Reusable knowledge by linkage-classifier in Accuracy-based Learning Classifier System
Kotaro Usui, Masaya Nakata, Keiki Takadama · 2014
Accuracy-based Learning Classifier System (XCS) can learn correct classifiers in a given environment, but they may not be reusable even in small environmental changes. To tackle this problem, this paper propose a new XCS, XCS with the linkage-classifier (XCSL), which can create reusable knowledge as a linkage of useful classifiers for a changed environment. The linkage-classifier represents the executed order of the classifiers, and is a set of classifiers which each must be reused as a sequence of actions to reach a goal of task. The intensive experiments on a benchmark sequential decision task have revealed that, XCSL performs as well as the conventional LCSs (Learning Classifier Systems) in the environments without any changes, while XCSL performs with fewer iterations than the conventional ones in the environments with some change.