A combined network architecture using ART2 and back propagation for adaptive estimation of dynamical processes

A neural network architecture called ART2/BP is proposed. The goal has been to construct an artificial neural network that learns incrementally an unknown mapping, and is motivated by the instability found in backpropagation (BP) networks: after first learning pattern A and then pattern B, a BP network often has completely `forgotten´ pattern A. A network using both supervised and unsupervised training is proposed, consisting of a combination of ART2 and BP. ART2 is used to build and focus a supervised backpropagation network consisting of many small subnetworks each specialized on a particular domain of the input space. The ABT2/BP network has the advantage of being able to dynamically expand itself in response to input patterns containing new information. Simulation results show that the ART2/BP network outperforms a classical maximum likelihood method for the estimation of a discrete dynamic and nonlinear transfer function