Author: Chris Whittington
Date: 05:34:43 11/03/97
Go up one level in this thread
On November 02, 1997 at 23:28:09, Robert Hyatt wrote: >On November 02, 1997 at 20:25:39, Joe Beck wrote: > >>What was with Crafty at this event? I really think it is a good >>program, and with running on a DEC Alpha, I really thought it might >>score a good result. Does anyone know what happend? Bob? >> >>Thanks, >>Joe > >I don't have a clue yet. Several possibilities come to mind: This is one of the interesting questions of the event. Firstly you have to note that Crafty didn't do 'badly', it scored 50%. This is an ok result. But I guess it came in below 'expectations'. Expectations based on the alpha 500 technology, and Bob's oft repeated assertions that 64 bit technology would give his 64-bit approach a major advantage. Certainly we could have expected, even with the main tournament result, that the program woudl have shown us its mettle at blitz. My view ? : 1. Its not chess knowledge - I don;t think that at the node rate indications there is a great deal of chess knowledge anyway in these faster programs; there's a limit to what can go in anyway and still maintain the speed. 2. Its not some kind of search bug or problem. It is that Bob releases Crafty source, all the 'amateurs' are able to use its concepts, and maybe even its code in places; plus they have more search ideas of their own. They then don't return the ideas to Bob. So Bob has a program with all the known ideas, all perfectly well executed, but containing no new, secret, bleeding edge ideas. These bleeding edgers are giving themselves a good lead over the Crafty technology. Result is that the bleeding edgers all perform better than Crafty at tournaments. Simple as that. Chris Whittington > >1. the top 3/4 of the field was *very* strong, with (probably) not a >whole lot of difference in strength. This makes the thing a crap shoot >anyway, since who would be favored in a group of 24+ fairly close >players? > >2. I've done lots of tuning in long games, but almost *all* of it is >against humans. Humans that are *very* good at playing anti-computer >chess. My attempts to thwart this style could easily backfire and make >it play oddly. > >3. the extra depth on the alpha could also work against this, letting >it >see even deeper, and get concerned about things that would likely not be >played by the opponent (IE the only program that I know of that plays >even *remotely* like some of the IM/GM players I see on ICC is CSTal... >where it goes for complex rather than normal positions.) Crafty might >well >have been defending against moves that would never have been played. > >4. It is certainly possible that Crafty is simply not playing well when >comparing it to other programs at the WMCCC. I find this hard to >believe, >but Crafty's "environment" (various chess servers) expose it to a type >of >chessplayer that most of the other programs there haven't seen, and >haven't >been tuned for. And that could cause lots of trouble for Crafty. > >I will know more after I see the log files, but the most worrisome thing >I saw was how poorly it did in the blitz event, which it should be >*very* >strong at, based on results against IM/GM players. I may have >introduced >a serious bug for all I know, although I have seen no evidence of one in >a couple of weeks of steady playing on ICC. > >One thing that is a potential problem is that in Jakarta we played on >the >same hardware we test on all the time. In Paris we were about 2.5 times >faster. Using a different compiler. It is certainly possible that we >had >unknown problems caused by this new compiler. Jason ran lots of tests >on the >alpha and compared node counts to the P6, so we "thought" we had covered >this >eventuality. Maybe we overlooked something. > >I believe it is better than it showed. But I also believe that the >programs >at the top are also *very* strong. There's always next year... :)
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.