org.chocosolver.solver.ParallelPortfolio Maven / Gradle / Ivy
Show all versions of choco-solver Show documentation
/*
* This file is part of choco-solver, http://choco-solver.org/
*
* Copyright (c) 2020, IMT Atlantique. All rights reserved.
*
* Licensed under the BSD 4-clause license.
*
* See LICENSE file in the project root for full license information.
*/
package org.chocosolver.solver;
import org.chocosolver.cutoffseq.LubyCutoffStrategy;
import org.chocosolver.solver.constraints.Constraint;
import org.chocosolver.solver.constraints.nary.sat.NogoodStealer;
import org.chocosolver.solver.constraints.real.RealConstraint;
import org.chocosolver.solver.exception.SolverException;
import org.chocosolver.solver.search.limits.FailCounter;
import org.chocosolver.solver.search.loop.lns.INeighborFactory;
import org.chocosolver.solver.search.loop.monitors.IMonitorSolution;
import org.chocosolver.solver.search.loop.monitors.NogoodFromRestarts;
import org.chocosolver.solver.search.strategy.Search;
import org.chocosolver.solver.variables.IntVar;
import org.chocosolver.solver.variables.RealVar;
import org.chocosolver.solver.variables.SetVar;
import org.chocosolver.solver.variables.Variable;
import java.util.ArrayList;
import java.util.Arrays;
import java.util.List;
import java.util.Spliterator;
import java.util.concurrent.ExecutionException;
import java.util.concurrent.ForkJoinPool;
import java.util.concurrent.atomic.AtomicBoolean;
import java.util.function.Consumer;
import java.util.stream.Stream;
import java.util.stream.StreamSupport;
import static org.chocosolver.solver.search.strategy.Search.*;
/**
*
*
* A Portfolio helper.
*
*
* The ParallelPortfolio resolution of a problem is made of four steps:
*
* - adding models to be run in parallel,
* - running resolution in parallel,
* - getting the model which finds a solution (or the best one), if any.
*
* Each of the four steps is needed and the order is imposed too.
* In particular, in step 1. each model should be populated individually with a model of the problem
* (presumably the same model, but not required).
* Populating model is not managed by this class and should be done before applying step 2.,
* with a dedicated method for instance.
*
* Note also that there should not be pending resolution process in any models.
* Otherwise, unexpected behaviors may occur.
*
*
* The resolution process is synchronized. As soon as one model ends (naturally or by hitting a limit)
* the other ones are eagerly stopped.
* Moreover, when dealing with an optimization problem, cut on the objective variable's value is propagated
* to all models on solution.
* It is essential to eagerly declare the objective variable(s) with {@link Model#setObjective(boolean, Variable)}.
*
*
*
* Note that the similarity of the models declared is not required.
* However, when dealing with an optimization problem, keep in mind that the cut on the objective variable's value
* is propagated among all models, so different objectives may lead to wrong results.
*
*
* Since there is no condition on the similarity of the models,
* once the resolution ends, the model which finds the (best) solution is internally stored.
*
*
* Example of use.
*
*
* ParallelPortfolio pares = new ParallelPortfolio();
* int n = 4; // number of models to use
* for (int i = 0; i < n; i++) {
* pares.addModel(modeller());
* }
* pares.solve();
* IOutputFactory.printSolutions(pares.getBestModel());
*
*
*
*
*
* This class uses Java 8 streaming feature, and may be not compliant with older versions.
*
*
*
*
* Project: choco.
* @author Charles Prud'homme, Jean-Guillaume Fages
* @since 23/12/2015.
*/
public class ParallelPortfolio {
////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
/////////////////////////////////////// VARIABLES //////////////////////////////////////////////////////
////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
/** List of {@link Model}s to be executed in parallel. */
private final List models;
/** whether or not to use default search configurations for the different threads **/
private boolean searchAutoConf;
/** This manager is used to synchronize nogood sharing.*/
private NogoodStealer manager = NogoodStealer.NONE;
/** Stores whether or not prepare() method has been called */
private boolean isPrepared = false;
private AtomicBoolean solverTerminated = new AtomicBoolean(false);
private AtomicBoolean solutionFound = new AtomicBoolean(false);
/** Point to (one of) the solver(s) which found a solution */
private Model finder;
////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
/////////////////////////////////////// CONSTRUCTOR //////////////////////////////////////////////////////
////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
/**
* Creates a new ParallelPortfolio
* This class stores the models to be executed in parallel in a {@link ArrayList} initially empty.
