A novel online AARS-based Pseudo Outer-Product Fuzzy Neural Network (online POPFNN-AARS)

Abstract

This report introduces an incremental learning algorithm for the Pseudo Outer-Product Fuzzy Neural Network (POPFNN) together based on the Approximate Analogical Reasoning Schema (AARS). Fuzzy neural networks have been applied to many domains due to their strong reasoning and optimization capabilities. But as training data changes, there is a need for most existing networks to retain previously evaluated data and retrain from scratch. The proposed online POPFNN-AARS avoids this problem by allowing learning to be performed incrementally. This is achieved by adapting the learning concept from the Incremental Backpropagation Learning Network, the structure and one-pass learning algorithm of the proposed online POPFNN-AARS system are presented in this dissertation. A suite of experiments from classification problems to nonlinear regression tasks are subsequently performed to evaluate the performance of the proposed online POPFNN-AARS system. The empirical results are encouraging and significantly demonstrate the benefits of incremental learning to solve complex problems. The novel online POPFNN-AARS system is also applied as a decision support system for ovarian cancer diagnosis, where incremental learning is observed to play a significant role in acquiring new domain knowledge.