Author: Drexel,Michael
Date: 11:11:57 07/22/03
Go up one level in this thread
On July 22, 2003 at 07:13:06, Uri Blass wrote: >On July 22, 2003 at 05:30:19, Drexel,Michael wrote: > >>On July 22, 2003 at 02:27:31, Dann Corbit wrote: >> >>>On July 22, 2003 at 02:22:44, Omid David Tabibi wrote: >>> >>>>On July 22, 2003 at 00:04:58, Dann Corbit wrote: >>>> >>>>>On July 21, 2003 at 23:50:38, Omid David Tabibi wrote: >>>>> >>>>>>On July 21, 2003 at 23:47:12, Derek Paquette wrote: >>>>>> >>>>>>>Hey, while I agree it would cost A LOT >>>>>>>i have to disagree with you on the point that it "isn't" possible. >>>>>>>While all wins might seem crazy, >>>>>>>If 15 million dollars was invested into it, >>>>>>>a machine that dwarfed deep blue, >>>>>>>with much larger openning books, (only 4000 lines for deep blue, and people >>>>>>>think DJ8 that has hundreds of thousands was just as strong? dont' get that, but >>>>>>>neway) >>>>>>> >>>>>>> >>>>>>>if its looking at a TRUE 20 ply ahead, I don't see how it could lose with that >>>>>>>much invested. >>>>>> >>>>>>It will lose very rarely, but will still draw quite frequently... >>>>> >>>>>Let's suppose that though a titanic bucket of money, the speed of Deep Blue were >>>>>increased thusly: >>>>> >>>>>Per-CPU speedup 100x (already achieved 10x, I think -- 100x should be doable) >>>>>CPU number increase 10x (480 to 4,800) >>>>> >>>>>so we have 1000x power increase. That would correspond to about 10 doublings in >>>>>power. >>>>>That would be 500 ELO. >>>> >>>>How did you calculate this? >>>> >>>>Assuming a branching factor of 3, log(1000)/log(3) is about 6 plies. 6 >>>>additional plies will increase the strength significantly, but not more than 200 >>>>Elo in my opinion (diminishing returns, etc). >>> >>>50 ELO per doubling of speed is the lower end of the standard estimates (50-70 >>>ELO is the heuristic figure). >> >>This estimate is for comp-comp. Contrary to popular believe the increase is less >>than 15 ELO against humans. There is of course the psychological factor. >>Chess players with no experience in Computer Chess might be scared playing the >>Super Computer and play much weaker (passive) as they would play against a mere >>PC. >> >>> >>>As far as diminishing returns go, it has been demonstrated by one study and >>>refuted by another. >>> >>>>Computers are already stronger than humans in tactics, so 6 additional plies >>>>will just be an overkill. On the other hand, those 6 additional plies will >>>>hardly make up for the program's inferior positional understanding in comparison >>>>to humans. >>> >>>With 6 more plies, tactics become strategy. If a program that can consider 12 >>>plies (6 full moves) can now consider 18 plies, that is 9 fullmoves ahead. >> >>9 fullmoves? This is almost nothing in a closed position. It wont help you a >>thing to find better moves in the Kings Indian defence for example. > >I do not think that you are tight here: >1)finding better moves in a lost position is also important because the opponent >with the advantage may fail to win the game. What do you mean with better moves? For a computer is a better move in a lost position always the move with the best (least worse) score. This move however might lead to a position the human is happy to play because win is easy and only a matter of time. Very often it is much better to choose a move which loses immediately but only due to complicated tactics (or due to a move which seems to be not logical at first sight) so that a human might miss the win. > >2)programs may avoid part of the blunders that they do if they search deeper. >They do not need to see that the blunder is losing and it is enough if they see >that another move is better. >If they are too optimistic they may reject the wrong move because the opponent >can force a draw. Of course it is always better to search deeper but it in some positions it is by far not deep enough. Michael > >Uri
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.