Applying genetic algorithms to neural network problems

Genetic algorithms perform nonlinear combinatorial optimization using 'intrinsic parallelism' to sample hyperplanes in an n-dimensional hypercube that corresponds to a binary representation of a problem solution space. 'Intrinsic parallelism' refers to the fact each binary 'genotype' of length L corresponds to a corner in the hypercube and is a sampling of 2L hyperplanes which intersect at that point in the search space. Genetic selection guides the search toward high performance hyperplanes, while genetic recombination creates new genotypes by reconfiguring genotype fragments which represent high performance hyperplanes. The ideas outlined guided a neural net design currently being implemented. The ideas address the kinds of net learning and optimization problems which can be attacked while using a simple problem definition that can be represented in binary form. The successful application of such a system would lay the groundwork for exploring the potential uses of genetic algorithms for solving more complex types of problems.