Genetic Algorithm-Driven Optimization Of Neural Network Architectures For Task-Specific AI Applications

Abstract

Artificial intelligence (AI) has attracted a great amount of interest in recent years. AI allows machines to perform different tasks that have traditionally required human intelligence. AI is useful in many different disciplines. One type of AI application is a neural network; task-specific AI consists of many different architecture types. Between neural network architectures and applications, there exists a complex relationship that makes selecting the right neural network architecture for the task a non-trivial challenge. Genetic algorithm optimization, a population-based evolutionary algorithm, is proposed to help find the most efficient neural network architecture for a specific task. We construct a performance prediction model that can predict the performance of an AI model based on the features of the problem. The performance prediction model is then used as a performance predictor to implement a behavior-learning strategy that can learn the search characteristics of an algorithm during the training process. The predictor is trained with machine learning models. By using the predictor, the learning speed of the final convergence process can be improved by making more informed decisions during the evolution strategy.