Reflective Learning Classifier Systems for Self-Adaptive and Self-Organising Agents

Learning is a key capability in self-adaptive and self-organising systems to deal with dynamically changing conditions, unanticipated operational incidences, and open system constellations. Learning Classifier Systems, especially variants of the X CS Classifier System, have been demonstrated to be successfully applicable to the self-adaptation task concerning condition-aware re-configuration of parameters of the productive system. Compared to recent alternatives such as control theoretic approaches or (deep) reinforcement learning, XCS has advantages in the interpretability and continual evolution of knowledge, rendering it particularly applicable to real-world control problems. In this paper, we argue that the algorithmic concept of X CS and its integration in cooperative system constellations needs to be augmented with concepts for self-reflection, flexibility and transferability of knowledge to cover still unsolved challenges of real-world control problems. We present a brief introduction to the field and derive core challenges leading to a research agenda to achieve extended, reflective XCS-based self-adaptive and self-organising agents.