Rango: An Intuitive Rule Language for Learning Classifier Systems in Cyber-Physical Systems
Melanie Feist, Martin Breitbach, Heiko Trötsch, Christian Becker, Christian Krupitzer · 2022
Self-adaptation is crucial for cyber-physical systems (CPS) to meet their requirements in environments characterized by complexity and uncertainty. As many situations that CPS encounter at runtime are not foreseeable at design time, (online) learning approaches are attractive for such systems. Learning classifier systems (LCS) are a promising learning approach for CPS thanks to their rather low computational complexity. They operate on a set of rules that describe potential adaptation behavior. So far, specifying rules for a learning classifier system is a tedious task that requires expert knowledge. In this paper, we present Rango — an intuitive rule language for learning classifier systems — to overcome this challenge. Compared to existing approaches, Rango has a strong focus on CPS and provides a large variety of corresponding keywords. In addition, Rango rules are automatically transferred into a representation that is usable in a learning classifier system without any modifications. Rango therefore empowers system administrators to formulate rules and, hence, leverage an online learning approach for their use case without having prior experience with learning classifier systems. We evaluate Rango extensively with (i) a complexity analysis of parsing and rule evaluation, (ii) a usefulness study which shows that Rango facilitates both the writing of rules and the understanding of LCS output and (iii) a usability study, which proves that basic programming knowledge is sufficient to understand and formulate Rango rules.