Genetic Algorithms and Neural Networks

Abstract

The basic genetic algorithm is introduced including the representation of individuals in populations, data structures for the representation of variables, binary strings, assessment of individual fitness, selection for recombination, crossover and mutation operators. The penalty function method of handling design constraints is introduced. The basic GA is illustrated by optimizing a simple structural design. We consider how we might improve the GA by on-line adaptation of the main controls. We then review string coding, the schema theorem and the formation of building blocks in the strings. We consider the coding of continuous-valued variables and bit array representations, elitism, methods of maintaining diversity in the population and introduce a further illustration in structural optimization. The application of the genetic algorithm is then extended into large scale-situations, particularly design situations involving a large number of variables. A combinatorial space reduction heuristic based on a record of parameter selection intensities is described. The allocation of fitness to partial strings is reviewed. Consideration is given to the multi-objective GA and pareto optimality. There follows a brief introduction to mathematical models of the GA. The GA is used to train a neural network as an alternative to back-propagation. We consider the ‘permutations’ problem and introduce the concept of ‘shift’. The method is illustrated by training a neural network for structural analysis. The chapter concludes with a brief review of the implicit parallelism of the GA and suggestions as to how the algorithm might be improved with parallel hardware and a further example application.