Technical reports

# Machine learning and automated theorem proving

**James P. Bridge**

November 2010, 180 pages

This technical report is based on a dissertation submitted October 2010 by the author for the degree of Doctor of Philosophy to the University of Cambridge, Corpus Christi College.

**DOI:** 10.48456/tr-792

## Abstract

Computer programs to find formal proofs of theorems have a history going back nearly half a century. Originally designed as tools for mathematicians, modern applications of automated theorem provers and proof assistants are much more diverse. In particular they are used in formal methods to verify software and hardware designs to prevent costly, or life threatening, errors being introduced into systems from microchips to controllers for medical equipment or space rockets.

Despite this, the high level of human expertise required in their use means that theorem proving tools are not widely used by non specialists, in contrast to computer algebra packages which also deal with the manipulation of symbolic mathematics. The work described in this dissertation addresses one aspect of this problem, that of heuristic selection in automated theorem provers. In theory such theorem provers should be automatic and therefore easy to use; in practice the heuristics used in the proof search are not universally optimal for all problems so human expertise is required to determine heuristic choice and to set parameter values.

Modern machine learning has been applied to the automation of heuristic selection in a first order logic theorem prover. One objective was to find if there are any features of a proof problem that are both easy to measure and provide useful information for determining heuristic choice. Another was to determine and demonstrate a practical approach to making theorem provers truly automatic.

In the experimental work, heuristic selection based on features of the conjecture to be proved and the associated axioms is shown to do better than any single heuristic. Additionally a comparison has been made between static features, measured prior to the proof search process, and dynamic features that measure changes arising in the early stages of proof search. Further work was done on determining which features are important, demonstrating that good results are obtained with only a few features required.

## Full text

PDF (3.5 MB)

## BibTeX record

@TechReport{UCAM-CL-TR-792, author = {Bridge, James P.}, title = {{Machine learning and automated theorem proving}}, year = 2010, month = nov, url = {https://www.cl.cam.ac.uk/techreports/UCAM-CL-TR-792.pdf}, institution = {University of Cambridge, Computer Laboratory}, doi = {10.48456/tr-792}, number = {UCAM-CL-TR-792} }