blog: AI backgrammon player got better with rollouts

[nikos]

here's the comment area for today's blog post found at
https://www.zabkat.com/blog/tesauro-bac ... kahead.htm

[Howard_Milano]

Thanks for these blog posts. Yes, there are other people who think writing an AI backgammon bot is interesting. My best performing bot uses TD(0) using a basic NN with a 80 hidden node layer, similar probably to what you use. The TD(0) code doesn't use the extra code from Tesauro (eligibilities), just train each move and at the end. I use an extended notation, some extra fields compared to Tesauro. Code beats PubEval around 58% of the time. This is my second bot research, my first bot I wrote years ago, before and during your first blog post time. Now I see there is a follow up. I'm trying to use your 2-ply search idea, still in the testing phase (takes forever). ChatGPT thinks high 50's% is about max without using extras (a 128 hidden layer performs as well as a 80 hidden layer). Using 2-ply search and a race board evaluator can push it to the low to mid 60's%. Professional bots are probably high 60's% to somewhere in the 70's% (no data I can see)?? PubEval is weak, of course, but it's pretty tough. Your blog was a great encouragement to keep going, it's definitely not an easy problem to solve, for me at least. But it's magic when the bot beats me easily playing backgammon.

[nikos]

as chatGPT itself, a good backgammon AI agent needs a good representation of the board. Tesauro's board state vectors are very important and whenever I tried other encoding schemes the player got rubbish.

[Howard_Milano]

I agree that a good points description plus bar and off counts, which Tesauro's notation is, is necessary. I used 6 inputs per point, versus 4 inputs for Tesauro, and added further inputs for pips, pip leads, blot counts and more. All those extra inputs the NN can figure out on its own, but adding extra inputs allows the NN to learn more.