Arimaa Forum - Print Page


    
      
        Arimaa Forum
        (http://arimaa.com/arimaa/forum/cgi/YaBB.cgi)
      

        Arimaa >> Off Topic Discussion >> Go
        
(Message started by: Nazgand on Apr 22^nd, 2012, 6:20pm)

Title: Go
Post by Nazgand on Apr 22^nd, 2012, 6:20pm

http://www.sciencedaily.com/releases/2012/04/120416100437.htm

http://iopscience.iop.org/0295-5075/97/6/68002/pdf/0295-5075_97_6_68002.pdf

Title: Go programming
Post by christianF on Apr 23^rd, 2012, 4:56am

Thanks, that's very much my area of interest, though unfortunately not of expertise. It is related to the question why some games are hard to program, but still (!) allow humans to play at a superior level, and the quest to unravel human decision making. I like to follow those developments, even if the underlying theory and its implementation go way above my head. :-/

The basic idea as I understand it, is to select a move based on statistical data of similar local positions in zillions of recorded games of amateurs, masters and grandmasters. The premiss in that case is the availability of such data. Go complies of course, but if the goal were to invent a game that is particularly nasty for bots while being 'human friendly', the development of the method would trail behind the development of human skill at the game.

Title: Re: Go
Post by hyperpape on Apr 23^rd, 2012, 7:18am

It always annoys me that these summaries can't get the basic data right: Go has a very good programming paradigm in MCTS. The past five years have seen very rapid progress, and a naive extrapolation suggests computers could be at the level of top professionals by 2016-2017. (A recent series featured a top bot beating Takemiya Masaki, who is a strong professional, with 5 and 4 stone handicaps).

Not that this paper isn't interesting, but I just get annoyed about the breathless "have we cracked go programming?" in the science daily. For all we know, we've already cracked go programming, and we just need to continue making incremental progress (fwiw, I'm actually pretty sceptical, and think MCTS will hit a wall somewhere around the professional-amateur boundary).

Title: Re: Go
Post by christianF on Apr 23^rd, 2012, 7:35am

on 04/23/12 at 07:18:49, hyperpape wrote:

Not that this paper isn't interesting, but I just get annoyed about the breathless "have we cracked go programming?" in the science daily. For all we know, we've already cracked go programming, and we just need to continue making incremental progress (fwiw, I'm actually pretty sceptical, and think MCTS will hit a wall somewhere around the professional-amateur boundary).

More than one 'wall' has proven to be illusionary in the past, and you can't tell what's around the corner. But I agree MCTS needs a lot of additional heuristics and it has the inherent drawback of being based on probability grounds, that is, good at general concepts, bad at long and thin tactical lines.

The basic drawback of the 'network method' seems to me that it tries to surpass human level while being based on it. Something like playing against the 'collectively best players'. That would make a very good bot, but could it top the best humans? ???