Robustness in reinforcement learning
Robustness in RL is a relevant research topic that has attracted the interest of several communities. The overaching idea is to design optimal RL policies that are able to withstand uncertainty in the environment, such as distributional shift and measurement noise. To address this issue, we have developed a notion of robustness and proposed an algorithm that leverages lexicographic optimisation to obtain a robust policy according to our definition. Relevant paper is:
- D. Jarne, L. Romao, L. Hammond, M. Mazo Jr, A. Abate, “Observational Robustness and Invariances in Reinforcement Learning via Lexicographic Objectives”. Technical report.