Surely pondering can't give more strength benefit than perhaps 1.5x more
CPU
power?
Which program did you use for your experiments?
David
Erica against GnuGo, with short time setting.
I think the upper bound is 2x CPU power. If the opponent is too strong or
too weak for our program, then pondering does not matter much for the game
result. But, since almost all the present programs have similar MCTS
implementation, the prediction rate over a program of "similar strength"
should be not be too slow. Doubling simulations against such a opponent
makes much sense.
But of course, I agree with Fotland that time management should not be the
most important in ToDos. :)
Aja
-----Original Message-----
From: [email protected] [mailto:computer-go-
[email protected]] On Behalf Of Aja
Sent: Thursday, September 16, 2010 4:53 PM
To: [email protected]
Subject: Re: [Computer-go] Speculatively pondering
Actually in my experiments on 19x19, pondering gives a VERY big strength
increase. This result is shown in our paper "Time Management for
Monte-Carlo
Tree Search Applied to the Game of Go" to be published.
Aja
----- Original Message -----
From: "David Fotland" <[email protected]>
To: <[email protected]>
Sent: Friday, September 17, 2010 1:58 AM
Subject: Re: [Computer-go] Speculatively pondering
> Many Faces does not ponder, but it's on the todo list. I don't expect
> pondering to give a big strength increase compared to all the other
things
> queued up.
>
> David
>
>> -----Original Message-----
>> From: [email protected] [mailto:computer-go-
>> [email protected]] On Behalf Of [email protected]
>> Sent: Thursday, September 16, 2010 6:45 AM
>> To: [email protected]
>> Subject: Re: [Computer-go] Speculatively pondering
>>
>> Quoting Brian Sheppard <[email protected]>:
>>
>> >> In other words: a strong opponent will cause a lot of ponder hits
and
>> >> speculative pondering is the best way to search effectively.
>> >
>> > This makes sense, but actual measurements on CGS showed that
>> > speculative pondering was worse. At least for Pebbles.
>> >
>> > That experimental result is consistent with mathematical
>> > models, so I have confidence.
>> >
>> > Have you tested Valkyria both ways?
>>
>> No, what I wrote is just what I believe. There is a lot of things I
>> would like to test but this is on the todo list.
>>
>> Currently pondering tend to be inefficient for a different reason
>> since Valkyria have no garbage collection and quickly fills memory
>> with 4 fast cores. In forced sequences it does not ponder at all
>> because the tree is already full.
>>
>> Magnus
>>
>> _______________________________________________
>> Computer-go mailing list
>> [email protected]
>> http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
>
> _______________________________________________
> Computer-go mailing list
> [email protected]
> http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
_______________________________________________
Computer-go mailing list
[email protected]
http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
_______________________________________________
Computer-go mailing list
[email protected]
http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
