Author: Robert Hyatt
Date: 08:33:36 11/10/02
Go up one level in this thread
On November 10, 2002 at 01:09:44, Russell Reagan wrote: >On November 10, 2002 at 00:11:16, Robert Hyatt wrote: > >>I don't see how you could use my PV to predict anything. I have already played >>my move before I kibitz my PV, so how will you use it? > >For example, I play 1. e4 against Crafty. Crafty thinks (no book) and replies >with 1. ... d5, and gives the PV line: > >11 22.98 0.34 1. ... d5 2. exd5 Nf6 3. c4 c6 4. dxc6 > Nxc6 5. Nf3 Bf5 6. d4 e6 > >Now my program takes this, and now my program knows that you're program is going >to ponder 2. exd5 (if that's of any use to me), and I can have a pretty good >idea that next time Crafty is going to play 2. ... Nf6, since that is what was >in the PV line. When I tried this manually, Crafty plays Nf6 like I thought, so >I would have pondered the correct move because of what you kibitzed. In forcing lines that may work. In normal lines, the 3rd move in the PV is not very likely to be played... I'd rather take that chance than have the operator nonsense we have suffered thru in past events. > >Maybe I could go ahead and assume that several of the moves in the PV are going >to be correct, and I'll start my search from 1. ... d5 2. exd5 Nf6 and get an >extra ply or two. So now I'm pondering almost every move you play correctly, and >I've got a few plies of extra search if it turns out that I can assume the first >few moves of the PV will be what you play, which it probably will. > >I'd like to know if this is possible with the current kibitzing requirement. If >it is, I think that's one of the worst rules I've ever heard of. My engine could >probably do something with the score you kibitz too. Maybe search longer if your >engine thinks it's better. Maybe it sees something I don't, so I'll search a >little longer. Or maybe, since I don't like this rule, my engine's PV's will be >1 move long, always, and my scores will be on some absurd scale, and in base 26. Then we use "whisper" rather than "kibitz". The opponent can't see that but all observers can. That ends that loophole quickly... Of course your engine could start a second thread, log on to ICC in a different account, and observe. But that is much harder to do than to operate manually where you can change _anything_ and noone is the wiser, including forcing moves, changing the time per move, etc... > >I don't see what's wrong with kibitzing trivial information, such as time, >nodes, nodes per second, the resulting FEN string, or whatever. For someone to >fake that, they would have to estimate how many nodes they were at, estimate the >time in say, 10 seconds from now, compute the nodes per second real quick, type >all of that in before you hit the "in 10 seconds" time, then kibitz that, and >manually make the move, get the FEN string right, and never make a mistake in 10 >games. If someone is consistently taking more time than usual between kibitzing >and making the move (IE not instantly) that's one tip off. If there's an error, >that's another tip off. If the time doesn't match up, that's another tip off, >because what are the chances that someone is going to get the time right that it >took between the opponent's last move and the time you kibitz, and not make a >mistake for the entire tournament? > >Do this, and provide logs immediately after each game, and I don't see how >kibitzing PV info is any better, at least from a cheating perspective, and now >you can't cheat by pulling off someone's next move from the PV. I think the log requirement is another good point. It would be in addition to "whispering" analysis however. > >This is why I posted earlier about not just posting suggestions, but posting the >reason for them. If the reason is to prove you're automatic, I don't see any >reason for providing PV and score info. To me it's no different than if a >baseball team has a guy out behind the center field wall picking up the signs >from the catcher and relaying them to the batter. If a team did that, they'd >have a significant advantage over the other team. I know this because I've seen >it happen. Knowing that a fastball is on the way makes a big difference. I >imagine that knowing Nf6 is on the way also makes a big difference. > The reason to whisper/kibitz analysis is to let everyone see what the engine is thinking. If a manual operator just makes a move, how can you tell that the engine played that move or not. Or did a commercial program play the move instead of the supposed amateur engine? Providing real-time analysis will increase the entertainment value to observers, and also provide insight into what is going on inside the program, and make it at least much more difficult for the human to take as active a role as he can when playing manually. No solution is perfect. But manual operation has to be the worst option by several light-years... >Require something that only the computer could provide automatically with a good >amount of accuracy. Things like time and nodes, then calculate the nodes per >second. A human would take longer to do these things, and fumble them up >eventually. Print the FEN string, that's something that would also take extra >time for the human. Maybe print a hash of the current position, and provide the >random numbers you used in a log file after the game. That would be almost >impossible for a human to hand produce accurately. Maybe print a list of all >squares with pieces on them. There are a lot of possibilities here, and no PV >info needs to be shared to acheive what you want. > >Russell
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.