A Proposal for a Leaner Narrative of Learning Classifier Systems

Learning classifier systems (LCS) are often regarded as complex and challenging to master despite the community's ongoing efforts to provide simplified educational models and detailed algorithmic descriptions. In this position paper, we argue that such perceived complexity is due to how LCSs are explained, which is still based on the narrative used to present the early models almost 50 years ago. Such a narrative centers around the system's interaction with the environment and how the information streams from detectors to actuators, creating increasingly focused classifier sets. Accordingly, it blends core LCS concepts with elements universal to all value-based reinforcement learning algorithms that are never included in the descriptions of competing methods. We suggest another, possibly leaner, narrative based on the view of XCS as an approximator of state-action value functions to solve reinforcement learning tasks. We show how abandoning the traditional narrative may result in a simpler description of XCS that can be easily extended to integrate known reinforcement learning extensions and tackle classification and regression problems. Our approach can be seamlessly applied to one-step and multi-step scenarios without modifications. It also provides guidelines for developing LCS implementations that might be more accessible to people approaching LCS for the first time.