A RECURRENT NETWORK IMPLEMENTATION OF TIME-SERIES CLASSIFICATION

An incremental credit assignment (ICRA) scheme is introduced and appli ed to time series classification. It has been inspired from Bayes' rul e, but the Bayesian connection is not necessary either for its develop ment or proof of its convergence properties. The ICRA scheme is implem ented by a recurrent, hierarchical, modular neural network, which cons ists of a bank of predictive modules at the lower level, and a decisio n module at the higher level. For each predictive module, a credit fun ction is computed; the module that best predicts the observed time ser ies behavior receives highest credit. We prove that the credit functio ns converge (with probability one) to correct values. Simulation resul ts are also presented.