Our NIPS 2017: Learning to Run approach
Nov 19, 2017 · 12 min read
For 3 months, from July to 13 November, me and Piotr Jarosik participated in the NIPS 2017: Learning to Run competition. In this post we describe how it went. We release the full source code.
tl;dr 22nd place in the end, the final skeleton has a cheerful gait, we used PPO trained on 80 cores in a couple of days with manually prepared observation vector + a bit of reward hacking. The final result: