Comparison of two learning networks for time series prediction

Abstract

Hierarchical mixtures of experts (HME) [JJ94] and radial basis function (RBF) networks [PG89] are two architectures that learn much faster than multilayer perceptrons. Their faster learning is due not to higher order search mechanisms, but to restricting the hypothesis space of the learner by constraining some of the layers of the network to use linear processing units. It can be conjectured that since their hypothesis space is restricted in the same manner, the approximation abilities of the two networks should be similar, even though their computational mechanisms are different. An empirical verification of this conjecture is presented, based on the task of predicting a nonlinear chaotic time series generated by an infrared laser.