[Top][All Lists]
[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: [Bug-gnubg] Question about the Hint Window Evals
From: |
Pablo Sangster |
Subject: |
Re: [Bug-gnubg] Question about the Hint Window Evals |
Date: |
Wed, 2 Feb 2005 13:13:34 -0000 |
Øystein O Johansen wrote:
Firstly my apologies for remaining anonymous.
My name is Pablo Sangster
Thanks you for your quick response.
Hi, Pablo! Nice to meet you!
You can of course press the "..." button and turn off the pruning
neural nets for the evaluations done with the "Eval" button. If I
coded the cache key calculation correctly to take the prune net
setting into account, the evaluation will be redone without use of
the pruning nets.
Yes thank you for this tip. It didn't occur to me.
This was the final argument for setting pruning nets default on.
Yes I fully understand how the pruning neural nets work.
Good!
However my speciality is backgames and I am finding that there
are enough differences between the prune and full nets to matter.
Yes, you're right. In some backgame positions the pruning neural nets
will possibly do some bad pruning. However it is not sure the normal
nets will do the right move either. ;-)
I wanted to check what move was chosen at it's optimum setting (at 2-ply)
and yes
you're right in that where it doesn't understand the position well, it may
not find the "best" play anyway.
If there is a position type where I believe that turning the pruning
nets off can give some improvments it must be in backgames.
BTW I think GNU could do with some more training in developing
backgames...
I've started to look at Josephs training code and tools and I do
some small test with differnt nets. Howevere I belive it's really
hard to improve the current nets much more as they are without
splitting further or introducing other inputs. (However I don't
believe introducing other inputs is the best strategy either.)
A long long long time a go (in this galaxy), Joseph tried to make
a net specially for backgames and prime position. However, this net
wasn't a success and Joseph abandoned the net. This must have been
at least 5 years ago. It must have been in the sourceforge period
of the project.
Joseph and the Gnubg team have done tremendous work in bringing the
program up to its current strength. I always compare it's moves to Snowie 4
(which
has specifically been trained for backgames) and I believe and other experts
that I have spoken to
agree that Snowie's moves appear "more correct".
So I presume Gnubg could be improved in this part of its game but as you
pointed out it isn't
that simple without affecting the nets overall.
What 'part' of the backgame do you find flaws? In developing a
backgame? I really find the positions where gnubg is developing a
backgame really strong. I see it make most mistakes after it hits.
It dosn't understand how contain the checker properly. Specially
if it's hitting late.
It will commit to a back game too early by playing loose kamikaze plays
where it
still has a very flexible position which could easily be won going forwards.
I believe Gnubg could also be slightly improved in "holding games" and " 1
checker back games".
I think it generally over values the holder.
Pablo
-Øystein
-------------------------------------------------------------------
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.