Some Experiments on the Use of Genetic Algorithms in a

While it appears to be a good idea to use Genetic Algorithm(GA) to train a Neural network, past results do not c o n f i i such optimism. The main problems encountered are the speed of convergence, convergence to the wrong answer, and failure to converge. In this paper we combine GA and Simulated annealing to form a Genetic Boltzmann Machme(GBM) and attempt to understand the propertks of such an architecture by experiments. We introduce the concept of weight reordering and demonstrate that it overcomes most of the convergence problems. Results of other experiments are also shown relating to the selection of parameters for the G A the effects of population, different crossover point operators, and hidden units are illustrated. We conclude that with careful design a GBM can perform nearly as well as a Boltzmann Machine in a scalar computer. However, GBM is easily amenable to parallel computation by process farming.