Our NIPS 2017: Learning to Run approach – ML Review – Medium

For 3 months, from July to 13 November (sometimes with long breaks though), me and my friend Piotr Jarosik participated in the NIPS 2017: Learning to Run competition. In this post we will describe how it went. We release the full source code. Competitors were given a model of a human skeleton and OpenSim simulator. The task was to write a program that activates legs muscles in order to maximize the number of meters passed in 1000 timesteps. A typical program will read the initial observation and output an action, activating some muscles. The simulator will update its state and give us the observation for the next timestep a reward (distance passed) for the previous timestep.