bot evaluation question - Backgammon Forum - Discuss Backgammon Strategy, Clubs and Books

Two Plus Two Forums Sports and Games Chess and Other Board Games Backgammon Forum hosted by Bill Robertie.

bot evaluation question

Post Reply Subscribe

...

08-09-2017 , 08:59 PM

7NTXX

newbie

Join Date: Jan 2010 Posts: 49

Maybe this is obvious but:

A 3-ply eval, as I understand, is looking 3 rolls into the future and taking an average of the positions that arise with ideal play.

But how is the bot evaluating those positions for winning chances? It is done recursively?

Does this make any sense?

Quote

08-10-2017 , 09:34 AM

Robertie

old hand

Join Date: May 2004 Posts: 1,954

A bot will have a static evaluation function that it will use to evaluate the end positions. The evaluation function gets constructed during the initial training of the neural net.

Quote

08-10-2017 , 02:45 PM

7NTXX

newbie

Join Date: Jan 2010 Posts: 49

Quote:

Originally Posted by Robertie

A bot will have a static evaluation function that it will use to evaluate the end positions. The evaluation function gets constructed during the initial training of the neural net.

How is that function tested for accuracy? Couldn't it potentially be very wrong in unusual or complex prime v. prime positions?

Quote

08-11-2017 , 09:21 AM

Robertie

old hand

Join Date: May 2004 Posts: 1,954

Quote:

Originally Posted by 7NTXX

How is that function tested for accuracy? Couldn't it potentially be very wrong in unusual or complex prime v. prime positions?

Yes, and in many kinds of position the static (0-ply) function will give wrong answers. That's why bots have a lookahead feature (1-pl, 2-ply, 3-ply, and 4-ply) as well as mini-rollouts (the XGR+ and XGR++ settings in Extreme Gammon). Combining these features with a pretty good 0-ply evaluation function gives the super-strong bots we know and love today.

Quote

Post Reply Subscribe

...