[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: [computer-go] Computer Olympiad photo
the point i'm making is that the advocated solutions at the 2003 advances
of the international computergames, is that several chess programmers who
showed interest in solving 9x9 go by brute force: "because some uni guy
with a programming technique from some decades ago can solve 5x5 go, we as
professional search guys will be able to do that to 9x9 hands down by brute
force", are with regards to go and especially in 19x19, overestimating
search and underestimating evaluation.
Who is talking about SOLVING 9x9 go? I just want to make a program
play 9x9 go better than current programs.
Global search is a reasonable attempt, because your smart super
intelligent evaluation only based approach has set such a low bar!
How many times have we heard that even the best 9x9 go programs are
terrible? These are the cream of the crop programs using the
evaluation based approach.
But when the issue of global searching comes up, all we hear about is
branching factors and how hard go is, ignoring the fact that current
approaches have not shown us anything either.
Show us the light. I want to see how you have solved the problems you
say global searching with intelligent evaluation cannot solve.
> The only thing we know is that the brute force search approach they propose
> is not working for computerchess anymore, because for any serious
> programmer the statistical evidence that only more and better debugged
> knowledge works has been delivered long time ago (i remember some clear
> experiments that you did yourself already years ago showing this kind of),
> so why would searching deeper work suddenly for computer-go?
I did some experiments that clearly showed that the quality of the
evaluation function mattered more as you searched deeper. A better
evaluation function is always better given a certain depth, but if you
plot the improvement graphically you will see that evaluation becomes
more important with depth, the graph lines are not parallel.
This verifies my conclusion that the search MAGNIFIES the evaluation
function. It pays to have a great evaluation function. But my
research NEVER indicated that the evaluation be so advanced that a
chess program should always be doing 1 ply searches.
- Don
What we know is that the tactical local search in computer-go is faster
than the global search of the knowledgeable chess software concepts and
that they can play quite quickly.
What we also know is that the proposed brute force techniques when applied
by the researchers mentionned, that those researchers all share that they
not only know very little from chess, but even far less from go.
With all respect, but i look like a go worldchampion in level when compared
to these programmers. It appeared that some even didn't know you do not
play in the middle of each square!
Now getting 1 dan within a year, knowing that designing a FPGA engine takes
2 years at least, is far from realistic again.
Of course you can try to buy a 1 dan certificate from a professional
player, but that's not really going to be much of a proof in such a bet.
>I don't advocate brute force either, I believe the approach that will
>eventually take hold as computers become faster is that global
>selective search will gradually be introduced into Go programs (the
>ones not already doing it.) These searches won't solve the 150 ply
>problem you mentioned, but they will solve lots of short term
>oversights and make the programs much stronger. YOU DON'T HAVE TO
>SOLVE EVERY PROBLEM TO MAKE IT A LOT BETTER. Chess programs improved
You will fail to find the first move of any fight without knowledge.
The proposed methods all share that they use zero knowledge.
When i proposed at the ICGA meeting that they should make something that
mirrors patterns, they started to think about it.
"No need, we'll hardcode it all, goes faster!!"
The slapstick continued daily there.
"This game is much simpler than chess at 9x9, because all the pieces are
equal, in chess you have so many strong players who know everything about
every piece".
Or to quote the reigning world champion computerchess, who by the way
doesn't know the rules from 9x9 go: "You can search this 9x9 computer go
till the end, no problem".
When i replied that there are like 70! possibilities in the crucial positions:
"No way, in chess there is on average 40 moves and my branching factor is
way less than 3 there too and i get 19 ply, and those commercial go
programs are happy with a few ply only"
What i feel is that several chess programmers are assuming that the average
go programmer knows less from search than some scientists in computerchess
from the start of the 70s knew.
Though this is true for university programs, in general this is dead wrong.
There is an arrogance there which after my opinion should be fought out
over the board. I feel some commercial go programmers should enter his
program in the 9x9 ICGA go olympiads to let the chess programmers feel
where their place is in history... ...in the far past
>very very slowly over a long period of time. They never solved chess
>or even came close. You can embarass any chess program by giving them
>problems that humans can easily solve but they can't. Knowledge
>engineering in GO will get better, but there is a point of diminishing
>returns where it takes a lot of extra computing time to make the
>smallest progress. Some of that time will be better spent doing a
>search.
>
>Let me ask the programmers this hypothetical question: If I gave you a
>computer 1 trillion times faster than what you have now, would your
>program play a lot better? If you changed your program, what would
>you do to utilize those extra compute cycles?
>
>
>
>- Don
>
>
>
>_______________________________________________
>computer-go mailing list
>computer-go@xxxxxxxxxxxxxxxxx
>http://computer-go.org/mailman/listinfo/computer-go
>
>
_______________________________________________
computer-go mailing list
computer-go@xxxxxxxxxxxxxxxxx
http://computer-go.org/mailman/listinfo/computer-go
_______________________________________________
computer-go mailing list
computer-go@xxxxxxxxxxxxxxxxx
http://computer-go.org/mailman/listinfo/computer-go