Title: Option Discovery in Hierarchical Reinforcement Learning

Speaker: Aditi Mavalankar (UCSD)

Abstract: Option discovery is one of the core challenges in reinforcement learning (RL). Most prior work in hierarchical RL involves a predefined set of options provided to an agent, which then uses these options in conjunction with primitive actions to learn a policy maximizing the long-term return. However, these options can also be composed in order to generate a combinatorially large number of behaviours in a zero-shot manner, as proposed by Barreto et al*. In this talk, I will first cover the background on options and zero-shot composition of options. I will then discuss some of our recent work on leveraging this idea of composing options to address the problem of option discovery. I will also demonstrate the malleability of the discovered options in their ability to achieve high reward, as well as facilitate better exploration.

*The Option Keyboard: Combining Skills in Reinforcement Learning, Barreto et al, NeurIPS 2019.