Compaction for Code Fragment Based Learning Classifier Systems

摘要

Learning Classifier Systems (LCSs) originated from artificial cognitive systems research, but migrated such that LCS became powerful classification techniques in single domains. Modern LCSs can extract building blocks of knowledge utilizing Code Fragments in order to scale to more difficult problems in the same or a related domain. Code Fragments (CF) are GP-like sub-trees where past learning can be reused in future CF sub-trees. However, the rich alphabet produced by the code fragments requires additional computational resources as the knowledge and functional rulesets grow. Eventually this leads to impractically long chains of CFs. The novel work here introduces methods to produce Distilled Rules to remedy this problem by compacting learned functions. The system has been tested on Boolean problems, up to the 70 bit multiplexer and 3×11 bit hidden multiplexer, which are known to be difficult problems for conventional algorithms to solve due to large and complex search spaces. The new methods have been shown to create a new layer of rules that reduce the tree length, making it easier for the system to scale to more difficult problems in the same or a related domain.