Questions for algorithm 2 in BOHB paper

[yjhong89]

Hi.

My name is YJ Hong and I am trying to use BOHB for my project.

I have few questions about how BOHB operates (especially in algorithm 2 in paper)

I followed 'optimize_mlp_digit example' and I set min_budget=5, max_budget=200, eta=2.

When I run the code, in the first time, BOHB finds the number of configuration to evaluate according to budget. In this case, 16 for smallest budget (12.5) (namely first step).

A major difference between BOHB and HB is that BOHB adopts BO (kind of TPE) to get configurations. So BOHB would find configurations after acquiring minimum number of data points by random.

And then next iterations, BOHB would select 8 best configurations (since eta=2) and evaluate those selected configurations for more budget (25) (second step).

My question is that how BO is applied after first step. According to 5 page in BOHB paper,

it seems like BO is continuously applied after first step.

If then, can you explain how BO is applied from second step please?

Thank you and have a good day.

Regards.