|Date||Tuesday, August 2, 2011|
|Speaker||Neville Mehta, Post-Doctoral Candidate|
|School/Dept.||School of Electrical Engineering and Computer Science|
|Affiliation||Oregon State University|
Reinforcement Learning (RL)
Reinforcement Learning (RL) studies sequential decision-making in unknown environments. Straightforward RL is grossly inefficient due to the combinatorial explosion of the state and action spaces. My talk will briefly cover some of the divide-and-conquer approaches that leverage structure to mitigate the explosion. It will focus on a particular paradigm, hierarchical reinforcement learning (HRL), which combines state abstraction and structured policies to expedite learning. A major detraction of most HRL systems is the reliance on domain experts for the structural information. I will present my research that aims to redress this issue by autonomously discovering hierarchical structure from action models and trajectories. Empirical evaluations in multiple domains,including resource collection in the real-time strategy game of Wargus, demonstrate that the discovered hierarchies are comparable to manually engineered hierarchies and also facilitate significant speedup in learning when transferred to related domains.