Author: Bruce Moreland
Date: 08:55:33 05/12/98
Go up one level in this thread
On May 12, 1998 at 03:52:25, Danniel Corbit wrote: >On May 12, 1998 at 03:32:01, Bruce Moreland wrote: > >> >>On May 11, 1998 at 20:24:42, Danniel Corbit wrote: >> >>>What if you had each move that a particular GM had ever made analyzed to >>>17 plies or more with a kibits and score saved to a database. Then, if >>>you ever hit any position that they ever played, you will know without >>>needing to calculate the best move. You can even use the time to >>>analyze something else, instead. >> >>A computer that spends a long time analyzing an opening position is like >>an unintelligent person who spends a long time thinking about nuclear >>physics. >> >>Opening preparation would help. >> >>Perhaps a program could find some cooks. >> >>But this isn't a magic bullet. >Not just the opening position. Every position they have ever played. >From first move to end of game. Our goals become much shorter. >Sometimes, we will be on a pre-analyzed position. Sometimes we will be >a single move away from a set of pre-analyzed positions. We can pick >one as a goal, based on its calculated value. This is like a little >tiny mate, and not far away. And if it cannot be achieved, we can >discover this quickly and compute as normal. But this new method plays >much more like a human. A bunch of pretty good moves [when the >technique does not work and we have to work 'as usual'] and occasionally >a brilliant one [when we have a database hit of a very good position or >figure out how to force one]. If we had every move ever made by every >GM analyzed and stored in a database, along with a 17 ply evaluation [or >whatever is feasible], we will very frequently hit upon a 'deja vu' or >at least know how to get there. Pick a position at random from a GM game, and figure out how many times it has come up elsewhere. The answer will almost always be zero. You'll get some endgame positions and a lot of opening positions, but almost no middlegame positions. And even if you get one, you get X minutes to search it, where X is not a lot. Because if you start talking about *every* position from *every* high quality game, you're talking about a lot more than the 500,000 positions that you mentioned before, you're talking about maybe a hundred million. I think it was 6000 computers for a week before, now it's 1.2 million computers for a week. And the result is almost meaningless. I don't think it would improve strength measurably to occupy two billion dollars worth of hardware for a week. At $10 an hour and one minute per install, it would cost you $200,000 just to install the software on these computers. If you want to talk about "siimilar" positions, then you have the same idea that Graham Laight presented in r.g.c.c. a couple of times. The mere act of identifying a similar position is very hard. Once you have something that is capable of understanding that a position is similar to another position, and verifying that a move made in the other position is still a good move here, you might as well just let it play chess and get rid of the pattern matching part, because it couldn't help but be very good just by itself. bruce
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.