Save the Princess
summary
Reimplements OpenAI’s evolution strategies to train a policy network without backprop, using Neanderthal for fast linear algebra, Cortex for the neural net and Onyx to distribute work; encodes game state as one long vector, single hidden layer feeding four action outputs, about 16,000 weights, perturbs a population with Gaussian noise (sigma), standardizes rewards and updates by weighted sum. Simulation built with re-frame, uses batches of ten runs, penalizes falling and long paths; converges slowly on a laptop but scales when distributed.
AI-written summary · about
chapters (8)
- 0:00Introduction to project
- 1:45Reinforcement learning basics
- 6:04Evolution strategies theory
- 7:59Neural network training
- 11:45Implementing the algorithm
- 21:45Massively parallel processing
- 29:21Policy network design
- 33:49Game mechanics and conclusion
related talks
47mHow we use Reinforcement Learning (...)
01h 17mTraining intelligent game agents using deep reinforcement learning
17mMatineh Shaker, Artificial Intelligence Scientist, Bonsai
49mML: Teaching machine to learn from its own mistakes (...)
37mTensorflow and deep reinforcement learning, without a PhD
20mAlpha2048
35mLet's play Space Invaders!
42mReinforcement Learning using Generative Models for Continuous State and Action Space Systems