by Douglas Zare
3 July 2012
Some of the first backgammon neural nets were trained to play properly in bearoff races. They were utter failures! Although races are simpler than contact positions, and it is simpler to bear off than to bear in, it turns out to be surprisingly difficult to train a neural net to understand the bearoff as well as people do, at least for relative evaluations.
Neural nets need a lot of help to be able to choose between checker plays up to our high standards. This is one of the reasons bots rely on databases for the end of the game.
In this column I'd like to show a few fun extreme examples in the bearoff. These are taken from a bearoff database.
Effective Pip Count
Estimating the effective pip count is not hard with practice, and it's a great way to tell how large your advantage is in the race. Can you determine whether a position is a take or pass from the effective pip counts?
|The rest of this article (28.21 K) is premium content. Please subscribe below.|
Article text Copyright © 1999-2018 Douglas Zare and GammonVillage Inc.