Download PDFOpen PDF in browser

We know (nearly) nothing!l But can we learn?

4 pagesPublished: November 8, 2017


The greatest source of progress in automated theorem proving in the last 30 years has been the development of better search heuristics, usually based on developer experience and empirical evaluation, but increasingly also using automated optimization techniques. Despite this progress, we still know very little about proof search. We have mostly failed to identify good features for characterizing homogeneous problem classes, or for identifying interesting and relevant clauses and formulas.
I propose the challenge of bringing together inductive techniques (generalization and learning) and deductive techniques to attack this problem. Hardware and software have finally evolved to a point that we can reasonably represent and analyze large proof searches and search decisions, and where we can hope to achieve order-of-magnitude improvements in the efficiency of the proof search.

Keyphrases: automated theorem proving, deduction, Heuristics, machine learning, search

In: Giles Reger and Dmitriy Traytel (editors). ARCADE 2017. 1st International Workshop on Automated Reasoning: Challenges, Applications, Directions, Exemplary Achievements, vol 51, pages 29--32

Download PDFOpen PDF in browser