Designing Artificial Intelligence (AI) is still reserved for experts, and the existing design paradigm follows a data-driven approach: domain experts start with a hypothetical model, verify the model on a task-specific dataset to acquire performance metrics, then revise the model based on prior experiences, hoping to improve the model in the next loop. This thesis seeks to build an intelligent agent to substitute domain experts in this design process. I start with formalizing the current design process as a computational model, upon which I further investigate issues to the algorithmic efficiency and system utilization to build an agent that algorithm and system can synergistically work together. Specifically, I propose a new black box solver, Latent Action Monte Carlo Tree Search (LA-MCTS), to address the sample efficiency and build a deep learning framework to expand the design space far beyond the available GPU DRAM. These results collectively provide a partial path toward AI democratization by creating a practical MCTS-based AI agent that efficiently designs complex AI without experts in a reasonable amount of time.
