Author: Robert Hyatt
Date: 18:54:35 06/27/01
Go up one level in this thread
On June 27, 2001 at 12:15:58, Uri Blass wrote: >On June 27, 2001 at 10:47:41, Robert Hyatt wrote: > >>On June 26, 2001 at 23:35:15, Uri Blass wrote: >> >>>On June 26, 2001 at 23:06:26, Robert Hyatt wrote: >>> >>>>On June 26, 2001 at 14:29:25, Uri Blass wrote: >>>> >>>>> >>>>>It also cannot be repeated against a chess program if it remembers the game and >>>>>has learning by position or if it is not deterministic. >>>>> >>>>>Uri >>>> >>>> >>>>This doesn't work quite like you think. For lots of well-known reasons. The >>>>most important is that if you go out of book very early, and don't see anything >>>>bad happening for a long while, it will take a _long_ while to propogate those >>>>scores back up the search tree to avoid a bad early move that doesn't lose for >>>>(say) 20 more moves. >>> >>>You can learn to remember scores of 0.1 or 0.2 pawns lower than the scores of >>>the game after losing so if the program has a logical alternative it is going to >>>choose it >> >> >>As I said, it doesn't work like that. Suppose that the first move looks like >>0.00 when you search it. And all the others look like -.3 at the same depth. >>Your -.2 score will _still_ be the best, since you can't see deep enough to >>discover that one of the -.3 moves will become +.3 in 4 more plies. >> >>That is a "local maximum" and you can't work your way around it >> >> >> >>> >>>For example if it lost by 1.e4 c5 2.Na3 Nc6 when the score was 0.24 for itself >>>After 2...Nc6 then it may remember only 0.04 for itself based on the fact that >>>it lost the game and it may help it to prefer 2...e6 that gives it 0.17 pawns >>>adavnatge for black. >>> >>>Uri >> >>But suppose all the others are < .04? you are stuck there. > >Only if there is a forced line from computer point of view after leaving book >and even in this case the program is not going to repeat the line again and >again because after 2 games with the same line it is going to prefer lines that >it played only once(it also prefer lines that it did not play but I assume that >it had no choce after enough games and it must repeat something that was >played). > >I believe that it is possible to avoid the problem of forced lines after book by >a book that leads to quiet positions when many positions have similiar score. > >people need to play a lot of game to get repetition of the same line. > >After 1.e4 c5 2.Na3 many moves have similiar score so at least this opening >cannot work again and again(c5 is not going to be more common then moves like >c6,e5 or d6 and after Na3 black has also many possibilities so you need many >games only to get repetition of the same first 2 moves. > >Uri The problem is that after e4 c5 Na3, black is probably out of book. It will take _forever_ for it to solve the problems here. The only answer will be to mark c5 as unplayable then you get e4 e5 Na3 and you are back to the same problem. Sooner or later you learn that _all_ your book moves are bad, and then what? Wait for a few thousand games for the scores from move 30 or beyond to propagate back up the tree to help? It simply won't happen. Position learning is a small crutch to help with a specific problem. It does't cure all the problems at all. If it did, developing a GM program would be easy...
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.