Haeone Lee

I am interested in building intelligent agents that can self-improve to be useful for humans. Specifically, it should generate useful problems and solve them by leveraging prior knowledge with critics to validate the success. I believe in the power of Reinforcement Learning, in that sense (1) it can autonomously come up with the solution given the goal (2) it interacts with and adapts to the changing world (3) it is the closest to how animals ‘emerge’ the intelligence as part of goal pursuit. To make RL successful, I deem there are plenty of challenges to solve such as sample efficiency, long-horizon control, and safe and autonomous learning. I believe that utilizing prior knowledge(e.g., common sense, offline data), and equipping the algorithms with long-term memorizing, hierarchical decision-making, and good abstraction can help to achieve my goal. For details, this briefly surveys my thoughts.