mirrors/z3
3
0
Fork 0
mirror of https://github.com/Z3Prover/z3 synced 2025-04-07 18:05:21 +00:00
Code Activity

adding maxlex, throttle use of asymmetric literal addition

Browse source 
Signed-off-by: Nikolaj Bjorner <nbjorner@microsoft.com>
This commit is contained in:
Nikolaj Bjorner 2019-01-24 19:26:44 -08:00
parent 8da1d6070b
commit ad81fee118
9 changed files with 274 additions and 27 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

1
src/opt/CMakeLists.txt
View file

@ -1,5 +1,6 @@
z3_add_component(opt
SOURCES
maxlex.cpp
maxres.cpp
maxsmt.cpp
opt_cmds.cpp

185
src/opt/maxlex.cpp Normal file
View file

@ -0,0 +1,185 @@
/*++
Copyright (c) 2019 Microsoft Corporation
Module Name:
maxlex.cpp
Abstract:
MaxLex solves weighted max-sat problems where weights impose lexicographic order.
MaxSAT is particularly easy for this class:
In order of highest weight, check if soft constraint can be satisfied.
If so, assert it, otherwise assert the negation and record whether the soft
constraint is true or false in the solution.
Author:
Nikolaj Bjorner (nbjorner) 2019-25-1
--*/
#include "opt/opt_context.h"
#include "opt/maxsmt.h"
#include "opt/maxlex.h"
namespace opt {
bool is_maxlex(weights_t & _ws) {
vector<rational> ws(_ws);
std::sort(ws.begin(), ws.end());
ws.reverse();
rational sum(0);
for (rational const& w : ws) {
sum += w;
}
for (rational const& w : ws) {
if (sum > w + w) return false;
sum -= w;
}
return true;
}
class maxlex : public maxsmt_solver_base {
struct cmp_soft {
bool operator()(soft const& s1, soft const& s2) const {
return s1.weight > s2.weight;
}
};
ast_manager& m;
maxsat_context& m_c;
void update_assignment() {
model_ref mdl;
s().get_model(mdl);
if (mdl) {
m_c.model_updated(mdl.get());
update_assignment(mdl);
}
}
void assert_value(soft& soft) {
switch (soft.value) {
case l_true:
s().assert_expr(soft.s);
break;
case l_false:
s().assert_expr(expr_ref(m.mk_not(soft.s), m));
break;
default:
break;
}
}
void set_value(soft& soft, lbool v) {
soft.set_value(v);
assert_value(soft);
}
void update_assignment(model_ref & mdl) {
for (auto & soft : m_soft) {
switch (soft.value) {
case l_undef:
if (mdl->is_true(soft.s)) {
set_value(soft, l_true);
}
else {
update_bounds();
return;
}
break;
case l_true:
break;
case l_false:
break;
}
}
update_bounds();
}
void update_bounds() {
m_lower.reset();
m_upper.reset();
bool prefix_defined = true;
for (auto & soft : m_soft) {
if (!prefix_defined) {
m_upper += soft.weight;
continue;
}
switch (soft.value) {
case l_undef:
prefix_defined = false;
m_upper += soft.weight;
break;
case l_true:
break;
case l_false:
m_lower += soft.weight;
m_upper += soft.weight;
break;
}
}
trace_bounds("maxlex");
}
void init() {
model_ref mdl;
s().get_model(mdl);
update_assignment(mdl);
}
public:
maxlex(maxsat_context& c, unsigned id, weights_t & ws, expr_ref_vector const& s):
maxsmt_solver_base(c, ws, s),
m(c.get_manager()),
m_c(c) {
cmp_soft cmp;
std::sort(m_soft.begin(), m_soft.end(), cmp);
}
lbool operator()() override {
init();
for (auto & soft : m_soft) {
if (soft.value == l_true) {
continue;
}
SASSERT(soft.value() == l_undef);
expr* a = soft.s;
lbool is_sat = s().check_sat(1, &a);
switch (is_sat) {
case l_false:
set_value(soft, l_false);
update_bounds();
break;
case l_true:
update_assignment();
SASSERT(soft.value == l_true);
break;
case l_undef:
return l_undef;
}
}
return l_true;
}
void commit_assignment() override {
for (auto & soft : m_soft) {
if (soft.value == l_undef) {
return;
}
assert_value(soft);
}
}
};
maxsmt_solver_base* mk_maxlex(maxsat_context& c, unsigned id, weights_t & ws, expr_ref_vector const& soft) {
return alloc(maxlex, c, id, ws, soft);
}
}

32
src/opt/maxlex.h Normal file
View file

@ -0,0 +1,32 @@
/*++
Copyright (c) 2014 Microsoft Corporation
Module Name:
maxlex.h
Abstract:
MaxLex solves weighted max-sat problems where weights impose lexicographic order.
Author:
Nikolaj Bjorner (nbjorner) 2019-25-1
Notes:
--*/
#ifndef MAXLEX_H_
#define MAXLEX_H_
namespace opt {
bool is_maxlex(weights_t & ws);
maxsmt_solver_base* mk_maxlex(maxsat_context& c, unsigned id, weights_t & ws, expr_ref_vector const& soft);
};
#endif

