Applying crossover operators to automatic neural network construction

The ability to automatically construct neural networks is of importance, since it supports reduction in development time and can lead to simpler designs than traditionally handcrafted networks. Automation is further required to take the step towards a more autonomous learning system. In this paper, we report further results involving the automatic network construction algorithm EGP (Evolutionary Growth Perceptron), which utilizes simple evolutionary processes to locally train network features using the perceptron rule. Emphasis is placed on determining the effectiveness of several types of crossover operators in conjunction with varying the population size and the number of epochs during which individual perceptrons are trained. The crossover operators considered and introduced are: simple random, weighted and blocked