Program Search for Machine Learning Pipelines Leveraging Symbolic Planning and Reinforcement Learning

Abstract

In this paper we investigate an alternative knowledge representation and learning strategy for the automated machine learning (AutoML) task. Our approach combines a symbolic planner with reinforcement learning to evolve programs that process data and train machine learning classifiers. The planner, which generates all feasible plans from the initial state to the goal state, gives preference first to shortest programs and then later to ones that maximize rewards. The results demonstrate the efficacy of the approach for finding good machine learning pipelines, while at the same time showing that the representation can be used to infer new knowledge relevant for the problem instances being solved. These insights can be useful for other automatic programming approaches, like genetic programming (GP) and Bayesian optimization pipeline learning, with respect to representation and learning strategies.