[RFC] AdaTune: Bring adaptive features to the auto-tuner

Thanks very much for your all attention and the naming suggestions.

  • For the enable_adaptive_evaluator, I agree it’s more reasonable to set the threshold value which early terminates the repeat with the args like min_repeat_ms does. How about replacing it with a min_converge_coef?
  • I meant to show the dynamic feature in the variables. And the key change is to replace the fixed 5% randomization when do the e-greedy. Setting the uncertainty_aware is to make this dynamic-searching an optional choice as some cost models are not supported. How about a fixed_uncertainty(Fasle for AdaTune) and a randomization_rate? Or you some suggestions? Thanks!