Developing locally linear models with XCSF: Scalability, Robustness, and Applicability

develops piece-wise overlapping, local linear models in function
approximation problems. We show that XCSF scales optimally in problems in
which a sufficient fitness signal is available. The evolutionary learning
approach ensures the high learning robustness and versatile applicability of
the system. Essentially, we show that XCSF's learning capabilities are even
partially superior to LWPR (locally weighted projection regression) while
being applicable to a wider range of problems. Moreover, we show that
further versatility in the kernel structures, which define the evolving
local models, broadens the search space but enables to search for
potentially highly useful problem regularities - thus increasing the
applicability of the system even further. We finish with a few exemplary
evaluations and applications of the system.