[Top][All Lists]

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

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

From: Øystein O Johansen
Subject: Re: [Bug-gnubg] Wildly different evaluations by CRASHED network
Date: Wed, 3 Jul 2002 14:38:04 +0200

Yes, I think the crashed network still lacks a bit of training.

The two position types gnu misevaluates the most, (in my opinion ) are:
- Crashed position (as the one you show)
- Low contact positions.

Crashed positions
Crashed position has its own network, which I think need some more training
to be accurate enough. You will often find positions like the one you
posted, where the evaluation of odd and even ply differs a lot. You will
also find positions where rollouts by gnu differs from rollouts of other
bots, because gnu misplays these positions more than other bots. Training
of the crashed network with databases of rollouts will therefore be a very
slow procedure. Joseph: Is there any progress in training of the crashed
neural net? If someone coordinates a database to be rolled out, the
rollouts could be distributed to users with fast computers to roll them
out. The user will after the rollouts return the result to the coordinator
who then will start the training. After this training a new crashed network
can be released. The whole of this procedure must then be repeated until we
have a good evaluation from the crashed neural net, since the rollout
results won't be very good in the first run-throughs either.

Low contact position.
Gnubg plays low contact positions like Snowie plays bear-in situations!
That's really bad! The race neural net seems to be OK though. I therefore
think this can be solved by adding a meta-phi like scheme for this kind of
positions. Maybe just using the I_BREAK_CONTACT input to make some kind of
interpolation between the contact neural net and the race neural net.

The overall strength of the program is probably improved most by improving
the crashed position neural net.

Just some ideas,

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.

reply via email to

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