Don Dailey wrote:
(snip)
In my opinion, the insight that Chrilly articulated was that all of
sudden we are now all using some type of global search - the very
idea was considered blasphemy just 2 or 3 years ago.
That may be too strong a statement. It may have not been popular but
many people consistently believed global search must be a big part of
any strong playing program, myself included. Not searching using the
same techniques as used for chess, but IMO certainly searching has not
ever been altogether dismissed nor considered blasphemy. Look back at
posts around 10 years ago (when I first joined the list) and probably
since its inception and you'll find this to be true. I personally wrote
about it on several occasions suggesting that to counter the evaluation
problem the search needed to go very deep and even talked about
"sampling" the tree. Other probability based searches have been studied
and written about in academic papers and on this list as well. The
crucial combination of techniques didn't bubble up, but not for lack
of trying.
But I have to admit that personally, I have many more ideas than time
with a full time job. Over the last 10 years all I've really done is
play around with various algorithms and ideas, study the game of go,
collect and read a lot of published papers, and keep up on this list -
occasionally posting. My wife still doesn't understand my putting this
much time into it! ;-) This is the kind of thing that could consume a
person. I don't know if particular ideas would pay off or not because I
haven't been able to put in the proper time to focus. In spare time, on
and off over the years I've only done a few experiments and algorithms
mostly focused on partitioning, goal directed and hierarchical searching
methods. This negligible computer-go work, some plans and a few ideas is
the extent of my would-be program "KatanaGo".
Regardless, it has been great fun watching the progress of computer-go
over the years and the current flurry of activity with MC/UCT is quite
exciting!
As I wrote in a post in early Feb of this year (paraphrasing from
memory), I think the main reason MC/UCT works is because it goes deep
(nearly always to the end) and tends to find paths with more favorable
possibilities and more importantly avoid paths littered with problems.
Though I'm still pretty amazed by how well it plays, but that's the
power of the law of large numbers at work. Correct?
As for the point that different paths may converge on similar methods, I
agree that could be very plausible scenario, but there is a very long
way to go yet...
_______________________________________________
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/
