Learning Curricula in Open-Ended Worlds
Minqi Jiang (Ph.D. Student)
Adaptive curricula have played a pivotal role in successfully applying deep reinforcement learning (RL) to the most challenging domains. By presenting RL agents with variations of the external world that best challenge their present capabilities, adaptive curricula can greatly improve how quickly agents learn, as well as the quality of the learned behavior. My research focuses on developing effective and theoretically-grounded methods for generating adaptive curricula to produce robust agents capable of succeeding across as many variations of the environment or task of interest as possible. By producing increasingly robust agents via adaptive curricula in open-ended environments, whose variations form an unbounded space of challenges, my work seeks to kickstart a co-evolutionary process between agents and environments that leads to generally capable AI.
|Primary Advisor:||Laura Toni (University College London)|
|Industry Advisor:||Tim Rocktäschel (University College London & Meta AI)|
|PhD Duration:||01 April 2022 - 01 October 2023|