Improved Cuckoo Search Based Neural Network Learning Algorithms for Data Classification Abdullah

Artificial Neural Networks (ANN) techniques, mostly Back-Propagation Neural Network (BPNN) algorithm has been used as a tool for recognizing a mapping function among a known set of input and output examples. These networks can be trained with gradient descent back propagation. The algorithm is not definite in finding the global minimum of the error function since gradient descent may get stuck in local minima, where it may stay indefinitely. Among the conventional methods, some researchers prefer Levenberg-Marquardt (LM) because of its convergence speed and performance. On the other hand, LM algorithms which are derivative based algorithms still face a risk of getting stuck in local minima. Recently, a novel meta-heuristic search technique called cuckoo search (CS) has gained a great deal of attention from researchers due to its efficient convergence towards optimal solution. But Cuckoo search is prone to less optimal solution during exploration and exploitation process due to large step lengths taken by CS due to Levy flight. It can also be used to improve the balance between exploration and exploitation of CS algorithm, and to increase the chances of the egg’s survival. This research proposed an improved CS called hybrid Accelerated Cuckoo Particle Swarm Optimization algorithm (HACPSO) with Accelerated particle Swarm Optimization (APSO) algorithm. In the proposed HACPSO algorithm, initially accelerated particle swarm optimization (APSO) algorithm searches within the search space and finds the best sub-search space, and then the CS selects the best nest by traversing the sub-search space. This exploration and exploitation method followed in the proposed HACPSO algorithm makes it to converge to global optima with more efficiency than the original Cuckoo Search (CS) algorithm.