Enhancing Multi-layer Perceptron Training with Prairie Dog Optimizer: Advancing Classification Accuracy in Medical Diagnostics

Multi-layer Perceptron (MLP) networks are widely used for classification and regression due to their flexibility and capacity. However, their performance depends heavily on the optimization algorithm used during training, with traditional backpropagation often failing to achieve globally optimal sol...

Full description

Saved in:
Bibliographic Details
Published in2024 IEEE Silchar Subsection Conference (SILCON 2024) pp. 1 - 6
Main Authors Debbarma, Tanzil, Shaikh, Azharuddin, Tarafdar, Anirban, Baidya, Abhijit, Majumder, Pinki, Bera, Uttam Kumar
Format Conference Proceeding
LanguageEnglish
Published IEEE 15.11.2024
Subjects
Accuracy
Backpropagation
Breast cancer
Diabetes
Dogs
Grasslands
Iris
multi-layer perceptron
Optimization
Prairie dog optimization
Search problems
Training
Online AccessGet full text

Cover

More Information
Summary:Multi-layer Perceptron (MLP) networks are widely used for classification and regression due to their flexibility and capacity. However, their performance depends heavily on the optimization algorithm used during training, with traditional backpropagation often failing to achieve globally optimal solutions. This paper presents a Prairie Dog Optimization (PDO) based backpropagation technique for MLP training, aimed at enhancing accuracy, especially in classification tasks. PDO improves MLP optimization by using a dynamic, population-based search strategy to navigate complex loss landscapes effectively, leading to better convergence to global optima. Experiments on breast cancer diagnosis and diabetes classification demonstrate that PDO-MLP significantly outperforms other optimization-based backpropagation methods, including flower pollination algorithm (FPA), ant colony optimization (ACO), standard gradient descent backpropagation (BP), and particle swarm optimization (PSO). Notably, PDO-MLP achieved 97.30% accuracy on the Iris dataset, 82.6% on diabetes classification, and 96.41% on breast cancer diagnosis, proving its effectiveness for various real-world applications.
DOI:10.1109/SILCON63976.2024.10910597