Optimizing Neural Network Weights and Biases Using Particle Swarm Optimization for Classification Task

The current digital era greatly demands reliable automatic classification systems, especially to handle large and increasingly complex data volumes. One attractive alternative is the Particle Swarm Optimization (PSO) algorithm, which is recognized for its effective global search. Nevertheless, the performance of PSO for training artificial neural networks with complex or large-scale data is still uncertain. The primary purpose of this research is to create and assess a classification engine based on MLP, which uses a PSO algorithm to generate weights and biases. The assessment was made on three different types of data - dummy data, Iris data, and Sasak script images. For the dummy and Iris datasets, the model successfully achieved 100% accuracy, demonstrating the effectiveness of the PSO-MLP approach on simpler data. However, the results differed significantly for the more extensive and complex image dataset, where the model experienced a drastic decline in performance. In the image classification test with 6 classes, the model with one hidden layer achieved 71% accuracy, while the model with two hidden layers only reached 56%. For 12-class classification, accuracy dropped to 35% and 25%, respectively, and for 18 classes, the model achieved only 27% and 7%. These results indicate that while PSO is effective in optimizing perceptron weights and biases for smaller and simpler datasets, its ability to handle large-scale image classification with increasing complexity remains limited. Therefore, there is a need for optimization strategies to enhance the accuracy of optimization for more complex data.