Combining Symbolic and Numeric Methods for Learning to Predict Temporal Series

Abstract

Radial Basis Function Networks axe universal function approximators which can be easely constructed from rule sets learned by a symbolic learner. This paper proposes the use of a more expressive concept description language, based on first order Ingics, and of a learning system (FLASH) working in such an environment, in order to incorporate the feature selection phase into the learning process. The main advantage of the first order description language used by FLASH, is that it allows to define and manipulate a large number of feztures while keeping simple the description of instances. The method is tested on two non trivial problems of functional regression represented by the Mackey-Glass temporal series and by the control function for a manipulation robot. The obtained results demonstrate the high accuracy of the proposed method.