Be like water: Adaptive Subgoal Search
Reinforcement learning involves training agents to interact with an environment through a feedback loop. To optimize performance, the agent must adapt to various information streams, such as environment changes, experience, uncertainty, and task complexity. In this talk, we will show how the concept of adaptivity can be applied in the context of subgoal search. The discussion will also cover open questions and potential future research directions.
Bio
Łukasz Kuciński is a Research Scientist at the Polish Academy of Sciences (PAS), where he leads a machine learning group. His research focuses on reinforcement learning and optimization, with publications in top machine learning venues. In addition to his research, Łukasz is actively involved in education, teaching courses on reinforcement learning and leading seminars on machine learning at the University of Warsaw. Prior to his current role at PAS, he served as the vice-director at the Polish Financial Supervision Authority, where he spearheaded the implementation of Solvency II internal models in Poland and the European Union. Łukasz holds a PhD and MSc in mathematics from PAS and the University of Warsaw, respectively, with a focus on optimization problems in control theory, stochastic processes, and game theory.