Author: Graham Laight
Date: 07:00:18 11/21/03
Go up one level in this thread
On November 21, 2003 at 08:50:50, Drexel,Michael wrote: >On November 21, 2003 at 07:28:41, Graham Laight wrote: > >>On November 21, 2003 at 07:09:16, Drexel,Michael wrote: >> >>>On November 21, 2003 at 06:09:20, Graham Laight wrote: >>> >>>>On November 21, 2003 at 05:23:45, Drexel,Michael wrote: >>>> >>>>>On November 21, 2003 at 05:07:32, Graham Laight wrote: >>>>> >>>>>>Everybody knows that as chess computers improve, the proportion of draws in >>>>>>their games becomes higher. >>>>>> >>>>>>The same is true of humans: the following graph suggests that at Elo 3600, all >>>>>>games will be drawn: http://math.bu.edu/people/mg/ratings/Draws.jpg . I also >>>>>>think that a player who plays at Elo 3600 would be unbeatable - no matter how >>>>>>good his opponent was. For a 3600 player, obtaining a draw would, IMO, be almost >>>>>>as easy as it would be for me to obtain a draw against Kasparov with only a king >>>>>>against a king and a knight. In this situation, Kasparov's extra skill and >>>>>>knowledge of the game (and his extra piece) would count for nothing. >>>>>> >>>>>>If what I'm saying is right (and I personally think that it is), then there's a >>>>>>serious problem ahead for the Elo rating system: the system measures chess skill >>>>>>by a player's likelihood of beating another player. However - if the computer >>>>>>that can see 50 ply ahead is unable to beat the machine that can only see 25 ply >>>>>>ahead, then, according to the Elo rating system, it would have the same Elo >>>>>>rating! >>>>>> >>>>>>Is this right, or is this wrong? >>>>> >>>>>It is clearly wrong. A computer that can only see 25 ply ahead will almost >>>>>always lose to a computer that can see 50 ply ahead. >>>>>Until the game isn't solved, a computer could always lose a game theoretical. >>>> >>>>Hmmmmmm. Sometimes, when a visionary looks too far ahead, his reasoning gets >>>>rejected out of hand. For example, if Moore's law continues, then supercomputers >>>>will match the human brain's computing capacity in less than 20 years from now. >>>>This is very easy to demonstrate, if you take 5 minutes to do the arithmetic (I >>>>am happy to do it for you if you want to see it - it's not at all difficult). >>>>Try telling people this, though, and you'll be dismissed out of hand - and >>>>usually by people who should know better! >>>> >>>>Do you think that money will become obsolete? I do - within 200 years. However - >>>>there has never been a time in recorded human history when there was no money - >>>>so people cannot conceive this idea, and reject it out of hand without giving it >>>>sufficient consideration. >>>> >>>>Let me see if we can open your mind to the possibility that a 3600 elo player >>>>will almost always obtain a draw. >>>> >>>>I'm sure you'd agree that in some positions, a bad player could get a draw >>>>against any opponent, just by sticking to a few simple principles. >>> >>>Yes, but the strong player will try to avoid these positions, most often with >>>success. >> >>Assuming his opponent does not also posses a certain level of strength... >> >>>>From here, the leap you have to make is that, with increasing knowlege, the >>>>range of positions in which you can obtain a draw increases. >>> >>>This applies for human chess. Computer Chess today is very different. >>>Search is most important. >>> >>>> >>>>In computer chess, depth of ply search substitutes for (or "provides") >>>>knowledge. >>>> >>>>Therefore, the deeper the ply search, the higher the range of positions in which >>>>the player can obtain a draw. >>> >>>No, computer programs today and in the near future will always get outsearched >>>by a program that can search much deeper. >> >>This is a self-contained truism! :-) >> >>>Your assumption is wrong. >>>The programmer of the program that searches much deeper (without dangerous >>>pruning) should be able to teach his program how to beat those opponents. >>>The deep searcher can explicitly head for lines that seem to be very good for >>>the opponent at lower search depths when in fact they are very bad. >>> >>>> >>>>Also, as the ply search deepens, it becomes increasingly difficult to get the >>>>player into a position in which he cannot achieve a draw. >>>> >>>>If it helps, think of the game of noughts and crosses. This game is known to be >>>>drawn, and, although I'm sure that you haven't sat down and crunched out all the >>>>possible games, I'm sure that you could obtain a draw against ANY player. >>>> >>>>You have to realise that, to a computer that can look 25 ply (or whatever depth >>>>is needed to achieve an elo rating of 3600) ahead, chess will look like a game >>>>of noughts and crosses. >>> >>>25 ply is by far not deep enough to achieve a rating of 3600. I assume you mean >>>3600 SSDF ELO. >>> >>>I am far from believing they can reach 3600 USCF or even 3600 ELO with a 25 ply >>>search. >>> >>>12 full moves and a half move is really not much in chess. >>>I promise you that I will still be able to beat a Computer that can look 25 ply >>>ahead occasionally. >> >>I have always said that I haven't properly calculated how many plies it would >>take to reach 3600 elo. > >Of course not, you simply can't calculate that. >Nobody can ... This is true. The best we can do is to extrapolate forward a "best fit" line (or curve) from various known ply depth / elo rating pairs. Even though I know that the result would only be a loose approximation, I would be interested to see it all the same. -g >You might have a look at this page: > >http://en2.wikipedia.org/wiki/Elo_rating_system > >You can only calculate expected scores between two rated players if the >difference is not higher than 400 points. > >The expected score for Garry Kasparov 2830 against a 3200 chess program would be >10.6%. >Assuming the program is allowed to partcipate in rated human tournaments and >matches. > >IMO we won't see a chess program that scores better than 90% on average against >a human chess player like Garry Kasparov for a loooong time. > >In order to reach 3600 ELO you need opponents (man or machine) which have at >least 3200 ELO and you have to win all games against everything rated below 3200 >ELO. > >Just forget it. You will not see a program rated 3600 ELO in your lifetime :) > >Michael > >I thought I'd read that fritz was looking 18 ply ahead >>against GK - but that may have been 18 moves. I'm sure that there are people >>here in CCC who can extrapolate the search depth that would probably be needed >>to reach 3600 elo. > >> >>btw - the 3600 number is only a visual approximation from this graph: >>http://math.bu.edu/people/mg/ratings/Draws.jpg . I should really do some >>curve-fitting in M$-Excel (which has curve-fitting functionality built in) to >>get a better approximation - but I don't think that 3600 is very far from the >>number that curve-fitting would give me. >> >>-g >> >>>You might have a look at some tablebase endings. >>>In many complicated positions there are not only 5 or 6 pieces but 32 pieces on >>>the board. >>> >>>Michael >>> >>>> >>>>>In some seemingly equal positions there might exist a forced mate in 1000 ply or >>>>>more. >>>> >>>>I said "almost always", not "always". You might get an Elo rating of >>>>3600.000001, rather than 3600, for example, if you win one game in a million. >>>>Not very exciting. >>>> >>>>-g >>>> >>>>>Michael >>>>> >>>>>> >>>>>>If it is right, then the Elo rating system has an upper bound of approximately >>>>>>3600. After this, even "solving" chess by computing out all the possible games >>>>>>will not give you an improvement in play, because the Elo 3600 will still almost >>>>>>always obtain a draw against you. >>>>>> >>>>>>-g
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.