I don't think your current work is not valuable, and fully agree with you
that MM/SB-based engines have a much greater effect on strength, which is
clearly proved by Erica. But recently my feeling is that current MCTS almost
reaches its limit on 19x19. We will need another breakthrough/good ideas to
overcome KGS 5d or 6d.
Aja
----- Original Message -----
From: "Brian Sheppard" <[email protected]>
To: <[email protected]>
Sent: Wednesday, April 06, 2011 7:00 AM
Subject: Re: [Computer-go] 7.0 Komi and weird deep search result
I spend a lot of time on computer Go, but probably not in the right places.
My work currently focuses on expressing positional features using 3x3
neighborhoods, so that domain knowledge is easier to express as data
rather than if/else code. This is useful stuff, to be sure.
Instead, I really ought to build two systems: an MM-based engine to
prioritize moves in the tree search, and an SB-based engine for random
search in the playout. I am sure that these would have a much greater
effect on strength.
But I started the other effort, and I should continue for a while before
changing. Probably I will do MM next.
Brian
-----Original Message-----
From: [email protected]
[mailto:[email protected]] On Behalf Of Aja
Sent: Tuesday, April 05, 2011 1:35 PM
To: [email protected]
Subject: Re: [Computer-go] 7.0 Komi and weird deep search result
I might not really catch what you meant, but I wonder why. :)
Aja
-----原始郵件-----
From: Brian Sheppard
Sent: Wednesday, April 06, 2011 12:29 AM
To: [email protected]
Subject: Re: [Computer-go] 7.0 Komi and weird deep search result
I don't know if the worst could be worse; UCT convergence for a 1-ply
search
is a probabilistic function with an exponential bound. The bound for an
N-ply search is a tower of N exponentials: Exp(Exp(Exp(...Exp()))). Ugh.
Because of this bound, guessing good moves quickly is absolutely vital for
strong play from UCT. Which calls into question why I haven't taken MM and
Sim Balancing more seriously. :-)
-----Original Message-----
From: [email protected]
[mailto:[email protected]] On Behalf Of Petr Baudis
Sent: Monday, April 04, 2011 10:27 PM
To: [email protected]
Subject: Re: [Computer-go] 7.0 Komi and weird deep search result
On Mon, Apr 04, 2011 at 12:56:54PM -0400, Brian Sheppard wrote:
>> MCTS using RAVE prioritization *does* converge to game theoretic
>> values
in a
>> binary-valued space.
>Can you reference some more detailed analysis claiming this?
Theorem: In a binary-valued game of finite length, the RAVE score of all
winning moves converges to 1, provided that 0 < FPU < 1.
Oh of course, it is obvious. Sorry for being slow and confused.
But it seems it should be possible to prove that even theoretical
convergence in case of RAVE discrepecancies is much slower than with
plain UCT... Might be a fun exercise.
Petr "Pasky" Baudis
_______________________________________________
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
_______________________________________________
Computer-go mailing list
[email protected]
http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
