Publication Date

8-2023

Advisor(s) - Committee Chair

Uta Ziegler, Mustafa Atici, Michael Galloway

Degree Program

Department of Computer Science

Degree Type

Master of Science

Abstract

This thesis examines the feasibility of implementing two simple optimization methods, namely the Weights Power method (Hagiwara, 1994) and the Tabu Search method (Gupta & Raza, 2020), within an existing framework. The study centers around the generation of artificial neural networks using these methods, assessing their performance in terms of both accuracy and the capacity to reduce components within the Artificial Neural Network’s (ANN) topology.

The evaluation is conducted on three classification datasets: Air Quality (Shahane, 2021), Diabetes (Soni, 2021), and MNIST (Deng, 2012). The main performance metric used is accuracy, which measures the network's predictive capability for the classification datasets. The evaluation also considers the reduction of network components achieved by the methods as an indicator of topology optimization.

Python, along with the Scikit-learn framework, is employed to implement the two methods, while the evaluation is conducted in the cloud-based environment of Kaggle Notebooks. The evaluation results are collected and analyzed using the Pandas data analysis framework, with Microsoft Excel used for further analysis and data inspection.

The Weights Power method demonstrates superior performance on the Air Quality and MNIST datasets, whereas the Tabu Search method performs better on the Diabetes dataset. However, the Weights Power method encounters issues with local minima, leading to one of its stop conditions being triggered. On the other hand, the Tabu Search method faces challenges with the MNIST dataset due to its predetermined limits and restricted scope of changes it can apply to the neural network.

The Weights Power method seems to have reached its optimal performance level within the current implementation and evaluation criteria, implying limited potential for future research avenues. In contrast, to enhance the dynamic nature of the Tabu Search method, further investigation is recommended. This could entail modifying the method's capability to adapt its stop conditions during runtime and incorporating a mechanism to scale the magnitude of changes made during the optimization process. By enabling the method to prioritize larger changes earlier in the process and gradually introducing smaller changes towards the conclusion, its effectiveness could be enhanced.

Disciplines

Artificial Intelligence and Robotics | Computer Sciences | Physical Sciences and Mathematics

Download

Share

COinS