Monday
January 14
4:00 - 4:50 PM 
Kelley 1001

 

Soumya Ray 
Postdoctoral Researcher
School of Electrical Engineering & Computer Science
Oregon State University

 

Knowledge Transfer in Reinforcement Learning 

Humans are remarkably good at using knowledge acquired while solving
past problems to efficiently solve novel, related problems. How can we
build agents with similar capabilities? In this talk, I focus on
"reinforcement learning" (RL)---a setting where an agent must make a
sequence of decisions to reach a goal, with intermittent feedback from
the environment about the cost of its current decision. I describe two
approaches that allow agents to leverage experience gained from solving
prior RL tasks. In the first approach, the agent learns a hierarchical
Bayesian model from previously solved RL tasks that it uses to quickly
infer the parameters of a novel RL task. In the second approach, the
agent learns a hierarchical task-subtask decomposition from the solution
of a previous task, and uses the decomposition to learn more quickly on
a novel task. I present empirical evidence on two RL problem domains,
maze-navigation and a resource collection problem in the real-time
strategy game, Stratagus, that show that leveraging experience from
prior RL tasks improves the rate of convergence to a solution in a new
task. 

This is joint work with Aaron Wilson, Neville Mehta, Alan Fern, Prasad
Tadepalli and Tom Dietterich at Oregon State University under DARPA
grant FA8750-05-2-0249. 

Biography

Soumya Ray obtained his baccalaureate degree from the Indian Institute
of Technology, Kharagpur, and his doctorate from the University of
Wisconsin, Madison in 2005. Since 2006, he has been a postdoctoral
researcher in the machine learning group at Oregon State University. His
research interests are in statistical machine learning, reinforcement
learning and bioinformatics.

_______________________________________________
Colloquium mailing list
[email protected]
https://secure.engr.oregonstate.edu/mailman/listinfo/colloquium

Reply via email to