Houdini analysis in Fritz 13--strange.

[ivoryknight]

The position:
r1b2rk1/1p2qppp/p1n1p3/2npP3/5P2/2N2N2/PPPQ2PP/2KR1B1R w - - 0 12

The mystery:


Full-sized image here:
http://westportchessclub.org/computer-c ... alysis.png

Houdini has come up with 3 different answers for the same problem. It's not a matter of plies, like it looks like at quick glance. Look deeper and read on.

So I analyzed that position a few days ago (Nov. 11th) to 30 plies, and Houdini said the "answer" was 12.Kb1. (See Exhibit D in the above image: maroon text.) Later that evening, I came back to that position, because I wanted to analyze the main book line: 12.Bd3. So I set it to that position, and the phone rang, so it analyzed to a depth of 26 while I talked (See Exhibit D again: blue text.) Houdini then said the "answer" was 12.Nd4. I ignored it and moved on to 12.Bd3. Then today I came back to that position and thought, let's analyze it again because I don't ever remember Houdini thinking the "answer" was 12.Nd4 the whole time it was originally analyzing and came up with 12.Kb1.

So, now it's analyzed to a further depth--28 plies--than the blue text (the 2nd time I analyzed the position) and it thinks--and always has thought during this analysis--that the "answer is 12.Qe3.

How can Houdini come up with three different answers to this problem? It's not because of the differences in plies. Today's answer, 12.Qe3, has been its answer since it started thinking. And it has now thought further than the blue text analysis by 2 plies. As I've typed this, the analysis has gone 29 plies deep, and Houdini still thinks the answer is the same as it has always thought this time: 12.Qe3. It hasn't changed its mind once.

Which brings me to my point. I've noticed that sometimes I feel like it picks an answer right away and sticks with it. It has done this same thing it's doing today a number of times. I can analyze a position, and I think it's "stuck", so I will back out or close the program and re-open it, and it will come up with a different answer, and start to actively "think", I mean show that it's considering different answers (you can watch them as it changes its mind), not "sticking" with one the whole time of analysis. I've seen it often enough, so I am mentioning it here. Has anyone else experienced this?

[Karol Majewski]

Hmm... You are using more than 1 thread, so you'll never get the same engine output twice. Engines on multiple procesors are non-deterministic. That's normal behavior.

[ivoryknight]

Hmm... You are using more than 1 thread, so you'll never get the same engine output twice. Engines on multiple procesors are non-deterministic. That's normal behavior.

They do not come up with the same answer each time? That is normal?

[ivoryknight]

I do not believe it is a Houdini problem. I have noticed it with Rybka, too. But not as much, probably because I only use it for more closed/positional situations as a secondary answer to Houdini.

[Karol Majewski]

Yes. To get the same output every time you have to reduce number of threads to 1. That's the nature of MP engines

[ivoryknight]

Yes. To get the same output every time you have to reduce number of threads to 1. That's the nature of MP engines

I believe you, but this blows me away. I just thought MP engines came up with the same answer as 1-core boxes, but faster. So for all those people utilizing MPs and turning in analysis for Fritz 13, it may be wrong?

[Karol Majewski]

Yes. To get the same output every time you have to reduce number of threads to 1. That's the nature of MP engines

I believe you, but this blows me away. I just thought MP engines came up with the same answer as 1-core boxes, but faster. So for all those people utilizing MPs and turning in analysis for Fritz 13, it may be wrong?

MP engines has their pros and cons. The most important pro is speed. MP engine will be always stronger than SP - because of speed. But if there are 2 or more equal moves - MP engine won't repeat it's choice every time.

[bob]

Yes. To get the same output every time you have to reduce number of threads to 1. That's the nature of MP engines

I believe you, but this blows me away. I just thought MP engines came up with the same answer as 1-core boxes, but faster. So for all those people utilizing MPs and turning in analysis for Fritz 13, it may be wrong?

It is nondeterministically different. It won't be the same each time. It will be far better than the single-cpu version, however...

[Houdini]

Houdini has come up with 3 different answers for the same problem. It's not a matter of plies, like it looks like at quick glance. Look deeper and read on.

If you run a multi-PV analysis you will see that the 3 moves are nearly equally good. Qe3 and Kb1 will transpose into the same main line, and Nd4 is only 0.05 point weaker.
As any of the 3 moves is equally good, the exact move choice will be more or less random. But once the choice has been made, Houdini will stick to the move because the other two possibilities or not better than the first choice.

Robert

[ivoryknight]

Houdini has come up with 3 different answers for the same problem. It's not a matter of plies, like it looks like at quick glance. Look deeper and read on.

If you run a multi-PV analysis you will see that the 3 moves are nearly equally good. Qe3 and Kb1 will transpose into the same main line, and Nd4 is only 0.05 point weaker.
As any of the 3 moves is equally good, the exact move choice will be more or less random. But once the choice has been made, Houdini will stick to the move because the other two possibilities or not better than the first choice.

Robert

Thank you for the clarification, Robert! Again, as I stated above, I never thought this was solely a Houdini "problem", and now that others have explained, I see it's a multi-processor "feature".