Backgammon Rollouts

Rollouts

Truncated rollouts

From:   Gregg Cattanach
Address:   gcattanach@prodigy.net
Date:   13 October 2002
Subject:   Re: Truncated rollouts
Forum:   rec.games.backgammon
Google:   Ennq9.1040$xU7.176@newssvr19.news.prodigy.com

> Could someone explain to me what truncated rollouts are? A truncated rollout just makes the number of moves specified and then uses the bot's evaluation function at the position obtained to accumulate the results that go into the rollout. For play vs. play comparisons (especially where both sides have simple or difficult plays), truncated rollouts probably give pretty reasonable results. *(Some caution is required when one side has very simple plays and the other is very difficult, like containment positions. In these the errors will pile up on one side, and that can skew the results.) To get the same accuracy for full rollouts that you get from a truncated rollout you usually have to do 40 times the number of games because a fully played out game has so much more 'noise' involved. And, of course, full rollouts take dramatically longer. So given limited CPU resources, on checker play comparisons truncation is probably quite useful and gives up only a small amount of accuracy. (I usually truncate at 11 moves.) For cube action rollouts, the absolute equity value is of critical importance, so any error the bot has on estimating the absolute equity of the resulting truncated position will find its way into the rollout results. Most people would say that truncation for cube action is probably the wrong way to go; full rollouts are required. Checker plays we're trying to answer which play is best, and we're not too worried what the absolute value of the best play is. For cube action the absolute equity is the only important value. Gregg C.

Did you find the information in this article useful?

Do you have any comments you'd like to add?