Author: Robert Hyatt
Date: 21:07:53 09/17/98
Go up one level in this thread
On September 17, 1998 at 18:46:02, Serge Desmarais wrote: >On September 17, 1998 at 09:05:04, Robert Hyatt wrote: > >>On September 17, 1998 at 02:05:43, Jeff Anderson wrote: >> >>>I suppose I can understand where you are coming from as the developer. I would >>>like to point out that Mr. Moreland allows Ferret to play unrated games against >>>anyone, and he is having no problem finding strong opponenets to play for his >>>creation. In fact the last 20 games in its history are against above 2600 >>>players. >>>Jeff >>> >>> >> >> >> >>I choose to not play unrated, to try to make the games "serious". Since most >>there treasure rating points, rated games tend to be more serious games. Also, >>I have seen "scams" where someone "fishes" by playing unrated games until they >>find a way to force a program to follow a bad book line, then they will play a >>rated game and win. Also there is a problem with playing rated as white, then >>an unrated when you get black, then rated with white, etc... >> >> > >Crafty would still learn after an unrated game? > > >Serge Desmarais > > it would depend. IE the most common way this used to happen was a guest would play it many games to fish around for a bad opening, then the guy would log on as a normal user and play that line. I let the rating of the opponent factor into the learning curve, and a guest would always get "0"... But that's not the problem... play it unrated, find a good position, then blunder. That defeats the learning by actually encouraging it to follow that line again (I handle this by reducing the learned result significantly, but as you can see, it won't learn anything bad if you pop it out of book, decide you like the position, then intentionally blunder so that it won't realize that the position was bad... Humans are clever animals. Some studied my "mercilous attack" code and found novel ways to take advantage of it, in fact.. A never-ending battle of wits, generally, but *only* when you operate "automated" on the servers. Manual operation eliminates the problems... But that's part of the challenge... get it right and it makes you feel like you've accomplished something important... > > >> >> >> >>>On September 16, 1998 at 22:16:18, Robert Hyatt wrote: >>> >>>>On September 16, 1998 at 00:49:17, Jeff Anderson wrote: >>>> >>>>>On September 14, 1998 at 22:13:42, Robert Hyatt wrote: >>>>> >>>>>>the most noticable affect is that "crafty" used to play a hundred games a >>>>>>day. Now it sometimes plays 10 or less, because there are so many crafty's >>>>>>on ICC. >>>>> >>>>>Well your formula is very strict! You eliminate 95% of all possible challengers >>>>> with insisting that there rating be above about 25001 >>>>> >>>>>Secondly you will only accepted challenges that propose rated games. >>>>> >>>> >>>> >>>>I do this for a reason. (1) the chances of a 2000 player beating Crafty on >>>>ICC are almost nil. It will happen now and then, but not often. I don't learn >>>>a lot from crushing 1500-2000 players, because the wins are tactical smashes >>>>that reveal nothing about problems I have. (2) I learn more from losing than >>>>from winning. Playing IM/GM players greatly increases the chances of losing, >>>>which is what I am looking for. (3) There are far more 1500 players than 2500 >>>>players on ICC. If I let 1500 players in, they will totally lock out the 2500 >>>>players. and finally (4) I have been specifically asked by some strong players >>>>there to keep my formula restrictive so that they can play when they want. >>>> >>>>If crafty operators want to ban together and work out mutually-exclusive >>>>formulas (IE I play players over 2500, someone else takes 2100-2499, etc.) >>>>then that would work. At present we simply have so many crafty's running that >>>>many have lots of open time because lower-rated players don't like to get >>>>drubbed tactically... >>>> >>>> >>>> >>>> >>>>>Now I think there is something to the suggested idea of ICC computers allowing >>>>>takebacks. For example you might lower the formula to allow those rated above >>>>>2000 play, and allow 2-5 takebacks a game. If you are really concerned about >>>>>seeing games where Crafty has losses against humans, you might consider this >>>>>approach. Also you could simply ask people to send you log files games where >>>>>Crafty lost against humans. >>>> >>>> >>>>this doesn't help as much, because those games come from the released version, >>>>while I am trying to evaluate changes for the *next* version to be released... >>>> >>>> >>>> >>>>> >>>>>This would be the sensible approach if your number one concern was having >>>>>valuable information for improving Crafty. But I'm sure it is not like the >>>>>adreniline rush you must get when you watch your program beat a GM. You say >>>>>your strict rating restrictions are in the name of science! You say you would >>>>>like to see games where Crafty has lost so you can improve Crafty, and very >>>>>strong players beat Crafty more frequently. Well this is non-sense, because two >>>>>perfectly reasonable alternatives have been offered, two that would give you >>>>>excellent data, and one that gives Joe Patzer a chance to play Crafty....and >>>>>win! >>>>> >>>> >>>> >>>>the above is *not* "nonsense". As I said, games from "other" crafty's are >>>>not particularly interesting for me. Folks are trying different extension >>>>options (tunable from the crafty.rc file), others are trying different book >>>>variations with wild gambits and stuff they are particularly interested in, >>>>even if the openings don't fit crafty's "style" very well. I don't have time >>>>to wander thru a hundred log files a day only to discover that 95 of them were >>>>lost due to book opening choices. >>>> >>>>As far as takebacks, that is complex. Chess is a game played from start to >>>>finish. The search is written with that in mind, with thinking on the >>>>opponent's time and so forth. Time controls. All of that is designed around >>>>the game of chess as defined in the rules. Takebacks add more to the code, >>>>and introduce code that is not needed in normal games, and this code could well >>>>be something that hurts something without it being known. So while takebacks >>>>would be interesting, it isn't chess. We can't do it at WMCCC events, or in >>>>real rated human events, so adding this to the engine is basically nothing for >>>>something, sort of. >>>> >>>>A year ago, Crafty was playing a hundred games a day, 90% of them against GM >>>>players, the remainder against IM players. Today it is playing 20-30 GM games >>>>a day, and 10-20 IM games. I'd still rather play a strong player that is going >>>>to push it in ways that a weaker player, assisted by takebacks, won't. >>>> >>>>As far as the "adrenaline rush" goes, that went away several years ago. I don't >>>>lose enough games against GM players to notice much any more, so the wins are no >>>>longer noteworthy. In fact, even the people watching have developed the same >>>>"expectations" and the occasional GM win produces far more chatter than the >>>>regular GM losses... >>>> >>>> >>>> >>>>> >>>>>Jeff
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.