Re: [computer-go] 9x9 search, tsume-go

[tesuji@xxxxxxxxx]

Comparing 9x9 Go and Tsume-Go is like comparing apples and oranges.

>
> On the other hand, quite a lot of effort has been put into solving
> tsume-go problems (unsurprisingly, since most strong players believe
> "if you want to get stronger, study tsume-go" is largely correct). To
> my knowledge the techniques which work (which definitely include a lot
> of search!) tend to blow up badly before the solidly enclosed area
> grows to 9x9. So while I agree with you (Mark) that the opponent
> making a whole-board fight is a key problem for existing software, I
> am more skeptical than you are that a reasonable amount of programming
> effort targeted at 9x9 Go will overcome the problem. Maybe 7x7...
>

Has a lot of effort been put into tsume-go? Maybe I'm a bit out of touch
nowadays, but I always felt most programmers put relatively little effort
into tsume-go. But I do understand why, it's hard to incorporate the
single-goal program usefully into a more complex whole and make it worth
the investment. I've found that PN-search is extremely promising to
overcome that problem, but haven't heard yet that anyone is actually using
PN-search in a playing program.

> To me the tsume-go thing is a very interesting contrast to computer
> Chess, and I can't help being reminded of it when people bring up the
> success of brute-force search in Chess. ISTR Bob Hyatt saying that by
> the early 1980s or so, Cray Blitz could blow through an entire book of
> hundreds of tactics problems before a human would be able to open the
> book and read the first problem. (And whether or not I'm remembering
> the quote right, the underlying fact seems about right.) In Go, no one
> seems to know how to do anything like that for problems at the level
> of _Graded Go Problems for Beginners_ volume 4.[1] (I mean the 90+% of
> problems there which have an easy-to-define sharp objective, excluding
> problems like #53, "what's the best move" in an early middlegame
> whole-board position.) And most of the solution diagrams in _GGPfB_
> seem to be 9x9 or smaller.
>
> I think I can speak for a substantial fraction of Go programmers here
> when I say that if a Chess programmer slumming in the Go programming
> world could use even a week of runtime, executing a search algorithm
> recognizably adapted from Chess and not overtuned/overtrained for a
> particular problem suite, to solve even the most tractable 80% of a
> set of hundreds of problems at that level, we'd be pretty damned
> impressed. (And for a smaller, but still significant fraction when I
> say that as long as the Chess programmers can't, we'll remain somewhat
> skeptical about weakness of Go programs being a reflection of
> insufficient attention paid to brute force search.:-)

Some years ago I've worked for a year and a half on Tsume-Go together with
three others and reached results that greatly exceeded the expectations I
had beforehand. Not in the least because one of us four was a "Draughts
programmer slumming in the Go programming world". I think with a few more
man-years work on Tsume-Go it would be possible to make a program that
would do better on an average Tsume-Go book than almost any amateur
player.


> (I'm aware that people -- including, I think, David Fotland on this
> list -- have reported higher than 80% success at choosing the right
> move in such problems. But what I mean by "solve" is to arrive at a
> 100% certain answer, so that the program either returns "this *is* the
> move" or "I'm not absolutely sure, I need to think more". That is,
> "solve" as pros solve such problems, or as computers so
> lve the
> corresponding chess problems.)
>

Maybe in Go you wouldn't be able to reach the 100% certainty level, but
it's not so relevant. You can reach a high level of play, which is what
it's all about.

Of course you can't translate Chess programming technigues straight into
Go programming and get immediate results. However, if a good Go programmer
would spend a few man-years on making a good evaluation and move-selection
for 9x9 Go and combine it with Chess search technigues (why reinvent the
wheel), I think you'd get a program much much stronger than any of the
exisitng Go programs (for 9x9 that is) which was the point of my previous
mail. I don't know if you'd reach 1-dan level but I think for 9x9 it would
probably be the best approach.

    Mark Boon

_______________________________________________
computer-go mailing list
computer-go@xxxxxxxxxxxxxxxxx
http://computer-go.org/mailman/listinfo/computer-go