Moskovitz, Theodore Harris; (2025) Structure, Learning, & Composition: Multitask Reinforcement Learning in Brains and Machines. Doctoral thesis (Ph.D), UCL (University College London).

Preview

Text Moskovitz_10211485_Thesis.pdf Download (28MB) | Preview

Abstract

This thesis centers around different forms of common structure that can be shared across tasks faced by reinforcement learning agents and how these types of structure can be leveraged to both learn new behavioral policies more efficiently and compose existing policies. Specifically, the first part of this thesis is concerned with how agreement among the optimal policies for some group of tasks constitutes a form of behavioral structure. This structure can be used as the basis for a regularized policy optimization approach to speed up policy learning on new tasks. One such approach proves to be an effective model of a number of animal and human behavioral patterns observed in neuroscientific studies of dual process theories of cognition. The second part of this thesis focuses on how consistent environmental transition dynamics across tasks can be exploited by agents to learn state representations which facilitate efficient policy evaluation and composition. In particular, prior work on this topic is extended to include several forms of biologically-inspired, non-Markovian, non-stationary reward functions, with applications to both machine learning and natural behavior.

Download activity - last month

Export as XMLCSVJSON

Download activity - last 12 months

Export as CSVXMLJSON

Downloads by country - last 12 months

Export as JSONXMLCSV

Archive Staff Only

View Item