Open Source Summit + ELC North America 2020

Reinforcement learning (RL) is good at solving some hand-designed challenges. However, in real-life applications such as robotics, there are many edge cases that are impossible to design by hand, and some of them are too challenging to solve from scratch.

We introduce a novel class of RL algorithms (named POET) which combines the concept of open-endedness with RL. Such algorithms create increasingly diverse and challenging problems all by itself, while automatically forming a curriculum that teaches RL agents to solve those problems. This talk covers our most recent significant enhancements to POET, which is NEVER presented before: we enable POET to be applied to any general domains, introduces quantitative measurement of open-ended progress, and formulate a much more expressive environmental encoding. The combination of all of these innovation leads to the most open-ended demonstration so far: in a single run of the algorithm, the capability of agents increase along with the difficult levels/diversity of the created tasks, finally they solving extremely challenging tasks that cannot be solved by direct optimization or by learning through a hand-designed curriculum.