This repository holds my backgammon-playing AI, which uses TD-lambda learning and a neural net, following Gerald Tesauro's TD-Gammon papers. I worked on this largely in May of 2020 as a quarantine self-edification project. To that end, I tried to do everything myself, instead of relying on external packages. That means everything is written more or less in stock python (with a little bit of NumPy), so it's not super fast—my five-year-old laptop can run about 4,000 games in an hour, or slightly more than one game a second.
If you'd like to try to play the trained net, you can run game.py. I'll warn that this interface is extremely rough, because I built it for debugging, not enjoyment. The rolls are given as tuples with each die. To enter your moves, type <point>-<num>, <point>-<num>. So, if I rolled a (3, 1) to open, I would type 8-3, 3-1 to make my 5-point. For doubles, you can type four sequences, and if you have no valid move you need to type something like 6-0. To enter off the bar, type 25-<number entering>. (Like I said, the user experience is, uh, unpleasant.)
Hope you have a great day!