BayeOpt stuck at local optima

[wqi]

First of all thank you for providing this wonderful tool. 

I've just started using it for my research which involves optimizing black-box problems. 

The first problem (8-D) blew my mind, BayesOpt was able to locate the best design faster than my previous genetic algorithm trial.

But something is happening for my second (11-D) design, it seems to get stuck at a sub-optimal point and started the parameter fine tuning (exploitation) thereafter.

My intuition is that my optimization problem resembles a 11-D version of the Ackley function...

Here are the key settings of this algorithm:

------------------------------

n_iterations=400

n_inner_iterations=500

n_init_samples=100

n_iter_relearn=1

init_method=1

surr_name=sGaussianProcess

sigma_s=1

noise=1e-10

alpha=1

beta=1

sc_type=SC_MAP

l_type=L_EMPIRICAL

l_all=0

epsilon=0

force_jump=10

kernel.name=kMaternARD5

kernel.hp_mean=[1](1)

kernel.hp_std=[1](10)

mean.name=mConst

mean.coef_mean=[1](1)

mean.coef_std=[1](1000)

crit_name=cEI

crit_params=[0]()

---------------------------------

I've already tried Hedge criterion, not much improvement was observed. Now I'm considering the annealed version of EI which seems to indicate later exploitation.

Any suggestions on how I can configure the algorithm would be greatly appreciated, thank you very much in advance! 

[Ruben Martinez-Cantin]

Hi,

The exploration/exploitation tuning can be problematic for some
functions. That's why BayesOpt is so customizable. The annealed EI can
be a good approach. I would also suggest using L_MCMC, trying
different surrogates or kernels, changing the kernel prior, etc.

Probably you don't need that many init samples.

Best,

Ruben

> --
> You received this message because you are subscribed to the Google Groups
> "BayesOpt discussion" group.
> To unsubscribe from this group and stop receiving emails from it, send an
> email to bayesopt-discus...@googlegroups.com.
> To post to this group, send email to bayesopt-...@googlegroups.com.
> Visit this group at https://groups.google.com/group/bayesopt-discussion.