Comparison between different methods for developing neural network topology applied to a complex polymerization process

In this paper, three methods for developing the optimal topology of feed-forward artificial neural networks are described and applied for modeling a complex polymerization process. In the free radical polymerization of styrene, accompanied by gel and glass effects, the monomer conversion and molecular masses are modeled depending on reaction conditions. The first proposed methodology is an algorithm which systemizes a series of criteria and heuristics on neural networks modeling. The method is laborious, but the practical considerations, structured in a 6-steps algorithm, and the criterion and formula used for calculating the performance indices give the method a real chance of obtaining a neural network of minimum size and maximum performance. The next two methods belong to evolutionary techniques and they are based on a classical genetic algorithm and differential evolution algorithm. They automatically develop the neural network topology by determining the optimal values for the number of hidden layers, the number of neurons in these layers, the weights between layers, the biases of the neurons and the activation functions. For all the three methods, a combination between training and testing errors was considered in order to evaluate the performance of the developed neural networks and to choose the best one among them. The relative percentage errors calculated in the validation phase registered good values, under 7%. A comparison between these methods pointed out both advantages and disadvantages, but, even if they lead to different network architectures, accurate results were obtained and, consequently, near optimal neural network topologies were developed.