Re: [Bug-gnubg] GNU backgame eval + bug

[Øystein O Johansen]

The averageing scheme that Kit suggest, is IMHO just a bandage. Better
would be if the program checked the difference of the 6-ply and 7-ply, and
then gave the user come kind of warning if the difference was more than,
say 0.05 equity.

Warning: The rollout results shown very inconsistent results at 6-ply
compared to 7-ply. It is therefore suggested that you extend your rollout
to a deeper truncation of a full rollout.

The real solution will be to have better trained nets. The crashed net
still needs some training, and I hope we can see some improvment when
Joseph returns.

-Øystein




                                                                                
                                   
                    Jim Segrave                                                 
                                   
                    <address@hidden        To:     Albert Silver 
<address@hidden>                               
                    .net>                cc:     GNUBackgammon bug reporting 
<address@hidden>, (bcc: Øystein O  
                    Sent by:             Johansen)                              
                                   
                    bug-gnubg-adm        Subject:     Re: [Bug-gnubg] GNU 
backgame eval + bug                      
                    address@hidden                                              
                                       
                                                                                
                                   
                                                                                
                                   
                    07.10.2002                                                  
                                   
                    10:22                                                       
                                   
                    Please                                                      
                                   
                    respond to                                                  
                                   
                    jes                                                         
                                   
                                                                                
                                   
                                                                                
                                   




[snip Neil's posting]

And the folloup from Kit:

  Yes, I have seen this oscillation effect often, and it can be
  irritating. It can also happen in a truncated rollout. The result
  after 7 rolls can vary quite a bit from what it would be after 8 (or
  6) rolls because of the oscillation effect -- keep in mind that we
  are talking about an evalution.

  The solution for the rollouts: When doing a truncated rollout (say 7
  rolls deep), take the AVERAGE of the equity estimate after 6 rolls
  and 7 rolls. This would require very little reprogramming, and would
  eliminate the oscillation effect and give us more accurate
  results. GNU people -- are you listening?

  Kit

Anyone have any opinions on whether this is a valid approach or just a
bandage to cover up this sort of inconsistancy? It certainly would
avoid some confusion for users, on the other hand, the very fact that
the equity is oscillating is a sign that either gnubg is weak here or
the position is one with a very unstable equity.

Should we assume it's an converging series which oscillates around the
terminal value and try to extrapolate from all the previous plies how
the convergence is actually working? (I'm not volunteering).


--
Jim Segrave           address@hidden


_______________________________________________
Bug-gnubg mailing list
address@hidden
http://mail.gnu.org/mailman/listinfo/bug-gnubg





-------------------------------------------------------------------
The information contained in this message may be CONFIDENTIAL and is
intended for the addressee only. Any unauthorised use, dissemination of the
information or copying of this message is prohibited. If you are not the
addressee, please notify the sender immediately by return e-mail and delete
this message.
Thank you.