/ifkash.dev

ML systems, open source, and useful tools.

new game
Score
00
Record
00
Games
0
Speed

The agent runs fully client-side: the exported best.pth weights (a 28 → 256 → 256 dueling MLP) are loaded from src/lib/rl/snake-dqn-weights.json and stepped at ~10 actions/sec. The game rules, 28-feature state encoding, and greedy action selection mirror snake_env.py / state.py / play.py exactly.