Distributed GPU Monte Carlo Tree Search

TL;DR

Solves checkers/reversi (why didn't I do GO?)

Includes self-play

Runs on GPU (CUDA) or CPU (C++)

Over 1 million game simulations/second on a single 8 year old 280 GTX GPU.

GPU version is quite efficient actually ( here is the score advantage plotted on the y-axis vs time, playing againt a single-core CPU)

Scales up nicely using MPI (Message Passing Interface) to a large distributed system (tested on a 2048-node supercomputer, up to 3.5M GPU threads)

Has a very minimal ssh-friendly interface

I used this code while working on my PhD thesis. The MPI version has been tested on the Japanese TSUBAME supercomputer.

THESIS + BIBTEX

Thesis Slides