On Fri, 2006-11-24 at 13:38 -0800, steve uurtamo wrote: > on a practical note, i think that MC is a great > idea for 9x9, and might even be a great idea as > a subset of a larger piece of code that employs > human knowledge, but that MC will never beat a > decent human at 19x19. the time/space limitations > are just too great. > > s.
Hi Steve, Just a few years ago these kinds of statements were made about computer chess and at the time they appeared to be REALLY SAFE predictions to make. Computers were doing 4 or 5 ply searches, they played absolutely horrible by master standards and it was absolutely inconceivable that a computer would ever search 2 or 3 ply deeper in the foreseeable future. Not only was that underestimated, but it was commonly believed that even 3 or 4 more ply wouldn't help it much. The improvement of even 1 ply was seriously underestimated. The reasons given are the SAME reasons given for computer GO. It was common to see mistakes that would require many extra ply for the computer to wake up. The common wisdom of the day was that an extra ply was close to worthless. I think the biggest myth of all was that humans were good at chess and that a computer had to be better at EVERYTHING in order to beat the human masters who were practically given god-like status. I'm probably a bit older than most on this list - and I have a very strong sense of historical perspective. I was there when "micro-chess" for the TSR-80 came out and I was bitterly disappointed with how badly it played. Back then I completely sucked at chess and yet I could beat the program without wasting many brain cells. The most common reason give was the same one you just made, "The time/space limitations are just too great." And that did not seem like a silly statement at all - the arguments were powerful and convincing ... and wrong! The typical argument started by showing a position that required a few extra ply to see that a human master knew almost intuitively. Then a projection was made about how much extra computing power was needed and the conclusion was completely obvious - it couldn't possibly happen in your lifetime! Then it went into wild speculation - even if a computer COULD look 10 ply deeper it wouldn't make much difference because computers can never have the positional understanding of even an advanced beginner, let alone a master. If you look back over the years, you will see that no matter how much progress had been made, many believed we had hit a wall - that further depth wouldn't help any while reluctantly admitting that the last increase did help a lot, but only because of tactics - after this tactics won't matter any more. This is just typical of us humans - we tend to have a very limited horizon ourselves and we only see the immediate surroundings - a shortcoming that computers are accused of. A lot of people take exception to this and say, "yes, but chess is nothing like GO" and the cycle starts all over again. And it's true - GO is more complicated and has it's own special difficulties. But I am not so pessimistic and I would never embarrass myself with claims like you made. You could turn out to be right - but I think it's rather bold of you. Perhaps in 40 years, if you are still alive - you will have to eat your words! Along the way there will be many other advancements. I can't envision a pure Monte/Carlo exactly like we have now doing the job - but even in chess there were numerous advancements to the basic alpha/beta idea. The old programs would not stand a chance on modern hardware against modern programs. This will surely happen in GO too and is part of the reason I think you are probably wrong and shortsighted just like the chess masters were who predicted it would take hundreds of years and basically embarrassed themselves. - Don _______________________________________________ computer-go mailing list [email protected] http://www.computer-go.org/mailman/listinfo/computer-go/
