Is BOLFI robust to scaling of the parameters? about elfi HOT 4 CLOSED

I'm not an expert in GPs, but I think you should use a different kernel. The default kernel has just one length scale for all parameters. So for example something like:

kernel = GPy.kern.RBF(2, lengthscale=[100, 1], ARD=True)
kernel += GPy.kern.Bias(2)
bounds = {'t1':(-20000, 20000), 't2':(-1, 1)}
gp = elfi.GPyRegression(['t1', 't2'], bounds=bounds, kernel=kernel)
bolfi = elfi.BOLFI(log_d, batch_size=1, initial_evidence=20, update_interval=10,
                   bounds=bounds, target_model=gp, acq_noise_var=[100, 0.1], seed=seed)

I hope this helps you in the right direction.

Also, please update ELFI to the newest version 0.7.3. :)

from elfi.

Thanks for your fast response. I tried out your suggestion and for the acquisition functions it looks like it helps.
Do you know, if I have to take a similar scaling into account for the NUTS sampling? Because as you can see in the plots below, the chains get stuck in the larger dimension and the posterior are still not that great.
The four chains:

And the resulting marginals of the posterior: True values are (6000,0.2)

from elfi.

I suppose very different scales can be problematic for NUTS as well, and we have considered implementing some kind of normalizing. You could also try the basic Metropolis sampler, which allows (requires, actually) manually setting the "scale" of each parameter.

Have you considered trying the opposite in your case: having roughly equal parameter scales and then upscaling in the simulator?

from elfi.

Yeah, I already did a normalization/rescaling in my simulator and it seems to work.
I just thought this is an issue other users could be interested in, since different scales in parameters are quite common for real world problems.
Maybe it would be nice to drop a hint in the documentation. Reading the BOLFI paper this issue wasn't clear to me.

from elfi.