Training of perceptron networks with sigmoidal activation functions is executed by a new evolutionary algorithm, Differential Evolution (DE). On this evolutionary optimization method, we apply a search space reduction algorithm and compare this approach to training by standard DE and training by the conjugate gradient method. The testbed is Austrian interregional telecommunication traffic data. In our experiments, DE with only mutation was able to achieve better suboptimal solutions than DE together with crossover and search space reduction. DE with mutation only outperforms the conjugate gradient method in terms of in-sample and out-of-sample performance. The experimental results support the hypothesis that mutation-based evolutionary algorithms (like DE) tend to be more suitable methods for training perceptron networks than crossover-based evolutionary algorithms and gradient based methods. An evolutionary mutation-based algorithm for weight training in neural networks for telecommunication flow modelling.

Top