Three-cornered coevolution learning classifier systems for classification tasks

Syahaneim Marzukhi, Will N. Browne, Mengjie Zhang

Abstract: The Three-Cornered Coevolution concept describes a framework where artificial problems may be generated in concert with classification agents in order to provide insight into their relationships. This is unlike standard studies where humans set a problem's difficulty, which may have bias or lack understanding of the multiple interactions of a problem's characteristics, such as noise in conjunction with class imbalance. Previous studies have shown that it is feasible to generate problems with one agent in relation to a single classification agent's performance, but when to adjust the problem difficulty was manually set. This paper introduces a second classification agent to trigger the coevolutionary process within the system, where its functionality and effect on the system requires investigation. The classification agents, in this case Learning Classifier Systems, use different styles of learning techniques (e.g. supervised or reinforcement learning techniques) to learn the problems. Experiments show that the realized system is capable of autonomously generating various problems, triggering learning and providing insight into each learning system's ability by determining the problem domains where they perform relatively well - this is in contrast to humans having to determine the problem domains.