| |
BGonline.org Forums
I think I understand it now (request for review)
Posted By: MaX In Response To: I think I understand it now (request for review) (Frank Berger)
Date: Monday, 27 July 2009, at 10:38 a.m.
Frank said:
"I did forget that Gnubg includes cube decisions in it's evaluation tree."
I tried to dig this Janowksi stuff on my side a while ago and I'm now relatively sure of what gnubg does: Janowksi formula is used only at the last ply.
Let's say you have a given position and roll, 10 candiate moves and you're using 2ply (on all the 10, let's forget about move filters and pruining).
Each move leads to a different subsequent position that gnubg has to eval 2ply. For each subsequent position and each possible roll, gnubg will have to find the best move 1ply. This means that for each subsequent-subsequent position and for each of the 21 different rolls, gnubg will have to find the best 0ply move.
This starts with the cubeless 0ply equities spitted by the Neural Net (let's forget bearoff dBs), then Janowski is applied to convert to cubeful. The a cube decision is considered comparnig the different Janowksi equities (centered, owning, etc) and this cubeful 0ply equities are passed back for the 1ply eval. The 1ply eval will once again consider the cube action, but since it already has cubeful equities, there's no need to apply Janowski formula again. 1ply equities are then passed back to 2ply eval and that's it.
Hence, if the number of ply is deep enough, gnubg will nkow that you can double next trun and telling you that is a 0.050 error not to double now he assumes you will double next turn (if the situation allows it).
MaX.
| |
BGonline.org Forums is maintained by Stick with WebBBS 5.12.