vampire-sys 0.5.2

Low-level FFI bindings to the Vampire theorem prover (use the 'vampire' crate instead)
Documentation
/*
 * This file is part of the source code of the software program
 * Vampire. It is protected by applicable
 * copyright laws.
 *
 * This source code is distributed under the licence found here
 * https://vprover.github.io/license.html
 * and in the source directory
 */
/**
 * @file InterpolantMinimizer.cpp
 * Implements class InterpolantMinimizer.
 * @author Bernhard Gleiss
 */

#if VZ3

#include "InterpolantMinimizer.hpp"

#include "Lib/VirtualIterator.hpp"

#include "Kernel/Unit.hpp"

#include <memory>
#include <unordered_set> // TODO: remove this after benchmarking
#include "z3++.h"

namespace Shell
{
    using namespace std;
    using namespace Kernel;

    std::unordered_map<Kernel::Unit*, Kernel::Color> InterpolantMinimizer::computeSplittingFunction(Kernel::Unit* refutation,  UnitWeight weightFunction)
    {
        using namespace z3;
        context c;
        optimize solver(c);

        std::unordered_map<Unit*, std::unique_ptr<expr>> unitsToExpressions; // needed in order to map the result of the optimisation-query back to the inferences.
        std::unordered_map<Unit*, std::unique_ptr<expr>> unitsToPenalties;
        int i = 0; // counter needed for unique names

        // note: idea from the thesis: we use x_i to denote whether inference i is assigned to the A-part.
        ProofIteratorPostOrder it(refutation);
        while (it.hasNext()) // traverse the proof in depth-first post order
        {
            Unit* current = it.next();

            // construct a new expression representing the current inference and
            // remember that the current inference is encoded by that expression
            std::unique_ptr<expr> x_i_ptr (new expr(c));
            *x_i_ptr = c.bool_const(("x_" + std::to_string(i)).c_str()); // ugly but necessary due to z3-API
            unitsToExpressions.emplace(std::make_pair(current, std::move(x_i_ptr)));

            // construct a new expression representing the current penalty and remember it
            std::unique_ptr<expr> p_i_ptr (new expr(c));
            *p_i_ptr = c.bool_const(("p_" + std::to_string(i)).c_str()); // ugly but necessary due to z3-API
            unitsToPenalties.emplace(std::make_pair(current, std::move(p_i_ptr)));

            // increase i in order to get unique names
            i++;

            expr& x_i = *unitsToExpressions.at(current);
            // if inference is an axiom we need to assign it to the corresponding partition
            if (current->inheritedColor() == COLOR_LEFT)
            {
                solver.add(x_i);
            }
            else if (current->inheritedColor() == COLOR_RIGHT)
            {
                solver.add(!x_i);
            }

            // if the conclusion contains a colored symbol, we need to assign the inference to the corresponding partition
            if (current->getColor() == COLOR_LEFT)
            {
                solver.add(x_i);
            }
            else if (current->getColor() == COLOR_RIGHT)
            {
                solver.add(!x_i);
            }

            // if any parent contains a colored symbol, we need to assign the inference to the corresponding partition
            VirtualIterator<Unit*> parents = current->getParents();
            while (parents.hasNext())
            {
                Unit* premise= parents.next();
                if (premise->getColor() == COLOR_LEFT)
                {
                    solver.add(x_i);
                }
                else if (premise->getColor() == COLOR_RIGHT)
                {
                    solver.add(!x_i);
                }
            }

            // TODO: the above may be adding the same constraint more than once

            // now add the main constraints: the conclusion of a parent-inference is included in the interpolant iff the
            // the parent inference is assigned a different partition than the current inference
            parents = current->getParents();
            while (parents.hasNext())
            {
                Unit* premise= parents.next();

                expr& x_j = *unitsToExpressions.at(premise);
                expr& p_j = *unitsToPenalties.at(premise);

                solver.add(!(x_i != x_j) || p_j);
            }
        }

        // add the function we want to minimise to the solver
        /*
         * Note: we want to use the weighted sum of literals as measurement,
         * but this doesn't exactly correspond to the size of the interpolant,
         * since we are not counting the connectives.
         * we can now use the fact that each included literal except the first one
         * introduces exactly one connective (from an NNF-perspective), so we therefore
         * add 1 to each weight. Afterwards we need to subtract 1, if there is at least one literal
         * in the interpolant, since the first element doesn't introduce a connective.
         */
        expr penaltyFunction = c.real_val(0);

        for (const auto& keyValuePair : unitsToPenalties)
        {
            expr& p_i = *keyValuePair.second;
            double weight = weightForUnit(keyValuePair.first, weightFunction) + 1;
            expr weightExpression = c.real_val(std::to_string(weight).c_str());

            penaltyFunction = penaltyFunction + ite(p_i, weightExpression, c.real_val(0));
        }
        solver.minimize(penaltyFunction);

        // set a time limit for z3 call in order to being able to fallback to heuristic splitting function for very big proofs
        /* ":timeout" seems not to by supported by (the current version of) Z3 in this context
        params p(c);
        p.set(":timeout", static_cast<unsigned>(60000)); // in milliseconds, i.e. 1 minute
        solver.set(p);
        */

        // we are now finished with adding constraints, so use z3 to compute an optimal model
        check_result result = solver.check();

        // if the optimization procedure finds any model
        // Note: doesn't have to be the optimal one (Assumption: a non-optimal model still produces a better splitting function than the heuristic approach)
        if (result == check_result::sat)
        {
            bool containsLeftInference = false; // needed for weight prediction
            bool containsRightInference = false; // needed for weight prediction

