Investigating the Adaptation and Forgetting in Fuzzy Neural Networks Through a Method of Training and Zeroing

Rules extraction, rules insertion and a method of alternative training and zeroing, based on zeroing of small connections in a fuzzy neural network, have been investigated in respect of adaptation and forgetting phenomena. The experiments show that updated (zeroed) fuzzy neural networks are more robust to forgetting, faster and better at adaptation and provide a good generalisation. The method may be used for building adaptive neuro-fuzzy systems. It may now be possible to control the level of adaptation and forgetting in practical systems for speech recognition, process control, decision making, time-series forecasting, etc. A specialised engineering environment which facilitates such experiments has been briefly described.