Tactic Script Optimisation for Aesop

Jannis Limperg

Abstract: White-box proof search tactics such as Coq's auto, Isabelle's auto and Lean's Aesop apply proof rules that often translate directly to lower-level tactics. When these search tactics find a proof, they can, at least in principle, generate a tactic script, i.e. a sequence of lower-level tactics that proves the goal. Such a script can then replace the typically much slower invocation of the search tactic, or it can be used to understand and debug the search. However, at least for Aesop, naively generated scripts are highly unidiomatic. For example, they use tactics that users would typically avoid; they apply these tactics in an unusual order; and they do not use Lean's constructs for structuring tactic scripts. To address these issues, I propose a three-stage optimisation pipeline that transforms naive scripts into scripts resembling a human-written Lean proof. The first stage replaces less idiomatic tactics with more idiomatic ones, checking that the two sequences of tactics produce the same result. The second stage permutes the tactics in the script to bring them into a natural, depth-first order, which then allows us to use Lean's structuring constructs. The third stage runs some straightforward post-processing passes.

External IDs:dblp:conf/cpp/Limperg25