Training neural networks using Multiobjective Particle Swarm Optimization

This paper suggests an approach to neural network training through the simultaneous optimization of architectures and weights with a Particle Swarm Optimization (PSO)-based multiobjective algorithm. Most evolutionary computation-based training methods formulate the problem in a single objective manner by taking a weighted sum of the objectives from which a single neural network model is generated. Our goal is to determine whether Multiobjective Particle Swarm Optimization can train neural networks involving two objectives: accuracy and complexity. We propose rules for automatic deletion of unnecessary nodes from the network based on the following idea: a connection is pruned if its weight is less than the value of the smallest bias of the entire network. Experiments performed on benchmark datasets obtained from the UCI machine learning repository show that this approach provides an effective means for training neural networks that is competitive with other evolutionary computation-based methods.