Building a Pong playing AI in just 1 hour (plus 4 days training...)
summary
Claim: build an AI to master Pong Pong in 1 hour (plus 4 days training...) using concrete methods: Q-learning, deep nets and Tensorflow. Implements tiny games — half‑pong 40×40 and mini‑pong — with binary frames, experience replay and epsilon‑greedy exploration. Code and checkpoints are on GitHub and slides link; training runs on CPU but benefits from GPU; convolutional nets with pooling boost score to roughly mid‑60s versus random baseline.
AI-written summary · about
chapters (10)
- 0:00Introduction
- 7:07Reinforcement learning basics
- 12:40Q-learning with maze
- 17:24Q-learning in TensorFlow
- 23:51Q-learning for Pong
- 27:23Deep learning for Pong
- 33:30Reinforcement learning libraries
- 39:54Debugging training
- 46:15Scaling to Atari games
- 49:18Q&A
related talks
37mTensorflow and deep reinforcement learning, without a PhD
56mCan I do AI?
46mAI Takeover
49mML: Teaching machine to learn from its own mistakes (...)
39mBuilding self evolving video game AI in Python
35mHow I Built an AI to Beat the 1990s Arcade Game Double Dragon
17mTraining My Rival in Java: A Deep Q-Learning AI to Play Azul
15mAI Self - Learning Game Playing