Colloquium: Lessons on Learning and Planning with Compact Domain Models

Dr. Thomas Walsh, Postdoctoral Researcher, University of Arizona

Venue: 220K CRMS

Time: 4:00-5:00pm

Host: Dr. J. Goldsmith      

Abstract:          

In this talk, I describe how classical ideas from reinforcement learning can be applied to large environments that have compact representations.  Specifically, I will describe how methods for online exploration, apprenticeship learning, and planning can be adapted for use in environments with structured representations (dynamic bayes nets, STRIPS models, etc.).  With online exploration agents can often actively discover conditional structure and important parameters of their compact models with a relatively small number of samples.  In other domains, where active exploration truly is intractable, I will describe how the addition of a teacher who can demonstrate "proper" behavior, can result in tractable learning algorithms.  Finally, I will describe how advancements in sample-based planning make action selection in large grounded  state spaces feasible.                                 

Host:  Judy Goldsmith

Speaker's website:  http://www.cs.arizona.edu/~twalsh/