Jowa Football AI v2
6v6 · DQN Reinforcement Learning · Formation 3-2-1
Previous
2026
Open Source
A newer version is out: v3 — full 11v11 with two AI brains, and apps for Windows and Linux.
What's New in v2
Upgraded from 3v3 to 6v6 with Formation 3-2-1
AI replaced from Q-table to DQN (Deep Q-Network) using PyTorch
Shared replay buffer with capacity 200,000 transitions
Player selection system — pick any player with ↑↓ then Enter
Auto epsilon decay with periodic model saving every 3600 frames
CUDA support — runs on GPU if available
Game Info
| Mode | 6v6 (Red vs Blue) |
| Formation | 3-2-1 |
| AI Type | DQN — Deep Q-Network (PyTorch) |
| Actions | 9 — Stop, N, S, E, W, NE, NW, SE, SW |
| Resolution | 800 × 600 |
| FPS | 60 |
| Device | CUDA (GPU) or CPU — auto-detected |
| Model Files | football_model_red.pth, football_model_blue.pth |
Requirements
Python 3.8+
Required to run the script
pygame
pip install pygame
PyTorch
pip install torch
NumPy
pip install numpy
How to Run
1
Install dependencies
pip install pygame torch numpy
2
Download and extract the ZIP
Extract jowa-football-ai.zip to any folder.
3
Run the game
python jowa-football-ai-with-players_PLAYMODEAI4_6p.py
Controls
Player Selection
↑ / ↓ — cycle through players | Enter / Space — control selected player | ESC — back to selection
Movement (after selecting a player)
W Up | S Down | A Left | D Right
↑ Up | ↓ Down | ← Left | → Right