Kernel Reinforcement Learning
Matthew Robards (SoCS CECS)
CS HDR MONITORING AI Research GroupDATE: 2010-04-16
TIME: 15:00:00 - 15:30:00
LOCATION: RSISE Seminar Room, ground floor, building 115, cnr. North and Daley Roads, ANU
CONTACT: JavaScript must be enabled to display this email address.
ABSTRACT:
Traditionally in reinforcement learning one must either hand engineer ones own features, or use awkward coding methods such as tile coding or RBF coding. Using the former almost certainly results in an agent which generalizes poorly across problems, whereas using the latter results in enormous redundancy in the feature vector.
In the broader area of machine learning kernel methods have become the standard way of automatically linearizing the state space, however they have been seldom explored in RL. This talk will present a novel kernelized SARSA $(lambda)$ algorithm. We present a very intuitive formula for the eligibility trace along with a method for achieving compact represetations of the value function.
BIO:
PhD student in Computer Science.
