SMAC-Optimizer-increment

What precisely is being increased by 5? There are many configuration settings for the SMAC optimizer.

Thanks.

In this case, with one integer parameter, it makes not much sense to use any optimizer different form "Exhaustive".

The more advanced optimizers are built for larger parameter spaces, say three or more parameters which may have floating point values.

In such a scenario the "Exhaustive" optimizer will either take a vary long time or needs big step-sizes to finish within practical time.

Here the "SMAC" and "Shrinking Window" optimizers shine. They find good parameter combinations with a low number of backtests. The SMAC optimizer also ignores the "Step" value and thus finds much more precise solutions.

@DrKoch

it works fine.
It was probably because I only took one parameter. Now I've done it with three parameters.

@Glitch - whilst SMAC ignores increments, saved (SMAC) strategy parameters are rounded however to the increment when loading the building block strategy - resulting in wrong default values. Can this please be corrected - thank you :)

We can’t, you’ll need to define finer increment values if you want to be able to save Strategies with those values.

Dr. Koch we now have a SMAC issue in that WL restricts loading saved float precision for defined parameter increments (also used with other optimizers). Can SMAC be changed to align with parameter increments ? or give the option to do so in the settings

Thank you

Ot is tecincally possible to restrict parameter values to step increments and in fact I have plans to implement this in one of the next builds. The same happens for integer parameterd, because SMAC originally uses float parameters only.

BUT: keep in mind that such restrictions worsen the overall result. It is qite possible that a too large step value causes SMAC to miss the true best parameter value.

QUOTE:
BUT: keep in mind that such restrictions worsen the overall result. It is quite possible that a too large step value causes SMAC to miss the true best parameter value.

I agree. It's better to let the optimizer find its own values. What would be the point of tying the hands of the optimizer?