Catálogo de publicaciones - libros

The XCS classifier system is only one among many evolutionary rule-based learning systems. With the successful facetwise analysis approach carried through in XCS, it needs to be considered how the theory may carry over (1) to analyze other LCSs and related systems in the same way and (2) to apply the knowledge to create new LCS systems, targeted to a specific problem at hand.

This chapter outlines how the derived Facetwise LCS approach can carry over to the design of cognitive learning systems. We propose the integration of LCS-like search mechanisms into cognitive structures for structural growth and distributed, modular relevancy identification. The mechanisms may be integrated into multi-layered, hierarchical learning structures and may influence solution growth interdependently using only RL mechanisms and evolutionary problem solution structuring.

This book has investigated rule-based evolutionary online learning systems, often referred to as learning classifier systems (LCSs). The proposed facetwise problem analysis approach showed that the XCS classifier system is an effective online learning and generalizing predictive reinforcement-based system that evolves a complete, maximally accurate and maximally general problem solution quickly and reliably. The scalability analysis showed that XCS can PAC-learn restricted k-DNF functions. However, XCS is not restricted to the world of k-DNF problems but can also learn in real-valued problem domains, regardless if the task is to classify data or to approximate functions. Moreover, XCS can be applied in multistep reinforcement learning (RL) problems, in which the system was shown to learn optimal behavioral policies effectively filtering noise and ignoring irrelevant attributes. The extension of the proposed facetwise analysis approach for LCS-based system design suggested strong potentials for integrating LCS systems and LCS-based learning components into modular, hierarchical, predictive learning structures.