Dan Ashlock, University of Guelph, Canada
Intended audience: Anyone producing strategies for games or agents
with a computational intelligence. While the topics of the tutorial
are founded on advanced mathematical techniques, the resulting
research methods require only a knowledge of computational
intelligence. The emphasis of the tutorial is on results and
techniques, not the mathematical foundations.
Description:
One of the down sides to evolving or training a game strategy on a
series of examples is that the resulting strategy often defies simple
analysis. This tutorial introduces a number of related helpful
techniques for understanding artificial agents or strategies produced
via computational intelligence techniques. The key tool is the
development of meaningful numerical signatures for the behavior of
agents that can then be clustered, classified, or visualized. These
techniques include:
Competative Analysis. This analysis technique checks to see how
well agents perform against one another. There are a number of
types of competative analysis, one-on-one, team-on-team, and
tournament rank analysis. These techniques permit a researcher to
understand the capabilities of an agent but are not as good at
understanding its behavior.
Fingerprinting is a method for testing simple game playing agent
against an infinite panel of opponents by using Markov chain
theory. Both advantages and shortcomings of fingerprinting will be
presented.
Dual metric space theory. This techniques induces a
behaviorally-based distance structure on both agents and boards
(instances of play). The technique generalizes far beyond the
analysis of agents and game boards but is especially useful in this
context. Dual metric space structures are useful in instances when
fingerprinting is too computationally difficult.
Applications of the techniques presented in this tutorial include:
(i) Behavioral sorting of agents so that behavioral groups can be
established and representatives of those groups subject to intensive
analysis.
(ii) Documentation that modifications of training techniques yield
a change in the part of the strategy space explored.
(iii) The ability to make representation-independent measurements
of the similarity or differences of agents.
Some techniques for analysis of high-dimensional data that are useful
for reducing fingerprint or dual metric space data to useful
information are also presented.
Presenter:
Dr. Daniel Ashlock is a Professor of Mathematics at the University of
Guelph, is a Senior Member of the IEEE, and is the author of over 160
peer reviewed scientific articles and two books. He serves as an
associate editor of the Transactions on Computational Intelligence in
Games and is a member of the IEEE Games Technical committee. His
research focus is the understanding of evolution as an algorithmic
process. This includes the analysis of the results of evolutionary
computation beyond simple performance measures, the focus of this
tutorial.
