A multilayered neural network composed of backpropagation and Hopfield nets and internal space representation

A multilayered neural network composed of backpropagation and Hopfield nets is proposed. In the network, the Hopfield net provides energy minimization, and two backpropagation nets put in front and in the rear of the Hopfield net function as mapping networks with learning capability. The input signals from the environment are mapped via one backpropagation net (BPN) into the internal space, where the Hopfield net minimizes the total energy according to the internal space representation. The output signals of the Hopfield net are mapped again to the environment via the other backpropagation net (BPN/sup -1/) with an inverse mapping. The indirect feedback loop via BPN/sup -1/ and BPN can be used if the constraints must be satisfied according to the distance measure in the environment. Simulation studies for manipulator configuration control show that the proposed network helps the end of the manipulator to reach the target point through the shortest pass in the internal space. Then parameters such as the constraints for the link lengths and the distances between the joints and the obstacles are conserved according to the distance measure not in the internal space but in the workspace.<>