Re: [Bug-gnubg] Training of the crashed neural net. What would help?

[Joseph Heled]

Thanks, at least I understand what it is all about.

My answer is the generic one,
 - GNUbg evaluation is wrong here.

 - Given time, I can come up with an unlimited number of such positions.

 - It is easy to pick on a specific position or even specific class of 
positions. If you want to improve, you need a net that does better over ALL 
range of positions in it's class.

 - All you suggestions have merit. However, from personal experience, it is 
extremely difficult to improve now. The gains are getting smaller and smaller 
for larger amounts of work and resources. For example, you can have 14 x 14 
nets, as per number of checkers off. But who can tell if it is worth it? 
Training all of them and storing them is a big hassle. Adding more inputs is 
another Gray area. I was never able to find a method to tell me which inputs to 
add and which to keep. There were several inputs I said to myself "They must 
improve the net ability to detect such and such", but was not able to get any 
improvement.

-Joseph


Øystein Johansen wrote:
> Joseph Heled wrote:
>
> |
> | While I am aware of the "containment" problem, this does not seem one of
> | them.
> |
> | The rollout give me 82%+0 18%+2.6%
> |
> | This seems reasonable since if you close the board and pick the second
> | checker you get 86%
> |
> | Can you please be more specific?
>
> The position didn't originate from me. It comes from a poster who calls
> himhelf (or herself) 'rew'. He/she writes:
>
> I've seen gnubg misevaluate this kind of position time after time..
>
> <image of the board />
>
> Alert: wrong double ( -0,1520)! [bad]
>
> Cube decision
> 2-ply cubeless equity    +0,3013   
> ~     67,40% 0,00% 0,00% - 32,60% 4,03% 0,64%
> Cubeful equities:
> 1.    No double    +0,5346   
> 2.    Double, pass    +1,0000    +0,4654
> 3.    Double, take    +0,3826    -0,1520
> Proper cube action:    No redouble, take (24,6%)
>
> A quick 0ply rollout indicate that the position is a small pass. As many
> know, after closeout, you would need 11 off to take..
>
> <end of quotation>
>
> My posting to the mailing list was more like a general suggestion and
> trying to bread some ideas... nothing else.
>
> -Øystein