Using genetic recombination to optimize neural networks

Summary form only given, as follows. The authors have successfully optimized neural nets using a genetic algorithm that differs in fundamental ways from standard genetic algorithms. This algorithm uses one-at-a-time reproduction and allocates reproductive opportunities according to rank to achieve the desired selective pressure. The authors refer to this as the GENITOR algorithm. A key advantage of genetic search is its global optimization capabilities; they show that the genetic algorithm easily optimizes a problem which backpropagation fails to solve due to local minima. The theoretical foundations of genetic search are presented as well as a thorough discussion of numerous experiments with four different neural network optimization problems. Other potential applications of genetic algorithms to problems of neural nets are also discussed.<>