8
src/opt/maxres.cpp
View file

@ -329,8 +329,8 @@ public:
verify_assumptions();
m_lower.reset();
for (soft& s : m_soft) {
s.is_true = m_model->is_true(s.s);
if (!s.is_true) {
s.set_value(m_model->is_true(s.s));
if (!s.is_true()) {
m_lower += s.weight;
}
}
@ -764,7 +764,7 @@ public:
TRACE("opt", tout << "updated upper: " << upper << "\nmodel\n" << *m_model;);
for (soft& s : m_soft) {
s.is_true = m_model->is_true(s.s);
s.set_value(m_model->is_true(s.s));
}
verify_assignment();
@ -878,7 +878,7 @@ public:
expr_ref n(m);
for (soft& s : m_soft) {
n = s.s;
if (!s.is_true) {
if (!s.is_true()) {
n = mk_not(m, n);
}
_solver->assert_expr(n);

22
src/opt/maxsmt.cpp
View file

@ -18,17 +18,18 @@ Notes:
--*/
#include <typeinfo>
#include "util/uint_set.h"
#include "ast/ast_pp.h"
#include "ast/ast_util.h"
#include "ast/pb_decl_plugin.h"
#include "opt/maxsmt.h"
#include "opt/maxres.h"
#include "opt/maxlex.h"
#include "opt/wmax.h"
#include "opt/opt_params.hpp"
#include "ast/ast_pp.h"
#include "util/uint_set.h"
#include "opt/opt_context.h"
#include "smt/theory_wmaxsat.h"
#include "smt/theory_pb.h"
#include "ast/ast_util.h"
#include "ast/pb_decl_plugin.h"
namespace opt {
@ -61,7 +62,7 @@ namespace opt {
rational k(0), cost(0);
vector<rational> weights;
for (soft const& s : m_soft) {
if (s.is_true) {
if (s.is_true()) {
k += s.weight;
}
else {
@ -80,13 +81,13 @@ namespace opt {
m_lower.reset();
m_upper.reset();
for (soft& s : m_soft) {
s.is_true = m.is_true(s.s);
if (!s.is_true) m_upper += s.weight;
s.set_value(m.is_true(s.s));
if (!s.is_true()) m_upper += s.weight;
}
TRACE("opt",
tout << "upper: " << m_upper << " assignments: ";
for (soft& s : m_soft) tout << (s.is_true?"T":"F");
for (soft& s : m_soft) tout << (s.is_true()?"T":"F");
tout << "\n";);
return true;
}
@ -234,7 +235,10 @@ namespace opt {
symbol const& maxsat_engine = m_c.maxsat_engine();
IF_VERBOSE(1, verbose_stream() << "(maxsmt)\n";);
TRACE("opt_verbose", s().display(tout << "maxsmt\n") << "\n";);
if (m_soft_constraints.empty() || maxsat_engine == symbol("maxres") || maxsat_engine == symbol::null) {
if (is_maxlex(m_weights)) {
m_msolver = mk_maxlex(m_c, m_index, m_weights, m_soft_constraints);
}
else if (m_soft_constraints.empty() || maxsat_engine == symbol("maxres") || maxsat_engine == symbol::null) {
m_msolver = mk_maxres(m_c, m_index, m_weights, m_soft_constraints);
}
else if (maxsat_engine == symbol("pd-maxres")) {

13
src/opt/maxsmt.h
View file

@ -59,10 +59,13 @@ namespace opt {
struct soft {
expr_ref s;
rational weight;
bool is_true;
soft(expr_ref const& s, rational const& w, bool t): s(s), weight(w), is_true(t) {}
soft(soft const& other):s(other.s), weight(other.weight), is_true(other.is_true) {}
soft& operator=(soft const& other) { s = other.s; weight = other.weight; is_true = other.is_true; return *this; }
lbool value;
void set_value(bool t) { value = t?l_true:l_undef; }
void set_value(lbool t) { value = t; }
bool is_true() const { return value == l_true; }
soft(expr_ref const& s, rational const& w, bool t): s(s), weight(w), value(t?l_true:l_undef) {}
soft(soft const& other):s(other.s), weight(other.weight), value(other.value) {}
soft& operator=(soft const& other) { s = other.s; weight = other.weight; value = other.value; return *this; }
};
ast_manager& m;
maxsat_context& m_c;
@ -84,7 +87,7 @@ namespace opt {
~maxsmt_solver_base() override {}
rational get_lower() const override { return m_lower; }
rational get_upper() const override { return m_upper; }
bool get_assignment(unsigned index) const override { return m_soft[index].is_true; }
bool get_assignment(unsigned index) const override { return m_soft[index].is_true(); }
void collect_statistics(statistics& st) const override { }
void get_model(model_ref& mdl, svector<symbol>& labels) override { mdl = m_model.get(); labels = m_labels;}
virtual void commit_assignment();

