Visiting Speaker: Tim Kovacs
ACCS Special Seminar
This week we have a special seminar linking in via a
video conferencing to the ACCS Node at ADFA to hear
Tim Kovacs who is a leader in the field of Learning
Classifier Systems (LCS).
Place: Room 621, GP South (Building 78)
Time: Wednesday 30th November, 9:45 morning Tea.
10:00am seminar (Brisbane time)
Title: "Strength or Accuracy? Credit
assignment in Classifier Systems""
Presenter: Tim Kovacs
Department of Computer Science, Bristol University
http://www.cs.bris.ac.uk/~kovacs/author.directory/publications.html
Abstract
This talk reviews some work on the problem of
crediting individual components of a complex
adaptive system for their often subtle effects on
the world. For example, in a game of chess, how did
each move (and the reasoning behind it) contribute
to the outcome? Application of adaptive methods,
whether to classification or control tasks, requires
effective approaches to this sort of credit
assignment problem.
A fundamental approach is to evaluate components of
solutions, rather than complete solutions, with the
intention of simplifying the credit assignment
problem. This is the approach taken by Michigan
Learning Classifier Systems, which combine
Evolutionary Algorithms, to generate solution
components, and Reinforcement Learning methods to
evaluate them.
Unfortunately, as will be outlined, serious
complications arise from this attempt at
simplification. Most significantly, it will be shown
that both of the main approaches (strength-based and
accuracy-based systems) have difficulties with
certain tasks which the other type does not. The
talk will also outline the causes of the main
difficulties each type of system faces, the types of
tasks which cause these ifficulties, and prospects
for addressing them.