STRUCTURE OPTIMIZATION OF ARTIFICIAL NEURAL NETWORKS USING SWARM INTELLIGENT

Abstract

The structure is a very important aspect in neural network design, it is not only impossible to determine an optimal structure for a given problem, it is even impossible to prove that a given structure is optimal. In this paper, PSO (Particle Swarm Optimization) are used to construct best ANN (Artificial Neural Network) architectures, and find an optimal pattern of connections and weights to reduce structure complexity by minimizing the number of connection weights in a Feed Forward Artificial Neural Network (FFANN). They are called Particle Swarm Optimization-Neural Network systems (PSONN). PSONN systems are examined through theoretical analysis and computer simulation using MATLAB package. They are tested by several different examples, where the tests show that PSO a more efficient and automated search method can be used to find an optimal topology of ANN. The best and trained network with few numbers of iteration is provided using PSONN for finding an optimal structure. Finally, a simpler network, faster training with higher accuracy than full connected network is obtained by using PSONN for finding optimal connections and weights.