QUASI_RANDOM_SEARCH no work about vizier HOT 4 CLOSED

Hi @BYTERHIT, I took a look at your code - just making sure, for the range you're describing, are you using LOG or REVERSE_LOG scale?

As as FYI, when using LOG or REVERSE_LOG, the algorithm will apply a log-scale to transform your range [a,b] down to [0,1], uniformly sample points in this underlying [0,1], and then convert back to [a,b].

Note also that Quasi-Random-Search is not like regular random search; the points sampled in the underlying [0,1] space will be fairly spread out and unlikely to cover the edges of the boundary; see https://en.wikipedia.org/wiki/Halton_sequence, which may be the cause of your observed behavior in [0.01, 0.013].

To double check this is a matter of the algorithm, can you try just regular RANDOM_SEARCH and see what the sampled points are?

from vizier.

Hi @xingyousong I have checked that when using RANDOM_SEARCH, I get the normal results, the points are in the range [0.0001,0.1]. As follows:

And I did not use LOG or REVERSE_LOG scale, instead, I use LINEAR scale. The QUASI_RANDOM_SEARCH results as follows:

Also I posted the config file for vizier client parameters and the python file to generate the figure above.
config&plot.zip

Since the sample points are in a small fraction of the target range. I think it is abnormal. Using Halton Sequence should not get the results?

from vizier.

@BYTERHIT Thanks for the notification. I've made the fix (it was mainly just needing to set scrambled=True) in #542 and pushed to 0.1.3.

from vizier.

It works!

from vizier.