Sequential Decision making in probabilistic models

点击翻译按钮获取中文摘要

This proposal considers the problem of robust sequential decision making in non-linear environments. Reinforcementlearning has demonstrated high potential for solving complex problems in non-linear environments but has lackedefficiency and robustness. We argue that in order to deploy reinforcement learning agents in the real world, it is essential todevelop similar efficiency and robustness properties that have been developed in control theory. We propose to leveragethe extensive control and probabilistic reasoning literature to improve RL algorithms and present two interesting researchdirections. The first one considers using Sequential Monte-Carlo methods to improve planning for non-linearenvironments. The second direction focuses on designing robust controllers by exploring the connections betweenadversarial learning, robust control theory, and uncertainty modelling.