Code · open source
Pac-Man DQN
An agent that learned Ms. Pac-Man from raw pixels, and more than tripled its score.
- 1,668mean score, vs 492 untrained
- 5 / 5evaluation seeds improved
- 2.36Mdecisions trained in 112 min
How it works
The agent sees four 84×84 grayscale frames and picks one of 9 moves. Starting from a course notebook, I added Double DQN, 3-step returns, a life-loss signal and a 100,000-frame replay buffer stored 5× more compactly. A baseline run (1,000 episodes) and a tuned run (3,000 episodes) were compared, with checkpoints picked on separate validation seeds.