browsing feasible space and population sizing

[bbd666]

2 questions, just to better understand :

1. My problem's got 75 bounded optimization variables. I monitor the set of variable in every trial and it seems they are all chosen between 0 and 1 irrespective of their range of variation given by cmaes_boundary_transformation. For example, one variable is set between 0 and 20, but each new search sets it lower than 1.

2. the size of the population is 16 = cmaes_Get(&evo, "lambda"). Can it be tuned or is it the result of the algorithm ?

[nikohansen]

the size of the population is 16 = cmaes_Get(&evo, "lambda"). Can it be tuned or is it the result of the algorithm ?

It is the input parameter lambda to the constructor cmaes_init or can be set via the https://github.com/CMA-ES/c-cmaes/blob/master/cmaes_initials.par file.

variation given by cmaes_boundary_transformation. For example, one variable is set between 0 and 20, but each new search sets it lower than 1.

I don't know. If the optimum is below 1, this is the expected behavior. Why is it not to be expected?

[bbd666]

thks for your prompt reply.

I asked the question because in the cmaes_initials.par file, lambda is set to 3. Then I looked for a hard-coding declaration of lambda in the source files and found no value of 16. Maybe it's a computed value of lambda.

I don't know. If the optimum is below 1, this is the expected behavior. Why is it not to be expected?

It's far far away from the optimum and it seems every new trial is between 0 and 1 among the 50 first iterations. But I confess it's not objective as an observation , I must carry on my search.

[nikohansen]

I asked the question because in the cmaes_initials.par file, lambda is set to 3.

This is (normally) not an eligible value for lambda, but it should be applied when read (which means it is probably not read). AFAICS, only values < 2 are replaced by the default.

It's far far away from the optimum and it seems every new trial is between 0 and 1 among the 50 first iterations.

Be award that the boundaries are unrelated to the initial condition for x0 (initialX) and sigma0 (initialStandardDeviations). Only the latter are expected to change the search behavior away from the bounds and they should be chosen with great care depending on your domain knowledge.

[bbd666]

This is (normally) not an eligible value for lambda, but it should be applied when read (which means it is probably not read). AFAICS, only values < 2 are replaced by the default.

sorry to bother you, I don't understand, where does the value of 16 come from ? the value 3 is set for lambda in cmaes_initials.par. You wrote this value is not eligible but it should be applied. What does it mean? t->lambda=16 indeed

Be award that the boundaries are unrelated to the initial condition for x0 (initialX) and sigma0 (initialStandardDeviations). Only the latter are expected to change the search behavior away from the bounds and they should be chosen with great care depending on your domain knowledge.

In understand better why searches are within [0-1]

[nikohansen]

I don't understand, where does the value of 16 come from ?

I don't know. I also don't have enough information to investigate or replicate. It would be the default in dimension 55.