Comparing Behaviour Trees and Deep Reinforcement Learning for Fighting AI
-
Updated
Mar 30, 2025 - Python
Comparing Behaviour Trees and Deep Reinforcement Learning for Fighting AI
RL environment for the 2005 GBA game 'Hot Wheels Stunt Track Challenge'
In-house Mortal Kombat tournament environment based on stable-retro
An agent that learns to play NES games from pixels and sound alone — the emulator's memory is off limits to the policy. Motion-based object tracking, on-screen counter reading, world model, planner, cross-game transfer.
Training engine where the Laya decision model learns to beat retro games: text game state, group-relative policy gradients, per-state checkpoints and a reverse curriculum.
Cookiecutter template for Retro Speedlab game packages: a skeleton that scheduled Claude Code lab runs grow until Laya beats the game.
A reinforcement learning A3C implementation trained to play Super Mario Bros
Train a PPO agent to play Super Mario Bros using Stable-Baselines3 on stable-retro.
Deep Q-Networks for Super Mario Bros 3, from a beginner-friendly basic implementation to advanced Rainbow-style improvements.
PPO agents for games
To associate your repository with the stable-retro topic, visit your repo's landing page and select "manage topics."