Since a few years it is possible to learn playing Atari games directly from pixels with reinforcement learning. A more recent discovery was that evolution strategies are competitive for training deep neural networks on those tasks. And even a simple genetic algorithm can work.
Those methods are fun to play around with. I have created a maze-world \u2026<\/p>","content":"\n\n\n\n
Since a few years it is possible to learn playing Atari games directly from pixels with reinforcement learning<\/a>1<\/a><\/sup>. A more recent discovery was that evolution strategies are competitive<\/a>2<\/a><\/sup> for training deep neural networks on those tasks. And even a simple genetic algorithm<\/a>3<\/a><\/sup> can work.<\/p>\n\n\n\n\n Those methods are fun to play around with. I have created a maze-world and trained a neural network to explore it. This task is much simpler than Atari but still gives interesting results.<\/p>\n I’m using Covariance-Matrix Adaptation (CMA<\/span>-ES<\/span>) and have tried a few simpler methods. The focus is on results but code is available.<\/p>\n\n\nMaze<\/h2>\n