Author: Robert Hyatt
Date: 09:53:37 07/09/02
Go up one level in this thread
On July 09, 2002 at 02:24:55, Uri Blass wrote: >On July 08, 2002 at 23:18:00, Robert Hyatt wrote: > >>On July 08, 2002 at 14:49:22, Chris Carson wrote: >> >>>On July 08, 2002 at 14:26:22, Christophe Theron wrote: >>> >>>>On July 08, 2002 at 12:36:01, Robert Hyatt wrote: >>>> >>>>>On July 08, 2002 at 12:15:06, Uri Blass wrote: >>>>> >>>>>>On July 08, 2002 at 11:32:38, Robert Hyatt wrote: >>>>>> >>>>>>>On July 08, 2002 at 00:32:42, Christophe Theron wrote: >>>>>>> >>>>>>>>On July 06, 2002 at 20:15:06, stuart taylor wrote: >>>>>>>> >>>>>>>>> >>>>>>>>>> >>>>>>>>>>I suspect that search may see that the right move help to push the opponent king >>>>>>>>>>closer to the corner relative to the wrong moves and it may be enough. >>>>>>>>>> >>>>>>>>>>Uri >>>>>>>>> >>>>>>>>>Yes, that looks like the best thing to try and work on, doesn't it? >>>>>>>>> >>>>>>>>>If not, can I ask two questions?: >>>>>>>>>1)What should be done during the near future to push computer elo forward as >>>>>>>>>much as possible? >>>>>>>>>2)If Deeper blue was really much stronger than todays tops, what was that due >>>>>>>>>to? Better long-term planning? Seeing deeper? >>>>>>>>>S.Taylor >>>>>>>> >>>>>>>> >>>>>>>>Huge speed. >>>>>>>> >>>>>>>>It was doing most things worse than the best micro programs, but it was doing it >>>>>>>>so fast that it was eventually stronger. >>>>>>>> >>>>>>>>Hum... Let me rephrase for the sensitive people out there. There was nothing >>>>>>>>Deep Blue did better than the best micro programs. But it was so fast that it >>>>>>>>allowed it to hide its defficiencies. >>>>>>>> >>>>>>>>Shit. That's not very diplomatic either. Let's try again: Deep Blue was build >>>>>>>>around a concept outdated by 2 decades but fortunately it was so fast that >>>>>>>>nobody noticed until their creators published their paper. >>>>>>>> >>>>>>>>Oops... OK, once again: >>>>>>>> >>>>>>>>Bob likes Deep Blue a lot, and that should be a reason good enough to convince >>>>>>>>you that it was well designed. >>>>>>>> >>>>>>>> >>>>>>>> >>>>>>>> Christophe ;-) >>>>>>> >>>>>>> >>>>>>>Er... excepting one game by Fritz in 1995, when was the last time you saw >>>>>>>any micro beat any predecessor of deep blue? When was the last time _your_ >>>>>>>program beat or drew them? Etc... >>>>>>> >>>>>>>Results speak far louder than prejudice... >>>>>> >>>>>>Results can only prove that they were better than their opponents but this is >>>>>>not the question. >>>>>> >>>>>>Uri >>>>> >>>>> >>>>>That is the problem. That was _the_ question. But since the answer is >>>>>clearly known, everybody wants to change the question to something that would >>>>>try to make deep blue look "less" than what it really was. But it was >>>>>unbeatable, considering that it lost to one micro in almost 10 years of >>>>>competition. Nobody _else_ has ever come close to that kind of dominance. >>>>> >>>>>I think it funny that _now_ the question becomes "was their search optimal"? >>>>>Implying that current micros _are_. Which is a joke. Both have enough holes >>>>>to supply a swiss cheese factory for years. The concept of "optimal" is a >>>>>joke. The concept of "results" is the only scientific way to measure the >>>>>programs against each other. The rest is only subjective opinion. >>>> >>>> >>>> >>>>There has been a big smoke fog spread around Deep Blue. >>>> >>>>At the time of the Kasparov match, we have been told that: >>>> >>>>1) it was extremely fast. >>>>2) it had much more knowledge than any other program around. >>>>3) it was using some revolutionnary search techniques. >>>> >>>>Now that we are able to see more clearly what it was, it turns out that: >>>> >>>>1) its superiority came from its speed. >>>>2) the rest was nothing new, and we are still trying to figure out what part was >>>>actually superior to what the best micro programs are doing. >>>> >>>>I don't think that noticing the above is against the interest of science. >>>> >>>> >>>> >>>> Christophe >>> >>>I will be happy to publish the steps to pass muster for human (including GM's) >>>experiments. One quick note is that any "scientific" test to be valid must be >>>reliable/published so that it can be shown to be repeatable by an independant >>>scientist. >>> >>>The DB project was a secret thing, it was very nice " h/w technology", but I do >>>not consider much about DB to be related to science. I am not sure the DB >>>results are reliable, I would expect significantly different results if the >>>Human GM played a few more game (say 100 prep like the 2700 GM had against Rebel >>>recently). I expect DB 1996/97 would get beat by the PC's today in a "true" >>>double blind match/tournament. >> >> >>You were doing OK until that last sentence. Do you _really_ think you could >>take _any_ program from 1997, run it at 200M nodes per second, and that program >>would lose to today's micro programs at 1M nodes per second. I _hope_ you don't >>believe that. And yet we _know_ that DB 97 was certainly stronger than any >>1997 micro, because deep thought was stronger than any micro of its time and >>DB took a quantum leap 100X faster than Deep Thought. > >You assume that 100X faster for deeper blue is the same as 100xfaster for the >top programs of 1997. > >I suspect that it is not the case and the programs of 1997 earn more from time >thanks to their search algorithm when the deep blue team simply believed in >wrong assumptions(for example null move pruning is too dangerous). > >I read that Hsu tried it and did not like the result and I guess that he has >some bug in the implementation. > >Uri He simply didn't like the errors and search inconsistencies it produces. And it produces errors and search inconsistencies no matter _how good_ your implementation. It is a characteristic of null-move search. IE chemo-therapy for cancer is sometimes worse than the actual disease. But the side-effects are known and unavoidable. The drugs attack good and bad cells. Sometimes it is worth the risk. Sometimes not. Null-move is exactly the same.
This page took 0 seconds to execute
Last modified: Thu, 15 Apr 21 08:11:13 -0700
Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.