Biasing evolving generations in learning classifier systems using information theoretic measures
Karthik Kuber, Chilukuri Krishna Mohan · 2009
This paper presents information-theoretic ideas to modify the course of evolution in Learning Classifier Systems. This approach exploits the possibilities that individuals in each generation contain potentially useful information that is not currently utilized. In particular, we look at the Sufficiency measure of a rule as an information theoretic indicator. We propose the modification of the XCS algorithm using this in the early formative stages of each run in view of these additional indicators of usefulness. The probability of selection during that period would be based on sufficiency.