2
src/opt/sortmax.cpp
View file

@ -114,7 +114,7 @@ namespace opt {
}
void update_assignment() {
for (soft& s : m_soft) s.is_true = is_true(s.s);
for (soft& s : m_soft) s.set_value(is_true(s.s));
}
bool is_true(expr* e) {

2
src/opt/wmax.cpp
View file

@ -122,7 +122,7 @@ namespace opt {
}
void update_assignment() {
for (soft& s : m_soft) s.is_true = is_true(s.s);
for (soft& s : m_soft) s.set_value(is_true(s.s));
}
struct compare_asm {

36
src/sat/sat_simplifier.cpp
View file

@ -1014,6 +1014,9 @@ namespace sat {
svector<bool> m_in_intersection;
unsigned m_ala_qhead;
clause_wrapper m_clause;
unsigned m_ala_cost;
unsigned m_ala_benefit;
unsigned m_ala_max_cost;
blocked_clause_elim(simplifier & _s, unsigned limit, model_converter & _mc, use_list & l,
vector<watch_list> & wlist):
@ -1021,8 +1024,11 @@ namespace sat {
m_counter(limit),
m_mc(_mc),
m_queue(l, wlist),
m_clause(null_literal, null_literal) {
m_clause(null_literal, null_literal),
m_ala_cost(0),
m_ala_benefit(0) {
m_in_intersection.resize(s.s.num_vars() * 2, false);
m_ala_max_cost = (s.s.m_clauses.size() * s.m_num_calls)/5;
}
void insert(literal l) {
@ -1034,6 +1040,10 @@ namespace sat {
return !s.s.is_assumption(v) && !s.was_eliminated(v) && !s.is_external(v) && s.value(v) == l_undef;
}
bool reached_max_cost() {
return m_ala_benefit <= m_ala_cost * 100 && m_ala_cost > m_ala_max_cost;
}
enum elim_type {
bce_t,
cce_t,
@ -1296,13 +1306,15 @@ namespace sat {
*/
bool add_ala() {
unsigned init_size = m_covered_clause.size();
for (; m_ala_qhead < m_covered_clause.size() && m_ala_qhead < 5*init_size; ++m_ala_qhead) {
for (; m_ala_qhead < m_covered_clause.size() && m_ala_qhead < 5*init_size && !reached_max_cost(); ++m_ala_qhead) {
++m_ala_cost;
literal l = m_covered_clause[m_ala_qhead];
for (watched & w : s.get_wlist(~l)) {
if (w.is_binary_non_learned_clause()) {
literal lit = w.get_literal();
if (revisit_binary(l, lit)) continue;
if (s.is_marked(lit)) {
++m_ala_benefit;
return true;
}
if (!s.is_marked(~lit)) {
@ -1333,6 +1345,7 @@ namespace sat {
}
if (!ok) continue;
if (lit1 == null_literal) {
++m_ala_benefit;
return true;
}
m_covered_clause.push_back(~lit1);
@ -1504,7 +1517,9 @@ namespace sat {
template<elim_type et>
void cce_binary() {
while (!m_queue.empty() && m_counter >= 0) {
m_ala_cost = 0;
m_ala_benefit = 0;
while (!m_queue.empty() && m_counter >= 0 && !reached_max_cost()) {
s.checkpoint();
process_cce_binary<et>(m_queue.next());
}
@ -1516,7 +1531,7 @@ namespace sat {
watch_list & wlist = s.get_wlist(~l);
m_counter -= wlist.size();
model_converter::kind k;
for (watched & w : wlist) {
for (watched& w : wlist) {
if (!w.is_binary_non_learned_clause()) continue;
if (!select_clause<et>(2)) continue;
literal l2 = w.get_literal();
@ -1543,9 +1558,13 @@ namespace sat {
template<elim_type et>
void cce_clauses() {
literal blocked;
m_ala_cost = 0;
m_ala_benefit = 0;
model_converter::kind k;
for (clause* cp : s.s.m_clauses) {
clause& c = *cp;
unsigned start = s.s.m_rand();
unsigned sz = s.s.m_clauses.size();
for (unsigned i = 0; i < sz; ++i) {
clause& c = *s.s.m_clauses[(i + start) % sz];
if (c.was_removed() || c.is_learned()) continue;
if (!select_clause<et>(c.size())) continue;
elim_type r = cce<et>(c, blocked, k);
@ -1563,7 +1582,10 @@ namespace sat {
break;
}
s.checkpoint();
}
if (reached_max_cost()) {
break;
}
}
}
void inc_bc(elim_type et) {