Optimized Machine Learning Model with Modified Particle Swarm Optimization for Data Classification

Metaheuristic search algorithms (MSAs) receive increasing popularity in recent year due to its excellent capability of solving complex real-world optimization problems without depending on gradient information. Particle swarm optimization (PSO), as one of MSAs, is widely used in optimization task due to its simple framework and quick convergence speed toward global optimum. However, conventional PSO suffers from premature convergence and quick diversity loss of population when the population is poorly initialized due to its random characteristics. In this paper, a new variant of PSO namely PSO with multi-chaotic scheme (PSOMCS) is introduced to train artificial neural network (ANN) by optimizing its neuron weights, biases and selection of suitable activation function based on the datasets obtained from UCI machine learning repository. Initial population generated using multi-chaotic system and oppositional-based learning ensure broader search space coverage, enabling PSOMCS to solve complex optimization problems effectively. Classification performances of ANN trained with PSOMCS are compared with other existing PSO variants. Based on simulation results, ANN optimized by PSOMCS outperformed its competitors in terms of classification performance for both training and testing datasets.