Local Minima Skip about optuna

[atoffano]

I have noticed the very same thing a couple weeks ago.
I assume that even running a RandomSampler for a couple iterations would still end with quickly falling back to the same minima since no other have been found.

I'm currently trying to first run a RandomSampler, and then switching to a TPESampler, the idea being that the RandomSampler samples multiple local minima that the TPE can then pick from.
No results to confirm yet, i am unsure whether that's actually a good workaround since i can't tell if data can be transfered from a sampler to another.

[ggmirandac]

Hi,

Yeah, I am doing the same, using the n_startup_trials = 100 (or some other bigger number) But I still found my self into the same local minima. :(

I am thinking that there is a problem with the dataset that I am analysing (has <100 data points). Do you know some way to make this better?

Bests