Machine Learning to Improve Cylindrical Algebraic Decomposition in Maple

Abstract

Many algorithms in computer algebra systems can have their performance improved through the careful selection of options that do not affect the correctness of the end result. Machine Learning (ML) is suited for making such choices: the challenge is to select an appropriate ML model, training dataset, and scheme to identify features of the input. In this extended abstract we survey our recent work to use ML to select the variable ordering for Cylindrical Algebraic Decomposition (CAD) in Maple: experimentation with a variety of models, and a new flexible framework for generating ML features from polynomial systems. We report that ML allows for significantly faster CAD than with the default Maple ordering, and discuss some initial results on adaptability.