How Hyper-opt can suggest a new configuration

[Phu Ha-Van]

Hi everyone,

I am reading the hyper-opt code to understand the use of TPE algorithm. I see the function rec_eval(...) in hyperopt/pyll/base.py could be a function to suggest a new configuration for evaluation. Let's say I run hyper-opt for greater than 20 iterations (because it randomizes selecting the configuration in 20 first iterations. I saw the parameter _default_n_startup_jobs as the number of points for initialization). 

I get lost in understanding the function rec_eval(...) and don't see why it can suggest a new configuration. Could you briefly summarize what happens inside this function?

Thank you very much,

Phu