            // convert computed model to splitting function
            model m = solver.get_model();

            std::unordered_map<Kernel::Unit*, Kernel::Color> splittingFunction;
            for (const auto& keyValuePair : unitsToExpressions)
            {
                Unit* current = keyValuePair.first;
                expr evaluation = m.eval(*unitsToExpressions[current]);

                if (Z3_get_bool_value(c,evaluation) == Z3_L_TRUE)
                {
                    splittingFunction[current] = COLOR_LEFT;
                    containsLeftInference = true;
                }
                else
                {
                    splittingFunction[current] = COLOR_RIGHT;
                    containsRightInference = true;
                }
            }

            // print weight prediction
            if (containsLeftInference && containsRightInference)
            {
                cout << "expecting interpolant weight " << m.eval(penaltyFunction - c.real_val(1)) << endl;
            }
            else
            {
                cout << "expecting interpolant weight " << m.eval(penaltyFunction) << endl;
            }

            return splittingFunction;
        }
        // otherwise use heuristic approach as fallback
        else
        {
            return Interpolants::computeSplittingFunction(refutation, weightFunction);
        }
    }

    void InterpolantMinimizer::analyzeLocalProof(Kernel::Unit *refutation)
    {
        // print statistics on grey area
        analyzeGreyAreas(refutation);

        // compute number of red subproofs
        const std::unordered_map<Unit*, Color> splittingFunction = computeSplittingFunction(refutation, UnitWeight::VAMPIRE);
        const std::unordered_map<Unit*, Unit*> unitsToRepresentative = computeSubproofs(refutation, splittingFunction);

        std::unordered_set<Unit*> representatives;
        for (const auto& keyValuePair : unitsToRepresentative)
        {
            representatives.insert(keyValuePair.second);
        }
        cout << "Number of red subproofs: " << representatives.size() << endl;

        // compute number of alternations between red and blue subproofs
        /*
         * Def.: if we have a refutation
         *          A
         *        -----
         *          B
         *        -----
         *          A
         * we count it as 2 alternations (although one could argue for 3 alternations)
         */
        std::unordered_map<Unit*, int> alternations;

        ProofIteratorPostOrder it(refutation);
        while (it.hasNext()) // traverse the proof in depth-first post order
        {
            Unit* current = it.next();

            // if current is an axiom
            if (!current->getParents().hasNext())
            {
                alternations[current] = 0;
            }
            else
            {
                int alternations_max = 0;

                VirtualIterator<Unit*> parents = current->getParents();
                while (parents.hasNext())
                {
                    Unit* premise= parents.next();

                    if (splittingFunction.at(premise) != splittingFunction.at(current))
                    {
                        alternations_max = std::max(alternations.at(premise) + 1, alternations_max);
                    }
                    else
                    {
                        alternations_max = std::max(alternations.at(premise), alternations_max);
                    }
                }
                alternations[current] = alternations_max;
            }
        }

        cout << "Number of alternations: " << alternations.at(refutation) << endl;
    }

    /*
     * compute both number of inferences which are necessarily assigned to red and to blue, and number of inferences which can be assigned arbitrarily
     * computes percentage grey / (red + blue + grey)
     */
    void InterpolantMinimizer::analyzeGreyAreas(Kernel::Unit* refutation)
    {
        int number_red = 0;
        int number_blue = 0;
        int number_grey = 0;

        /*
         * note: reuses code from heuristic splitting function
         */
        ProofIteratorPostOrder it(refutation);
        while (it.hasNext()) // traverse the proof in depth-first post order
        {
            Unit* current = it.next();
            ASS((!current->getParents().hasNext() && (current->inheritedColor() == COLOR_LEFT || current->inheritedColor() == COLOR_RIGHT)) || (current->getParents().hasNext() &&  current->inheritedColor() == COLOR_INVALID));

            // if the inference is an axiom, assign it to the corresponding partition
            if (!current->getParents().hasNext())
            {
                if (current->inheritedColor() == COLOR_LEFT)
                {
                    number_red++;
                }
                else
                {
                    number_blue++;
                }
                continue;
            }

            // if the inference contains a colored symbol, assign it to the corresponding partition (this
            // ensures requirement of a LOCAL splitting function in the words of the thesis):
            // - this is the case if either the conclusion contains a colored symbol
            if (current->getColor() == COLOR_LEFT)
            {
                number_red++;
                continue;
            }
            else if (current->getColor() == COLOR_RIGHT)
            {
                number_blue++;
                continue;
            }

            // - or if any premise contains a colored symbol
            Color containedColor = COLOR_TRANSPARENT;
            VirtualIterator<Unit*> parents = current->getParents();
            while (parents.hasNext())
            {
                Unit* premise= parents.next();

                if (premise->getColor() == COLOR_LEFT || premise->getColor() == COLOR_RIGHT)
                {
                    containedColor = premise->getColor();
                    break;
                }
            }
            if (containedColor == COLOR_LEFT)
            {
                number_red++;
                continue;
            }
            else if (containedColor == COLOR_RIGHT)
            {
                number_blue++;
                continue;
            }

            number_grey++;
        }

        cout << "Proof contains " << number_red << " / " << number_blue << " / " << number_grey << " inferences (red/blue/grey)" << endl;
        cout << "Percentage of grey inferences: " << static_cast<double>(number_grey) / (number_red + number_blue + number_grey) << endl;
    }

}

#endif // VZ3