedExperiments

What it solves

This project provides a framework and scripts to train reinforcement learning (RL) agents to play the video game Pokemon Red. It allows users to observe how an AI learns to navigate the game world and progress through the story.

How it works

The project uses RL agents (via Stable Baselines 3) and the PyBoy emulator to interact with the game. It includes two versions of training scripts: an original baseline and a "V2" version that trains faster, uses less memory, and utilizes a coordinate-based exploration reward to help the agent reach further in the game, such as Cerulean City.

Who it’s for

Researchers and enthusiasts interested in reinforcement learning, game AI, and the application of RL to complex, tile-based environments.

Highlights

• Pretrained Models: Includes the ability to run a pretrained model interactively to see the AI in action.
• Training Broadcast: A wrapper that allows users to stream their agent's training progress to a shared global game map.
• V2 Improvements: A coordinate-based exploration reward system that improves training efficiency and reach.
• Visualization: Tools for tracking progress via TensorBoard, Weights & Biases, and static map visualization.