Question about parameter bounds and sudden drops in objective value when using ‘make_genopheno_pwqb’ and ‘make_genopheno_pwqb_ls’

[gaofz121]

Dear authors,

I’m currently using libcmaes for parameter optimization in an image enhancement experiment,
and I’ve encountered an issue related to the genotype/phenotype mapping and parameter bounds.

When I use the following code:

gp = lcmaes.make_genopheno_pwqb_ls(low_bounds, up_bounds, len(up_bounds))
p = lcmaes.make_parameters_pwqb_ls(initial_params, sigma, gp, olambda, seed)

and

gp = lcmaes.make_genopheno_pwqb(low_bounds,up_bounds,len(up_bounds))
p = lcmaes.make_parameters_pwqb(initial_params,sigma,gp)

Parameter Settings:

olambda = 50
seed = 32
sigma = 0.1

low_bounds.append(low)
up_bounds.append(up)

the optimization runs, but I noticed two unexpected behaviors:
The optimized parameters sometimes go outside the specified bounds.
The objective value (evaluated using best_params through my objective function)
shows occasional abnormal drops, as shown in the figure below.

Questions：

1. How can I ensure the parameters remain strictly within the given bounds?
2. What might cause sudden drops in the objective value during optimization?

Thank you very much for your time and for maintaining this excellent library!

[nikohansen]

The optimized parameters sometimes go outside the specified bounds.

We don't see where or how this is the case. This is expected for the internal representations of the parameters.

What might cause sudden drops in the objective value during optimization?

At this point I would not have enough information to make an educated guess.

[gaofz121]

Thank you for your help!
It turns out the issue was on my side. I was retrieving the best parameters using:

bcand = cmasols.best_candidate()
best_params = lcmaes.get_candidate_x(bcand)

instead of
best_params = lcmaes.get_best_candidate_pheno(cmasols, gp)
After correcting this, the results are now within the expected parameter bounds.
Everything works perfectly now — thank you again for your clarification!

[beniz]

Make sure you are retrieving the solution after pheno is applied, as shown here: https://github.com/CMA-ES/libcmaes/wiki/Defining-and-using-bounds-on-parameters

I can't comment on the drops, but I'd be interested in the underlying application type, I like optimizing PSNR, in super-resolution typically (but not with CMA-ES) ;)