An Empirical Study on the Efficacy of Evolutionary Algorithms for Automated Neural Architecture Search
M.S. Computer Science
College of Science & Mathematics
Shen-Shyang Ho, Ph.D.
Committee Member 1
Bo Sun, Ph.D.
Committee Member 2
Ning Wang, Ph.D.
evolutionary algorithm, genetic algorithm, machine learning, neural architecture engineering, neural network
The configuration and architecture design of neural networks is a time consuming process that has been shown to provide significant training speed and prediction improvements. Traditionally, this process is done manually, but this requires a large amount of expert knowledge and significant investment of labor. As a result it is beneficial to have automated ways to optimize model architectures. In this thesis, we study the use of evolutionary algorithm for neural architecture search (NAS). Moreover, we investigate the effect of integrating evolutionary NAS into deep reinforcement learning to learn control policy for ATARI game playing. Empirical classification results on the NASBench image dataset using different population selection and drop methods for the evolutionary NAS are presented to validate the usefulness of simple evolutionary algorithms in optimizing neural architectures, showing that even basic evolutionary algorithms are a well performing and easy to use approach. We also show the feasibility of using evolutionary NAS to extract good features that can improve the pong game playing score when the computational resource is limited.
Cuccinello, Andrew D., "An Empirical Study on the Efficacy of Evolutionary Algorithms for Automated Neural Architecture Search" (2022). Theses and Dissertations. 2964.