This enables our method to learn from failed attempts where we do not solve the problem outright, but solve some of the subgoals. The results show that this approach helps the system learn faster and more efficiently, without sacrificing the quality of the policy.

This paper looks at a class of algorithms called policy tree search, which combines policies from reinforcement learning with traditional tree search. We show how one can decompose a problem into learnable subgoals, without any prior knowledge of the environment.

Thanks for these lists! Is the grumpy list an inside joke or am I missing something 🤣