/* Copyright (c) 2015 Microsoft Corporation. All rights reserved. Released under Apache 2.0 license as described in the file LICENSE. Author: Leonardo de Moura This module implements a heuristic for selecting relevant theorems based on the approach described at "Lightweight relevance filtering for machine-generated resolution problems" Jia Meng and Larry Paulson Journal of Applied Logic 7 2009 */ #include #include "util/sexpr/option_declarations.h" #include "kernel/environment.h" #include "library/decl_stats.h" #include "library/private.h" #ifndef LEAN_DEFAULT_MENG_PAULSON_P #define LEAN_DEFAULT_MENG_PAULSON_P 0.6 #endif #ifndef LEAN_DEFAULT_MENG_PAULSON_C #define LEAN_DEFAULT_MENG_PAULSON_C 2.4 #endif namespace lean { static name * g_meng_paulson_p = nullptr; static name * g_meng_paulson_c = nullptr; void initialize_meng_paulson() { g_meng_paulson_p = new name{"meng_paulson", "p"}; g_meng_paulson_c = new name{"meng_paulson", "c"}; register_double_option(*g_meng_paulson_p, LEAN_DEFAULT_MENG_PAULSON_P, "(theorem selection) control parameter for the Meng&Paulson theorem selection heuristic" "(for details see paper \"Lightweight relevance filtering for machine-generated resolution problems)\""); register_double_option(*g_meng_paulson_c, LEAN_DEFAULT_MENG_PAULSON_C, "(theorem selection) control parameter for the Meng&Paulson theorem selection heuristic" "(for details see paper \"Lightweight relevance filtering for machine-generated resolution problems)\""); } void finalize_meng_paulson() { delete g_meng_paulson_p; delete g_meng_paulson_c; } double get_meng_paulson_p(options const & o) { return o.get_double(*g_meng_paulson_p, LEAN_DEFAULT_MENG_PAULSON_P); } double get_meng_paulson_c(options const & o) { return o.get_double(*g_meng_paulson_c, LEAN_DEFAULT_MENG_PAULSON_C); } class relevant_thms_fn { environment m_env; double m_p; double m_c; name_set m_relevant; double get_weight(name const & n) const { double r = get_num_occs(m_env, n); return 1.0 + 2.0 / log(r + 1.0); } double get_thm_score(name const & n) const { name_set s = get_use_set(m_env, n); unsigned IR = 0; double M = 0.0; s.for_each([&](name const & F) { if (m_relevant.contains(F)) { M += get_weight(F); } else { IR++; } }); if (M > 0.0) return M / (M + IR); else return 0.0; } public: relevant_thms_fn(environment const & env, double p, double c, name_set const & rel): m_env(env), m_p(p), m_c(c), m_relevant(rel) { lean_assert(c > 0.0); } name_set operator()() { name_set A; while (true) { name_set Rel; m_relevant.for_each([&](name const & c) { name_set used_by = get_used_by_set(m_env, c); used_by.for_each([&](name const & T) { declaration const & T_decl = m_env.get(T); if (A.contains(T)) return; // T is already in the result set if (!T_decl.is_theorem() && !T_decl.is_axiom()) return; // we only care about axioms and theorems if (is_private(m_env, T)) return; // we ignore private decls double M = get_thm_score(T); if (M < m_p) return; // score is to low Rel.insert(T); A.insert(T); }); }); if (Rel.empty()) break; // include symbols of new theorems in m_relevant Rel.for_each([&](name const & T) { name_set uses = get_use_set(m_env, T); uses.for_each([&](name const & c) { declaration const & c_decl = m_env.get(c); if (c_decl.is_theorem() || c_decl.is_axiom()) return; // we ignore theorems occurring in types if (is_private(m_env, c)) return; // we ignore private decls m_relevant.insert(c); }); }); m_p = m_p + (1.0 - m_p) / m_c; } return A; } }; name_set get_relevant_thms(environment const & env, double p, double c, name_set const & relevant_symbols) { return relevant_thms_fn(env, p, c, relevant_symbols)(); } name_set get_relevant_thms(environment const & env, options const & o, name_set const & relevant_symbols) { return relevant_thms_fn(env, get_meng_paulson_p(o), get_meng_paulson_c(o), relevant_symbols)(); } }