A supervised neural network for dynamic systems identification

This paper describes a neural network for identifying discrete dynamic systems using only input-output relationships. The network is based on the ART2 network. An improved adaptive resonance topology has been developed which achieves a robust structure for dynamic systems identification. A mapping field has been implemented for the system to be modelled. A new output short term memory (STM) has been added to the neural network model and the connection between the new field and the category field has been made by long term memory (LTM) adaptive filters. Top-down adaptive filters in the new field assume full responsibility for coding the output expectation. New feedback connections have been added to provide recurrent properties. The modified network has been used successfully to model dynamic systems. Results are presented to demonstrate the effectiveness of the network