XCSF under limited supervision
Markus Görlich-Bucher, Jörg Hähner · Proceedings of the Genetic and Evolutionary Computation Conference Companion · 2022
XCSF is a well-researched function approximator used for supervised, online regression learning tasks. Hereby, XCSF relies on receiving a ground truth after each prediction step in order to update its internal rulebase. As there exist various conceivable scenarios where the amount of external supervision might be limited, we investigate how XCSF can be adapted in order to be applicable in such environments. We change the update mechanism such that only inexperienced classifiers are updated, and introduce an external covering procedure as well as a cache for saving external supervisions. Our experiments show that, despite performing significantly worse than an non-limited XCSF, our adaptations still provide acceptable results and limit the necessity for external supervision noticeably.