Reinforcement Learning: Pac-Man
summary
Builds a Pac-Man agent with Python using TensorFlow/Keras and Jupyter on Azure, first as a table-based Q-learning agent (works on a 12-position map, converges in ~100–1,000 episodes, model files from kilobytes to ~3 GB) then as a neural deep Q-network with experience replay to handle 2.7×10^21 states; DQN trains for thousands of episodes (4,500 tested), uses decaying exploration (≈0.3 start), reward shaping, and a GTX 1080 Ti (~26 hours for full run) and finally beats medium maps reliably but needs more tuning on hardest levels.
AI-written summary · about
chapters (5)
- 0:00Start
- 9:26A few challenges of Reinforcement Leaming
- 13:50Reward shaping
- 41:00Experience Replay
- 43:33State to neural network input mapping
related talks
01h 31mReinforcement Learning- Pac-Man
46mAI Takeover
56mAn Introduction to Reinforcement Learning - Snake Your Way Out of a Paper Bag
01h 02mTeach your Pacman to play with ML and Reactive Streams
49mML: Teaching machine to learn from its own mistakes (...)
35mLet's play Space Invaders!
28mDeep Reinforcement Learning: theory, intuition, code
01h 17mTraining intelligent game agents using deep reinforcement learning