I was going to diligently read through the many pages, but I must concur that correlation suggests but does not prove cheating.  They would have to monitor the given player for a period and see if their correlation falls.  They simply are not going to do that; it would be hard enough for a game lasting an hour.  For correspondence chess it seems impossible.  Personally I never care if an opponent uses a computer against me, but then I am a lowly pawn in the chess world.  I also like others think that pre-computer analysis is likely to cause an increase in the correlation to computer moves.  In one line I am studying in Avrukh's 1.d4 vol 1, he clearly admits that one awkward and surprising move is computer generated.  If I face it, correspondence or not, I will play it.  Does that make me a cheater?

Bonsai wrote on 09/26/10 at 08:45:11:

When there are equally good moves the ordering of top choices is all the more sensitive to depth and the engine used.

Seth_Xoma wrote on 09/24/10 at 00:04:43:

Nice to read that the discussion hasn't made any progress:

MNb wrote on 09/20/10 at 10:08:23:

Of course I agree. It's sad that some jugglers can't accept this.

Göran wrote on 09/24/10 at 02:03:45:

Exactly.

TJ, That's expressly forbidden on the site. (Except perhaps in unrated games, I don't know.)

Regarding Dembo's name being released, my guess is that was a condition of her free account.

But the grounds for a lawsuit are defamation, which I suppose is also a breach of privacy in some esoteric way.

Having followed this thread for a bit, I would say that something like the following might work reasonably well:

Take all games from a person, but look only at the part of the game where computer help was most likely (i.e. exclude part of the game where there are already games in the database and also once computer evaluations are hugely in favour of one side - even a cheater might play out the K+2P vs. K endgame without assistance)

Then you want to see how often the person matches the computer's (let's say Rybka's) choice particularly in positions where there are a lot of nearly equally good moves. As observed previously it is more plausible that decisive tactics might more often be seen than that someone evaluates positional nuances exactly as Rybka does.

Top 3 matching or similar methods make sense in part because a cheater might let his computer run for hours, which is of course infeasible for someone trying to check lots of games for potential cheating. As a result the shorter time you can invest in checking a position will result firstly in some more or less random fluctuations in the evalutions as well as in the checker having somewhat poorer moves available.

All of this then needs to be put into a proper statistical model, which will then be able to tell you how comparable the percentage of computer moves is compared to

• pre-computer age high-level correspondence chess
• very recent computer age correspondence chess (perhaps not too high-level, so as to look at players who might be very likely to accept a lot of computer suggestions - let's say 2100 to 2300 rating?)

In particular you might use the comparison between those two things to actually figure out what kind of statistical model really fits this data well. Some logistic regression model (fit/no fit) might be an option.

However even if you conclude that the play matches computer play better than past high-level CC, then you would have the problem to decide why that is, either

• the player in question has a different style that more closely matches the computer (I guess less likely, but at least possible)
• the player in question is more precise than past CC players (hopefully accounted for if one e.g. looks at only situations where there were a lot of very similarly evaluated moves, but the person let the computer make the "positional" decisions)
• The player is quite strong and the game sample just happens to be "unlucky" so that there are a lot of matches in moves with the computer.
• The player used computer help.

What you would of course have a hard time detecing is a player that makes all the positional decisions himself only checking whether the computer sees something important...

TalJechin wrote on 09/25/10 at 08:03:42:


This entire batch job was run without removing dB games? I know you'd have to run the dB for games played only up to the date of the games in question, but how "hopeful" can we be that the statistics aren't skewed?

An interesting point is that the player in question made fewer engine-based moves than his opponents!

Volcanor wrote on 09/24/10 at 12:54:16:

Let's say that I'm quite in the middle, maybe in the lower half. I've heard about hypothesis testing, variance and student, but I didn't hear of risk alpha. I heard of Bonferroni inequalities, but not of corrections. I had to take a stats course, unfortunately it was very theoretical and not hands-on (aka it emphasized measure theory proofs).

Zygalski wrote on 09/24/10 at 13:15:55:


Hold your horses;

I don't think you efforts are in vain. Someone with such high top 3/4 values does warrant some extra attention. It on itself is probably not enough to draw definitive conclusions. But hopefully chess.com backs it up with more reliable methods.

You are doing a commendable job as a volunteer.