Training product unit neural networks with genetic algorithms

Product unit neural networks are useful because they can handle higher order combinations of inputs. When trained using traditional backpropagation, however, they are often susceptible to local minima. The use of genetic algorithm exploratory procedures that can often locate near-optimal solutions to complex problems to overcome this, is discussed. The genetic algorithm maintains a set of trial solutions and forces them to evolve toward an acceptable solution. A representation for possible solutions must first be developed. Then, with an initial random population, the algorithm uses survival of the fittest techniques as well as old knowledge in the gene pool to improve each generation´s ability to solve the problem. This improvement is achieved through a four-step process of evaluation, reproduction, breeding, and mutation. An example application is described.<>