Distributed Monte Carlo Tree Search With Applications To Chip Design

Abstract

Monte Carlo Tree Search is a classic method in AI that builds up a search tree asymmetrically using random rollouts on a game tree. The work detailed in this thesis expands upon traditional implementations by allowing the capability of fully distributing each node onto different physical machines while enabling them to keep in constant communication. The ability to distribute work to other machines is a highly desirable capability that will allow users to save on single computer resources, enable an almost arbitrary level of scaling, and allow for the processing of states which previously would have been too large to run on a single computer realistically. When applied to the problem of automating the design of Printed Circuit Boards (PCB) from just a list of desired board specifications, this fully distributed search will allow increased search breadth and depth. This expands the computational limits of each action applied to the state, increasing the probability of finding an improved final state when compared to running the search on one physical machine. In this thesis, we discuss our motivating problem and the infrastructure changes necessary to enable this capability increase. We show results highlighting the potential improvements these changes will have on the process of generating a PCB design and identify significant areas for improvement.