Abstract

Machine learning pipeline potentially consists of several stages ofoperations like data preprocessing, feature engineering and machine learningmodel training. Each operation has a set of hyper-parameters, which can becomeirrelevant for the pipeline when the operation is not selected. This gives riseto a hierarchical conditional hyper-parameter space. To optimize this mixedcontinuous and discrete conditional hierarchical hyper-parameter space, wepropose an efficient pipeline search and configuration algorithm which combinesthe power of Reinforcement Learning and Bayesian Optimization. Empiricalresults show that our method performs favorably compared to state of the artmethods like Auto-sklearn , TPOT, Tree Parzen Window, and Random Search.

 

Quick Read (beta)