bug-gnubg
[Top][All Lists]
Advanced

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

RE: [Bug-gnubg] Wildly different evaluations by CRASHED network


From: David Montgomery
Subject: RE: [Bug-gnubg] Wildly different evaluations by CRASHED network
Date: Wed, 3 Jul 2002 10:08:21 -0700

I think the easiest thing that could be done to
improve the performance of the crashed network
would be to split it into two networks.

The idea of having multiple networks is to have
specialist nets that understand a particular
type of situation.

The problem with the position like Ian Shaw posted
is that the position is completely different from
the two sides of the board.  

A network's category should be defined so that all 
the positions in its category are of the same type.
So I would suggest an "I'm crashed" net and an
"Opponent's crashed" net.

You can still get wildly different evaluations
with this kind of system, because odd and even
ply lookaheads are actually combining evals from
different nets.  But each net has a better chance 
of correctly evaluating its positions,  because 
it has been trained only on positions similar to
them.

>     GNU Backgammon  Position ID: AAAAt+3uAAAAAA
>                     Match ID   : cAkAAAAAAAAA
>     +24-23-22-21-20-19------18-17-16-15-14-13-+  O: GnuBg
> OOO |                  | O |                  |  0 points
> OOO |                  | O |                  |  
>  OO |                  | O |                  |  
>  OO |                  |   |                  |  
>  OO |                  |   |                  |  
>     |                  |BAR|                  |v (Cube: 1)
>     |                  |   |                  |  
>     |                  |   |                  |  
>     |          X  X  X |   |                  |  
>     | X  X  X  X  X  X |   |                  |  On roll
>     | X  X  X  X  X  X |   |                  |  0 points
>     +-1--2--3--4--5--6-------7--8--9-10-11-12-+  X: ian





reply via email to

[Prev in Thread] Current Thread [Next in Thread]