this should show just how strong Rybka is - it performs well in handicap matches even though that should be its weak spot.

Actually, depending on how flexible Rybka (the program) is, its programmers could have easily tweaked its evaluation function to make it prefer complicated positions.

This is true, and they generally do so in such matches.

Whoops, tense slip-up. At first I had it worded, "the programmers could easily tweak", but then I decided to refer to some specific match ("the programmers could have easily tweaked"), but I left it as "the programmers could have easily tweak".

Anal retentive? Yes.

I was going to mention this but decided to dumb it down a little.

It's already tweakable, they have released several different binaries with different styles. The problem is that every tweak makes the program objectively weaker, so the point is still valid -- it's excellent in handicap matches even though it has to be crippled somewhat to play at it's most effective in them.

I'm skeptical as to how much weaker such tweaking really makes it. It would probably be enough to have it choose the move which leads to the "more complicated" position if the difference between it and the #1 move is less than .1 pawn or so. Obviously it's not going to make significantly worse moves just to complicate.

Alternatively, you could have it continue to play #1 moves but modify its evaluation function to slightly overvalue (the weakening of) the opponent's king safety, and perhaps to slightly overvalue the net difference between its total mobility and its opponent's total mobility, which will tend to be larger with more pieces on the board, in open positions, and when the opponent has less space. All good things in an odds match.

I think option 2 is closer to what they do.

but the bolded bit is vital. I haven't seen the Kasparov-Chapman games in ages but I'm pretty sure Kasparov made a lot of moves which were significantly worse (by a computer's standards) just to complicate, but he can judge when to do that so much better than the computer.

I can agree that the problem of when to complicate and to what extent is a much harder problem from an AI standpoint than finding the (probable) best move in a chess position. I just don't think that it's such an obstacle that the fact that Rybka's done well in odds matches speaks volumes about its playing strength. I will admit that it speaks well of it, though.

By what standard can one call a particular move objectively worse? If we know that a more complicated position increases our EV, then if we weigh the benefit of the complication against any other losses that a particular complicating move might bring, then isn't that all that counts?

Is it possible, truly, to separate the value of a move from the setting of the game? That is: the nature of the opponent, the time controls, and other factors?

I thought it was clear that the standard is Rybka's own (normal) evaluation function. Obviously there are many positions where there's no "objectively best" move, because chess is so complex. Even if 90% of GMs think that a particular move is best in a particular position and 10% like a different move, the minority could in theory be right, but we'll never really know. Ultimately only moves which lead to a very clearly won or lost position can be called objective best or worst.

So what we're talking about is how Rybka might adjust (or, rather, be adjusted) for the fact that it not only needs to play good moves, it needs to keep the position complicated. RoundTower suggested that Kasparov is, in all likelihood, much better at pulling off such a balancing act than Rybka could ever be. I agreed, because the task of balancing the need to play good moves against the need to keep the position complicated is a task that is itself much more difficult than chess from an AI perspective.