Pattern Search Using Genetic Algorithms and a Neural Network Model
Shigetoshi Nara
Department of Electrical and Electronic Engineering,
Faculty of Engineering, Okayama University,
Tsuchimanaka 3-1-1, Okayama 700, Japan
Wolfgang Banzhaf
Current address: Department of Computer Science, Dortmund University and Informatics Center Dortmund, Joseph-von-Fraunhofer-Strasse 20, 44227 Dortmund, Germany
Central Research Laboratory, Mitsubishi Electric Corporation,
Tsukaguchi Honmachi 8-1-1, Amagasaki,
Hyogo 661, Japan
Abstract
An information processing task that generates combinatorial explosion and program complexity when treated by a serial algorithm is investigated using both genetic algorithms and a neural network model. The task in question is to find a target memory from a set of stored entries in the form of "attractors'' in a high-dimensional state space. The representation of entries in the memory is distributed ("an auto associative neural network'' in this paper) and the problem is to find an attractor under a given access information where the uniqueness or even existence of a solution is not always guaranteed (an ill-posed problem). The genetic algorithm is used for generating a search orbit to search effectively for a state that satisfies the access condition and belongs to the target attractor basin in the state space. The neural network is used to retrieve the corresponding entry from the network. The results of our computer simulations indicate that the present method is superior to a search method that uses a Markovian random walk in state space. Our techniques may prove useful in the realization of flexible and adaptive information processing, since pattern search in a high-dimensional state space is common in various kinds of parallel information processing.
