Applications of Neural Architecture Search to Deep Reinforcement Learning Methods

Abstract

This research aims to investigate the impact of various Neural Architectures (NAs) on the performance of machine learning models in the context of the deterministic game of Othello, with the goal of providing insights into selecting optimal neural architectures for a given problem description. We constructed several fundamentally different neural architectures, including fully connected networks and convolutional networks, and assessed their performance across various metrics such as win rate, training time, and computational resource consumption. By correlating these performance changes between NAs with their relevant structural differences, we sought to offer technical insights that can guide the selection and composition of ML architectures for other problem-solving contexts with deterministic constraints analogous to those in Othello. Furthermore, this research emphasizes the importance of understanding the trade-offs between different NAs in terms of resource efficiency, learning capability, and adaptability to specific problem domains, thereby facilitating more informed decisions in the development of machine learning models.