Reinforcement Learning- Pac-Man
summary
Code and results: a working Pac-Man agent is implemented with a Q-learning on tiny map that reaches playable behavior after ~250 episodes and a Deep Q-network with Keras/TensorFlow trained with experience replay (memory size 5,000) and batch updates. Implementation uses Python and Jupyter notebooks, GPUs for training, and a bundled Pac-Man environment; deep models require thousands of episodes (one run ~5,000 episodes took ~26 hours on a GTX 1080) but ultimately beat larger maps and sometimes exploit reward quirks.
AI-written summary · about
chapters (19)
- 0:00Reinforcement learning intro
- 1:56Agenda
- 3:00Presenters
- 6:43What is RL
- 8:12Key RL terms
- 11:03Exploration vs exploitation
- 14:18Q-learning explained
- 25:11Setting up the coding environment
- 28:00Q-learning for Pac-Man
- 33:33Training Q-learning
- 42:28Deep Q-learning intro
- 44:00Supervised learning recap
- 47:11Neural networks for DQN
- 51:51Why DQN for Pac-Man
- 1:00:43Implementing deep Q-learning
- 1:01:58Training DQN model
- 1:17:25Evaluating agents
- 1:22:05Next steps in DQN
- 1:23:58Q&A
related talks
01h 01mReinforcement Learning: Pac-Man
01h 02mTeach your Pacman to play with ML and Reactive Streams
46mAI Takeover
49mML: Teaching machine to learn from its own mistakes (...)
56mAn Introduction to Reinforcement Learning - Snake Your Way Out of a Paper Bag
35mLet's play Space Invaders!
26mHacking Reinforcement Learning
28mDeep Reinforcement Learning: theory, intuition, code