Title: Recommending Symbol Precedences by Machine Learning

Abstract:

    Proof search in state-of-the-art automated theorem provers for first-order logic is constrained by a simplification ordering on terms, for example the Knuth-Bendix ordering. Such ordering is typically parameterized by a symbol precedence—a permutation of the predicate and function symbols of the input problem. The choice of the symbol precedence may have a substantial impact on the amount of work required to complete a proof search. Neural Precedence Recommender is a system that, given an input problem, produces a precedence that is estimated to yield a good performance with the theorem prover Vampire. The recommender, consisting namely of a graph neural network, is trained on proof searches with random precedences. This presentation describes the design and an evaluation of the recommender