CLOP for Noisy Black-Box Parameter Optimization

[zamar]

Very impressive achievement Remi. I will try your new tool at some point in the near future!

[Rémi Coulom]

Thanks Joona.

I added some screenshots of the program to the web page, and a description:

These are some screenshots of an old version in action. You can also run the program from the command line, which is more convenient for use on a remote cluster. The program can deal with chess outcomes (win/draw/loss), and integer parameters. The program is written in C++ with Qt, so it can be compiled and run on Windows, Linux, and MacOS.

http://remi.coulom.free.fr/CLOP/

Rémi

[Zach Wegner]

Very impressive achievement Remi. I will try your new tool at some point in the near future!

+1

I am very glad that you decided to re-release your source code, even though you are a "commercial" now. Thanks!

I will try and understand your paper. Do you know how well your ideas could be applied to non-game-playing applications (i.e., a floating point objective function)?

[mcostalba]

Thanks Joona.

I added some screenshots of the program to the web page, and a description:

These are some screenshots of an old version in action. You can also run the program from the command line, which is more convenient for use on a remote cluster. The program can deal with chess outcomes (win/draw/loss), and integer parameters. The program is written in C++ with Qt, so it can be compiled and run on Windows, Linux, and MacOS.

http://remi.coulom.free.fr/CLOP/

Rémi

As you perhaps know Stockfish has been tuned in an automatic way by means of an (not disclosed) algorithm written by Joona, so I am very happy that he finds your work useful because he is the expert of tuning in the SF team...and so I hope this will give him good hints to further tune the tuner

[Rémi Coulom]

Do you know how well your ideas could be applied to non-game-playing applications (i.e., a floating point objective function)?

I did not try, but I expect the basic idea of CLOP would work well in many situations, even the completely noiseless case.

I am really enthusiastic about this algorithm, because it is extremely simple, and very universal. The problem of optimizing a function from noisy (or noiseless) observations has really been researched a lot for more than 50 years. It is really difficult to contribute anything significant to this field. I am looking forward to the feedback of optimization specialists. Time will tell if CLOP makes an impact.

Rémi

[Rein Halbersma]

Hi Rémi,

Can CLOP also be applied to LOS as an objective function? Suppose that a tournament has a very skewed prize money distribution (e.g. in poker). In such cases, I can imagine that programs slightly below the absolute top might want to maximize a different mean-variance combination than the best programs. E.g. optimize their chance of winning a tournament, rather than their ELO.

Rein

[Rémi Coulom]

Hi Rémi,

Can CLOP also be applied to LOS as an objective function? Suppose that a tournament has a very skewed prize money distribution (e.g. in poker). In such cases, I can imagine that programs slightly below the absolute top might want to maximize a different mean-variance combination than the best programs. E.g. optimize their chance of winning a tournament, rather than their ELO.

Rein

I don't really understand your question.

The objective function can be the expected value of any random variable that depends on parameters. So, if, instead of playing one game and getting the result, you play a tournament and observe LOS over a specific opponent, you can use CLOP to optimize it. But that would be a strange way to use CLOP. If you wish to optimize a program against a set of opponents instead of just one opponent, you can use the "Replications" option of CLOP to play a game against each opponent, and then CLOP will maximize the average winning rate against all these opponents.

Regarding your example, if I understand correctly, you mean letting the program be more agressive when it needs to win, and safer when a draw is OK. For that, you could tune your evaluation with CLOP, considering that a draw is a loss (resp. a win) to make it play agressively (resp. defensively). Make sure your evaluation function is asymmetric, then. It may be more efficient than using just contempt.

Rémi

[Andres Valverde]

Can anybody explain in a nutshell how can one use it in engine tunning? I find the doc very interesting but pretty abstract.

[Rémi Coulom]

Can anybody explain in a nutshell how can one use it in engine tunning? I find the doc very interesting but pretty abstract.

Did you read the doc in the "README" file?

Did you manage to open DummyExperiment.clop with CLOP?

Once you manage to run this experiment, you should be all set up. All you have to do is write your own scripts to replace DummyScript.py. Run DummyScript.py without arguments (or look at the source) for an explanation.

Rémi

[Rein Halbersma]

I don't really understand your question.

The objective function can be the expected value of any random variable that depends on parameters. So, if, instead of playing one game and getting the result, you play a tournament and observe LOS over a specific opponent, you can use CLOP to optimize it. But that would be a strange way to use CLOP. If you wish to optimize a program against a set of opponents instead of just one opponent, you can use the "Replications" option of CLOP to play a game against each opponent, and then CLOP will maximize the average winning rate against all these opponents.

I don't mean to maximize the number of points against all these opponents. I mean to maximize the chance that I end up first in a tournament. E.g. playing thousands of games against opposition with an elo of 2950 +/- 1, I would prefer my program to be rated 2940 +/- 20, rather than 2945 +/- 1. The former should have a much larger chance to finish first in a tournament than the latter, even though it would score less points on average.

Is there any way to do such optimization? Can you simply make LOS the objective function?