Author: Robert Hyatt
Date: 14:03:41 10/23/99
Go up one level in this thread
On October 23, 1999 at 15:56:14, Michel Langeveld wrote: >On October 23, 1999 at 15:27:15, Robert Hyatt wrote: > >>On October 23, 1999 at 15:04:44, Michel Langeveld wrote: >> >>>On October 23, 1999 at 12:52:11, Robert Hyatt wrote: >>> >>>>On October 23, 1999 at 11:14:07, Michel Langeveld wrote: >>>> >>>>>Thanks in advance. >>>>> >>>>>I want to try it myself ;-)!!! >>>>> >>>>>Michel Langeveld >>>> >>>> >>>>It is too early to say. Compared to the opening book I use, less than 100K >>>>of the positions have CAP scores. Which means that many positions have no >>>>data at all. >>> >>>I have no idea how much positions your book contains but it sounds terrible. >>>Maybe we have to add an project for CAP which is called "Crafty Book". >>>All positions of the Crafty book has to be examined. >>> >>>CAP is a great project but unfortunately it's not going as fast as I would like >>>to see it going. Does someone know some very fast machines which are free for >>>this project?? >>> >>>>Another problem is that the CAP data is of the form >>>> >>>> <position> <score> <best move> >>>> >>> >>>In the CAP data is also included: >>><ply> >>><number of nodes examined> >>><analysis variation> >>> >>>which might be very interesting to use these in Crafty. >>> >>>I have also no idea if all positions of a certain variation are also in the CAP >>>base. For example if 1.e4 h6 is a found in CAP as the best line that also that >>>also 1.e4 and 1.e4 h6 are examined. And It might be that when doing a minimax >>>again over all the data that 1.e4 e5 is found the very best. >> >>I used to minimax eval/material from tips back to root. I started that in >>Cray Blitz, but it has obvious flaws. I even did searches from all the 'tip' >>positions, but that didn't catch blunders in the middle of the line. I also >>tried searching _every_ position so I could catch blunders, but searches deep >>enough weren't possible... due to the time required. > >I hope Dann will add the Crafty book as project to CAP so that every book >position is analysed at least for 100M nodes. It might be good to collect all >learn information on a website if the full-cap-analysed-book so that errors can >be CAP-ed again. > >>>>I can use that... but what I _really_ need to do, and I don't have the code >>>>to do yet, is to do a simple minimax of the CAP scores back up thru the book. >>> >>>Right this is what I described above too! >>> >>>>IE if the best move in position P is Rh1 with a score of +1.3, then I need to >>>>know that any move that leads to position P (with no cap score in the file) >>>>ought to have a score of -1.3. I don't do this yet, which hurts. I think it >>>>will be even more useful once I have time to write that code. >>> >>>Do you have an idea of how many positions are then filled with CAP data in the >>>book? >> >>The book is about 13.5 megs. at 20 bytes per book position, that means about >>700K positions, where the cap import found about 100K 'matches'. So about 1 >>of every 7 possible moves has a CAP score. > >So if we go with the same speed it will take a year to analyse all >crafty-book-position with CAP. (The CAP-data is now about 850K positions, and we >talk about a book of 700K). > >>>Are you goining to extend the book with an extra flag which indicates a score is >>>is a real CAP-data-score or a score is a "moved-up" one? (By the way I remember >>>BOOKUP is has facilities to do this) >>> >>>Is having the book minimax with your algorithm a think which HAS to be done >>>before releasing Crafty 17.0? >>>In other words what has to be done before 17.0 can be released? >> >>this won't be a 17.0 feature (minimaxing). Main thing for 17.0 is to test the >>new eval code and the eval tuning. It is playing pretty well, but there are >>still quirks as the new candidate/majority code can influence things in odd >>ways and it takes some watching to spot these and fix things. > >Ok, I see. If you need any help let us know! > >>> >>>CAP data are positions which has roughly 100.000.000 of nodes examined. Do you >>>know on what hardware Crafty reach such numbers in a normal game? And is it >>>possible to add those positions in a seperate file so than can immediatly added >>>to CAP? >> >>100M nodes? On my quad xeon it will search deeper than that in a real game, >>as it is searching close to 1M nodes per sec on the quad already. That would >>take about 100 seconds to do... :) > >Cheeze. So I normal game (which is 3min/move) on your quad will produce higher >quality of data than CAP. I don't know If I had to cry or laugh about this. > >I also calculate that all CAP data until now will take 590 days on your single >computer. Is this your developing machine? Or is your testing machine? > It is both. I both test/debug on this machine while crafty is actually running on ICC. I try to do things using 'nice 20' if I know it is playing... >>>Will crafty 17.0 also scan in the CAP-data in a normal minimax when it has left >>>to book is is the book the only place wher CAP data is used at the moment? >> >>only in the book. I don't 'trust' the scores that much, since different >>versions of different program produce different scores. Choosing book moves >>is one thing, but I want "the crafty search" to be responsible once the book >>has ended. > >Ok. seems fairly. Maybe that's why only 100K positions matched your book because >it has also plenty of middlegame, endgame end EPD-testset positions. > >By the way... is it difficult to write a program that converts an book.bin to a >ascii file with all kind of EPD positions. Can you give me some hints in doing >this. Then I will try to write it. It might be of some help in future.... > since I store hashed positions, you have to 'search' just like I do to find opening moves. Make each move.. if a resulting hash key matches the book, that move is a book move. So you would basically 'search' but only follow moves that lead to known book positions, outputting FEN (savepos does this) after each successful hit. >Kind regards, > >Michel Langeveld
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.