James Worth Bagley College of Engineering
Department of Computer Science and Engineering
Bachelor of Science (B.S.)
Artificial neural networks, which mimic the human brain's ability to learn from experiences, are increasingly being used to analyze complex datasets. However, proper structural configuration requires intuition, trial-and-error, and frequent human attention. This study investigates an automated alternative that uses a Darwinian evolutionary strategy to optimize the structure of a small-scale Modified National Institute of Standards and Technology (MNIST) image classification network. Using accuracy as a measure of fitness, it examines the effect of varying the amount each network learns prior to differential reproduction. The accuracy of optimized networks was significantly higher than random initial networks, being increased by up to 0.133% (1.435 standard deviations above initial mean accuracy). The use of evolutionary strategies in design holds promise for producing networks that are appreciably more accurate than randomly-generated networks without a large ongoing input of human attention.
Bethel, Cindy L.
Dinep-Schneider, Nicholas, "On the feasibility of using genetic algorithms to optimize the structure of small multilayer perceptrons" (2018). Honors Theses. 31.