Absumption and Subsumption based Learning Classifier System for Real-World Continuous-based Problems

Abstract

Learning Classifier Systems (LCSs), a series of rules-based evolutionary computation techniques, which have solved a wide range of discrete-feature-based applications over their 40 years of history. Yet, adapting LCSs to complicated continuous-feature-based domains is still an unsolved challenge. This paper proposes new LCS methods specialized for continuous problems. Concretely, phenotype-orientated Absumption, Subsumption, and Mutation are proposed and employed to form and revise rules directly in a single iteration according to the target problems' inherent data distribution, allowing rules to be released from the burden of directly carrying the information of previous instances. Furthermore, a novel representation format supporting fine-grained generalization degree modification is also proposed. Experiments demonstrate for the first time that LCSs are promising techniques in efficiently producing models with satisfactory prediction performance for complicated continuous problems.