Author: Robert Hyatt
Date: 12:13:31 09/20/99
Go up one level in this thread
On September 20, 1999 at 12:15:43, Christophe Theron wrote: >On September 19, 1999 at 17:17:27, Robert Hyatt wrote: > >>On September 19, 1999 at 15:04:53, Christophe Theron wrote: > >>>I think it's a very bad idea to target for a given speed. Sorry, but I see no >>>reason to do this. When I make a change in Tiger I make various tests on >>>different computers to be sure that nothing is broken. These tests include blitz >>>games against Genius5 on my 386sx20 and longer games on a 300MHz computer. >>> >>>I think that making sure that the strength is unaffected at very different >>>speeds is a way to make sure that you are going into the right direction. At >>>least I think it works for me... >>> >>> >>> Christophe >> >> >>I personally believe that doing this is impossible. IE I can't imagine a >>program that plays equally well if it searches to 6 plies or to 10 plies. I >>had this problem for years as my Cray Blitz development was mainly done on a >>VAX, and then we would run on the Cray for tournaments. And on occasion, it >>was very obvious that things that were helping at 6 plies were killing us at >>10... > >This kind of problem can be solved. Generally I use the same algorithms whatever >the speed is, but in some cases I adapt. IE there are things I do in the first >iterations I don't do anymore at deeper depths. > > here is an example: In 1985 (Summer) I was getting ready to move to Birmingham to work on my Ph.D. Bert and I were also getting ready for the 1985 ACM computer chess tournament in Denver late that fall. We played a _bunch_ of games, using cray blitz on a vax, vs his super constellation and a par excellence (I think). the vax was winning most games, but we did lose some endgames where the opponent infiltrated because our pawns were scattered around and we had lots of holes. We looked at this and added a simple scoring term to penalize holes since once made, they can never be repaired. We stopped losing the endgames, and were happy. When we moved to the Cray, and went from 4-5 plies to 9-10 plies at the ACM event, we got into serious trouble and lost two games that year. And lost the second round of the WCCC the next year (1986) as well. Turns out that a 10 ply search could usually see how to exploit those holes, while a 4-5 ply search couldn't. So it would _never_ advance pawns on the 10 ply search depth, because it got a penalty for the hole, and then it also saw something negative happening to exploit the hole. Often you advance a pawn to get something _and_ give up something. It was giving up so much (it thought) that it wouldn't advance for any reason. We took those four lines of code out and Cray Blitz started playing like Cray Blitz again and won the 1986 event. I was very careful after that point to be sure that such things didn't happen again. I have seen exactly the same thing happen with Crafty. Prior to the Paris WMCCC event, Crafty was sitting on the highest ratings it had ever had on ICC. But it turns out that king safety was turned up too high. Rather than doing 9-10 plies in blitz, it was doing 12-14 plies in the WMCCC. And it was so very paranoid about king safety, that it would do _anything_ to fend off threats that really didn't exist. I don't see how to avoid the depth dependency when the eval gets _very_ large (as mine is getting and as Cray Blitz's was...) > >>just my opinion, of course. But you certainly can't get away with null move >>R=2 on a 386. > >That's right. Null move is the kind of algorithm that reduces the branching >factor, and needs at least 5 or 6 plies of depth to begin to be really >effective. > >That's why I have other selective algorithms that are designed to work even at >depth 2-4 (and they keep working very well at deeper depths BTW). > > > >> It will be so blind to king-side attacks that a good expert will >>eat it alive. > >OK, I would not play a grandmaster on a 386 anyway... > >The problem is to be equal or better than the best software running on the same >computer. I don't mean I achieved this, but I think that trying to do it helps a >lot to go into the right direction. > > > > Christophe
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.