Performance evaluation of Hopfield neural networks for overlapped English characters by using genetic algorithms

Abstract

In this paper we are analyzing the performance of the Hopfield neural network as an associative memory feature for pattern storage and recalling purposes. A genetic algorithm is employed for recalling of the stored patterns corresponding to the presented input overlapped patterns. The training pattern set considered is the English characters and the Hebbian learning rule is used for encoding the pattern information in the Hopfield network. The recalling of patterns is accomplished with a genetic algorithm by settling the network in appropriate stable states for the presented overlapped input pattern. This is achieved by one set of weights used for recalling of stored patterns. If an overlapped pattern is presented, the network goes into a stable state which represents one of the characters for the presented pattern, and then this recalled character is eliminated from the overlapped pattern and a new input pattern is formed. The network will then be assigned this new input pattern as an initial state and the network goes into a state representing the other character of the pattern. The simulated results demonstrate the performance of the network for the pattern recalling corresponding to the presented overlapped input patterns.