Improving Extreme Learning Machine Algorithm Through Optimization Technique

Abstract

Classification and function approximation tasks are used widely in artificial neural network (ANN), and there are a number of algorithms by which a neural network can be trained. ELM is known as extreme learning machine; a novel algorithm has come under the spotlight in the last decade, and the reason for its popularity is a rapid learning rate and a more efficient generalization capability, in comparison with orthodox gradient descent learning algorithms for (SFLN) single-hidden-layer feed-forward neural network. Even though it has many advantages, the single-hidden-layer feed-forward neural network trained using ELM requires a number of hidden neurons. Due to this, it faces poor condition problem because of the randomly created hidden biases and input weights. The solution to this problem is to use an optimization technique, using it to find a set of inputs that are optimal. We have proposed a hybrid learning scheme that uses one such optimization technique which will be used to figure out an optimal set of input-hidden node weights, and then a set of optimal input weights are used to determine the output weights analytically. Few of the techniques that are used for optimization like competitive swarm optimization (CSO) and particle swarm optimization (PSO) have been proposed in the past. The hybrid learning scheme provides (i) help in enhancing the search variety, which encourages us to explore the solution search space better and (ii) improve the structure of ELM to solve the problem of random input. The proposed scheme in addition to other multi-objective optimal algorithms requires smaller amounts of hidden neurons compared to the conventional ELM for achieving higher classification accuracy. The proposed hybrid methodology has the ability to produce good condition SLFNs with an improved accuracy with better generalization.