pokemon-rl

an end-to-end deep-RL pipeline that trains agents to autonomously play Pokémon Red. custom Game Boy emulator (rlgb) + Stable Baselines 3 PPO, CUDA acceleration, and 64 parallel training environments.

architecture

• core emulator — rlgb at 3,200+ fps
• 64 parallel envs — multi-process pool, shared-memory IPC
• PPO — Proximal Policy Optimization via Stable Baselines 3
• CUDA — GPU-accelerated inference and training
• reward shaping — badges, exploration, levels
• deterministic checkpoints — reproducible runs

stack

ml: python, pytorch, stable-baselines3, gymnasium
emulation: c (rlgb) + python bindings
infra: cuda, multiprocessing, docker

results

• autonomous map exploration without explicit guidance
• gym-badge progression in near-optimal order
• team composition and leveling strategy
• stable convergence across seeded runs