Massively parallel non-linear system identification techniques: A committee architecture

Antonio Carlos Gay Thome, Purdue University

Abstract

In this thesis homogeneous committee architectures, where each agent consists of an independent neural network structure, are exploited for time series prediction. The motivation for this work is based on the fact that many real prediction problems are too complex for a global approximation, but can be better handled through a decompose-solve-recombine approach. Combination of multiple predictors is therefore, the major subject of this thesis, which develops a complete combining methodology covering the following steps: (1) systematic decomposition of the original problem based on a criterion of spatial similarity; (2) a parallel learning of the individual subproblems; and (3) a combination strategy of the multiple predictions. To support both, the systematic decomposition of the input space and the learning effort demanded for the parallel training of many networks on large data sets, a novel clustering algorithm is developed and innovative learning accelerating techniques are proposed. Empirical results are reported and discussed.

Degree

Ph.D.

Advisors

Tenorio, Purdue University.

Subject Area

Electrical engineering

COinS