Tensorflow and deep reinforcement learning, without a PhD
summary
Trains a policy-gradient agent in TensorFlow to play Pong from raw pixels using frame differences, a tiny two-layer dense network and sampling of actions via tf.multinomial. Uses batches of 20,000 moves, rewards of ±1 with time discounting, RMSProp optimizer and backprop on the collected trajectories. The model learns to beat a scripted paddle, shows neurons that track the ball, and generalizes to examples like a pancake-flipping robot and DeepMind locomotion; the implementation is simple and empirically effective.
AI-written summary · about
chapters (16)
- 0:00Introduction
- 6:13Cross entropy loss
- 10:03Data set
- 11:44Tensorflow
- 12:26Sample operation
- 13:40Placeholders
- 14:11Observations
- 15:06Lowlevel Tensorflow
- 16:05Training
- 20:35Visualization
- 21:07Examples
- 23:05Pancake flipping robot
- 24:27Skeleton walking robot
- 27:52Alphago
- 29:26Neural network architectures
- 33:03Questions
related talks
50mBuilding a Pong playing AI in just 1 hour (plus 4 days training...)
01h 17mTraining intelligent game agents using deep reinforcement learning
49mML: Teaching machine to learn from its own mistakes (...)
01h 01mReinforcement Learning: Pac-Man
56mCan I do AI?
01h 47mGamified Reinforcement Learning: Mastering AI through Video Games
31mDesigning game AI with reinforcement learning
39mTF-Agents: A Flexible Reinforcement Learning Library for TensorFlow