Stockfish 9 is OUT!

[syzygy]

According to the online manual, Houdini's "Analysis Contempt" simply forces contempt to be from white's pov

According to "actually firing up the engine and using Analysis Contempt", this is not the case.

Reminds me of my experience in the MultiPV thread, many people kept claiming that Stockfish behaved in a way it didn't. Can't people go and actually check how the engine behaves instead of telling me that it behaves like it doesn't or that some manual says it behaves differently?

I only have the manual. But the behaviour that you describe is... at least hard to believe without seeing it myself. But I can't check it myself, so I am not going to dispute anything here.

In the other thread the confusion was created by the terminology: you said SF does not resolve fail lows, but it does.

[syzygy]

But still according to the manual:

Houdini 6.01
Released September 24, 2017.

Corrections
* Houdini 6 Pro: A correction made to Nalimov code to prevent multi-threaded crashes.

Changes to UCI Options
* A new "Hide Redundant Output" option makes the engine output cleaner by removing some output lines. The option hides information about fail high and fail low situations, and statistics at the end of each search iteration.
* When analyzing positions with the "Analysis Contempt" option checked, the Contempt value will now apply to the White side (instead of to the side that is to move). To use contempt that is optimistic for the Black side you should enter a negative Contempt value.

So you are saying that Houdart wrote up one thing and implemented something that is entirely different (and vastly more complicated to boot)?

[Ovyron]

Yeah, Houdini doesn't behave like the manual claims. I wonder if I discovered a bug, then, since I never have to use negative contempt to analyze the black side, I think this is not a bug, but a feature. But if Robert doesn't know about it (I'm used to this, Vas was frequently surprised by Rybka's unexpected behavior when I beta tested it), it's a glitch.

Did he accidently implement the optimal Analysis Contempt?

you said SF does not resolve fail lows, but it does.

Where? I said it moves to look for alternative moves without resolving the score of the move first, thus, if e4 fails low and Stockfish switched to d4 when looking for alternative moves, e4's fail low will not be resolved. Joerg Oster's patch fixed this. It turns out that if e4 is the main move of the last MultiPV line, then it's faster to not resolve it, so SF and SF+Patch wouldn't resolve the fail low if the move is replaced, but SF+Patch would always resolve the fail low if e4 was in a previous PV of MultiPV and show a score for it.

Unless we're arguing semantics and for you the fact that SF would drop e4 entirely without a score is "resolving" its fail low.

[Ovyron]

So you are saying that Houdart wrote up one thing and implemented something that is entirely different

Um, okay, so the behavior I'm depicting is with a thread devoted to the Tactical setting, I have never tested it with Tactical=0.

So I wonder if Contempt is messed up because the Tactical Thread(s) take over the evaluation in certain circumstances or if the Tactical Thread(s) do not use Contempt, use Contempt differently, or use it in a reversed way. But last time everything worked out after I switched to all caps and bold text, so let me try again:

IN EQUAL POSITIONS THAT WOULD BE 0.00 OTHERWISE, HOUDINI WITH "ANALYSIS CONTEMPT" ON, WITH CONTEMPT=2, IS SHOWING A SCORE OF -0.09, WHICH MEANS 9 CENTIPAWNS ADVANTAGE FOR THE BLACK SIDE, WHICH SHOULDN'T BE POSSIBLE UNLESS HOUDINI BEHAVES DIFFERENTLY FROM ITS MANUAL.

(in certain circumstances, specifically, those where white is actively looking to draw the position)

If this is a bug I hope it's not fixed, it does make sense that the side with the advantage uses Contempt, and the other side only does after equalizing, regardless of colors.

The manual should explain this behavior, but not change it.

[syzygy]

So we can safely assume that Houdini indeed does what the manual says it does. If Houdart changed a specific setting between two versions and documented the change in the manual, it is pretty likely that he tested it carefully.

In any particular position funny things can always happen without contradicting the manual.

[syzygy]

Unless we're arguing semantics and for you the fact that SF would drop e4 entirely without a score is "resolving" its fail low.

Well, that's the point: two sides were talking about different things.

I'm not blaming you for the confusion (and I think you did give enough information to resolve any possible ambiguity... at least I think I understood what you were trying to say).

[Ovyron]

So we can safely assume that Houdini indeed does what the manual says it does.

Nope, in no circumstance Houdini would apply Contempt to the black side of a position for analysis if the Contempt setting is positive according to the manual.

Yet, it does.

You can't just assume that the engine does what the manual says when you use the engine and it does the opposite of what the manual says.

I wonder if this is just an instance where the manual is outdated.

[corres]

I am afraid a debate on contempt is a worthless thing because every kind of contempt decrease the power of a well-tuned engine. Against an engine having the same power of Stockfish, Stockfish 9 with contempt = 20 is about 5 Elo weaker then with contempt = 0. The real difference depends on the power of enemy and the type of opening.

[Michel]

I am afraid a debate on contempt is a worthless thing because every kind of contempt decrease the power of a well-tuned engine. Against an engine having the same power of Stockfish, Stockfish 9 with contempt = 20 is about 5 Elo weaker then with contempt = 0. The real difference depends on the power of enemy and the type of opening.

Can you back that up? contempt=20 was shown to be neutral in self play.

[syzygy]

So we can safely assume that Houdini indeed does what the manual says it does.

Nope, in no circumstance Houdini would apply Contempt to the black side of a position for analysis if the Contempt setting is positive according to the manual.

Yet, it does.

You are just misinterpreting a single position.

As I said, in a single position anything can happen. A search tree can change completely if an evaluation parameter (like contempt) is changed by just 1cp. You're not seeing positive contempt for black, you're just seeing the engine finding another line.