use hyperband / bayesian probability for tuner instead of RandomSearch

[Palashio]

currently we're using random-search from keras-tuner. Instead we should be using probabilistic models