Department of Computer Science and Technology

Technical reports

Machine learning and computer algebra

Zongyan Huang

April 2016, 113 pages

This technical report is based on a dissertation submitted August 2015 by the author for the degree of Doctor of Philosophy to the University of Cambridge, Lucy Cavendish College.

DOI: 10.48456/tr-884

Abstract

Computer algebra is a fundamental research field of computer science, and computer algebra systems are used in a wide range of problems, often leading to significant savings of both time and effort. However, many of these systems offer different heuristics, decision procedures, and parameter settings to tackle any given problem, and users need to manually select them in order to use the system.

In this context, the algorithm selection problem is the problem of selecting the most efficient setting of the computer algebra system when faced with a particular problem. These choices can dramatically affect the efficiency, or even the feasibility of finding a solution. Often we have to rely on human expertise to pick a suitable choice as there are no fixed rules that determine the best approach, and even for experts, the relationship between the problem at hand and the choice of an algorithm is far from obvious.

Machine learning techniques have been widely applied in fields where decisions are to be made without the presence of a human expert, such as in web search, text categorization, or recommender systems. My hypothesis is that machine learning can also be applied to help solve the algorithm selection problem for computer algebra systems.

In this thesis, we perform several experiments to determine the effectiveness of machine learning (specifically using support vector machines) for the problem of algorithm selection in three instances of computer algebra applications over real closed fields. Our three applications are: (i) choosing decision procedures and time limits in MetiTarski; (ii) choosing a heuristic for CAD variable ordering; (iii) predicting the usefulness of Gröbner basis preconditioning. The results show that machine learning can effectively be applied to these applications, with the machine learned choices being superior to both choosing a single fixed individual algorithm, as well as to random choice.

Full text

PDF (1.5 MB)

BibTeX record

@TechReport{UCAM-CL-TR-884,
  author =	 {Huang, Zongyan},
  title = 	 {{Machine learning and computer algebra}},
  year = 	 2016,
  month = 	 apr,
  url = 	 {https://www.cl.cam.ac.uk/techreports/UCAM-CL-TR-884.pdf},
  institution =  {University of Cambridge, Computer Laboratory},
  doi = 	 {10.48456/tr-884},
  number = 	 {UCAM-CL-TR-884}
}