I still don't understand your point. Are you just trying to say computers have a long way to go to beat really strong humans?
If so, that doesn't have anything to do with what Chrilly said or my response to him. On Wed, 2007-01-10 at 19:10 -0800, steve uurtamo wrote: > i'm saying that if a factor of 50 extra time can make your program > strong enough to be 'impressive', then you know that you're within > a reasonable hardware 'scaling distance' of playing that strong at > regular timeframes. Ok, I understand this and agree. > if, on the other hand, you're *not* (which would be easy to see > by performing my experiment with the putatively scalable player > getting the factor of 50 time advantage over a few otherwise > similarly-ranked computer players who have normal time constraints) > seeing a massive advantage with a factor of 50 time advantage, > then you're *not* within a small scalable factor of having a piece > of code which, whatever the technique and hardware, is objectively > 'strong'. This is not what we were discussing, but I think I see your point. All you are saying is that we can measure how far we are from some arbitrary goal by playing very long time-control matches. > given that the second is the case, worrying about performance as a > function of time as opposed to worrying about absolute strength is > a little bit silly. No, your argument is wrong. It's ONLY about performance as a function of time. Absolute strength has nothing to do with it as we already know how to make a perfect program. If your goal is to make a stronger program then you have to think in terms of how to make it play better moves in less time. > to state this more simply, it only really matters to measure strength > as a function of thinking time if changing the thinking time affects > the level of play significantly. and given that it doesn't, it does make > good sense to emphasize the 'absolute' level of play. Who said it doesn't? Hiroshi Yamashita posted the run times of the program. Multiply all of those times by 100 to see how LONG it would take to run on computers of just a few years ago. Either the programmers are getting stupid, or they actually have scaled their program up to modern hardware. Average expended hours. KCC 17m51s Gnugo 3m14s, Opteron248(2.2GHz) Haruka 11m53s Gnugo 4m28s, Opteron248(2.2GHz) + AthlonXP 2100 +(1.73GHz) Go4++ 4m59s Gnugo 2m18s, Opteron248(2.2GHz) Handtalk 2m42s Gnugo 5m22s, AthlonXP 2100+(1.73GHz) > to see what i'm saying in action, if we were to dump a few copies > of gnugo into a slow-motion tournament, one of which always played > black with 2 stones handicap, one of which always played black > with 3 stones handicap, etc., etc., and all of which were told that > they only had, say, 30 minutes to make all of their moves (while > their opponents were given 24 hours), we could see just how much > stronger everyone's programs are *even with* a factor of 50 time > advantage. GnuGo is probably not a very good example of a scalable program. It's a reasonably good program by today's standards, but it probably wouldn't play a lot better on a computer 1 billion times faster. I don't know where the tradeoff is with mogo vs gnugo, but Mogo is the strongest future program of the two - assuming neither one changes. Mogo would overtake gnugo on a computer you will buy in a few years. However, by then gnugo will have improved in some way - perhaps by being more mogo-like or doing something completely different - but whatever it is that it does different it will not run on today's computers - it will do a lot more calculations. > my guess, after watching the last tournament, is that all of the > players involved last time would still be within a few stones of gnugo -- > that the factor of 50 time advantage didn't make anyone massively > stronger. heck, i doubt that a factor of 50 time advantage would make > *gnugo* more than a few stones stronger than itself. of course, all > of this is easily verified. :) A "few stones" is a heck of a lot. If 1 stone is approximately 100 ELO points then a "few stones" is huge. A chess program gets about 60 ELO for a doubling - a factor of 50 times is only a few doublings, perhaps 300 ELO or 3 stones in GO terms. Of course this is all pretty crude estimates and it's hard to compare. 50 X speedup sound rather impressive but it's not that much. It's probably made go programs about 2 or 3 stones stronger over the few years that it took to get hardward 50X faster about what you would expect. You are not thinking straight if you think 50X speedup is supposed to make it more than 2 or 3 stones stronger. It's no wonder you are disappointed. I think the programs probably did play a lot stronger - it's just not that impressive when you compare it to 1 dan players (which I don't know why we insist on doing that.) It's a human perception issue - when someone sucks, we usually don't distinguish how much they suck so even if the improve a lot, we still think they suck and if you suck no one cares how much. I noticed that good players can't tell the difference between bad play and random play. I discovered that there is a huge gap in random play and play with almost immeasurable intelligence added. If you can beat a random player 19 out of 20 times, a good player still thinks your play is random and have no appreciation for a massive improvement - it's still random to them. It's why people smoke cigarettes. They have no respect for the odds. They believe the chances of it killing them on any given day is very tiny - so tiny that it's like round-off error and inconsequential. Any kind of risk taking is like this. Something twice as dangerous or life threatening doesn't matter when the chance of dying is less than 1 percent. > s. > > > > > > ____________________________________________________________________________________ > Cheap talk? > Check out Yahoo! Messenger's low PC-to-Phone call rates. > http://voice.yahoo.com _______________________________________________ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
