Comparing different metaheuristics for model selection in a supervised learning classifier system
Jonathan Wurth, Michael Heider, Helena Stegherr, Roman Sraj, Jörg Hähner · Proceedings of the Genetic and Evolutionary Computation Conference Companion · 2022
In the context of tasks that highly involve human interaction and expert knowledge, i.e., operator guidance in manufacturing, the possibility of decision verifications by the user is a key requirement for inspiring confidence in a system and its predictions. Rule-based Machine Learning offers one way to create such systems, with Learning Classifier Systems being a family of algorithms whose models are by design human-interpretable. Obtaining a rule base as compact and accurate as possible is a mandatory prerequisite to increasing comprehensibility, and metaheuristics such as Genetic Algorithms or Particle Swarm Optimization are powerful approaches to reducing the size of large rule bases. In particular, this paper will analyze five population-based metaheuristics and their ability to compose solutions (rule subsets) as part of a newly developed rule-based learning system, the Supervised Rule-based Learning System (SupRB). The experiments suggest that all metaheuristics can significantly reduce the complexity and filter out obstructive rules, increasing the prediction quality in the process.