*
* @param searchAutoConf changes the search heuristics of the different solvers, except the first one (true by default).
* Must be set to false if search heuristics of the different threads are specified manually, so that they are not erased
*/
public ParallelPortfolio(boolean searchAutoConf) {
this.models = new ArrayList<>();
this.searchAutoConf = searchAutoConf;
}
/**
* Creates a new ParallelPortfolio
* This class stores the models to be executed in parallel in a {@link ArrayList} initially empty.
* Search heuristics will be changed automatically (except for the first thread that will remain in the same configuration).
*/
public ParallelPortfolio() {
this(true);
}
////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
/////////////////////////////////////// API //////////////////////////////////////////////////////
////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
/**
* Calling this method will ensure that workers equipped with a restart policy not only
* record nogoods from themselves (based on {@link NogoodFromRestarts}) but also based on
* other workers of the portfolio.
* @implSpec
* It is assumed that all models in this portfolio are equivalent (ie, each variable has
* the same ID in each worker).
*/
public void stealNogoodsOnRestarts() {
this.manager = new NogoodStealer();
}
/**
*
* Adds a model to the list of models to run in parallel.
* The model can either be a fresh one, ready for populating, or a populated one.
*
*
* Important:
*
* - the populating process is not managed by this ParallelPortfolio
* and should be done externally, with a dedicated method for example.
*
* -
* when dealing with optimization problems, the objective variables HAVE to be declared eagerly with
* {@link Model#setObjective(boolean, Variable)}.
*
*
*
*
* @param model a model to add
*/
public void addModel(Model model){
this.models.add(model);
}
/**
* Run the solve() instruction of every model of the portfolio in parallel.
*
*
* Note that a call to {@link #getBestModel()} returns a model which has found the best solution.
*
* @return true if and only if at least one new solution has been found.
* @throws SolverException if no model or only model has been added.
*/
public boolean solve() {
getSolverTerminated().set(false);
getSolutionFound().set(false);
if (!isPrepared) {
prepare();
}
ForkJoinPool forkJoinPool = new ForkJoinPool(models.size());
try {
forkJoinPool.submit(() -> models.parallelStream().forEach(m -> {
if (!getSolverTerminated().get()) {
boolean so = m.getSolver().solve();
if (!so || finder == m) {
getSolverTerminated().set(true);
}
}
})).get();
} catch (InterruptedException | ExecutionException | SolverException e) {
e.printStackTrace();
}
forkJoinPool.shutdownNow();
getSolverTerminated().set(false);// otherwise, solver.isStopCriterionMet() always returns true
if(getSolutionFound().get() && models.get(0).getResolutionPolicy()!=ResolutionPolicy.SATISFACTION) {
int bestAll = getBestModel().getSolver().getBestSolutionValue().intValue();
for (Model m : models) {
int mVal = m.getSolver().getBestSolutionValue().intValue();
if (m.getResolutionPolicy() == ResolutionPolicy.MAXIMIZE) {
assert mVal <= bestAll : mVal + " > " + bestAll;
} else
assert m.getResolutionPolicy() != ResolutionPolicy.MINIMIZE || mVal >= bestAll : mVal + " < " + bestAll;
}
}
return getSolutionFound().get();
}
/**
* Returns the first model from the list which, either :
*
* -
* finds a solution when dealing with a satisfaction problem,
*
* -
* or finds (and possibly proves) the best solution when dealing with an optimization problem.
*
*
* or null if no such model exists.
* Note that there can be more than one "finder" in the list, yet, this method returns the index of the first one.
*
* @return the first model which finds a solution (or the best one) or null if no such model exists.
*/
public Model getBestModel(){
return finder;
}
/**
* @return the (mutable!) list of models used in this ParallelPortfolio
*/
public List getModels(){
return models;
}
/**
* Attempts to find all solutions of the declared problem.
*
* - If the method returns an empty list:
*
* - either a stop criterion (e.g., a time limit) stops the search before any solution has been found,
* - or no solution exists for the problem (i.e., over-constrained).
*
* - if the method returns a list with at least one element in it:
*
* - either the resolution stops eagerly du to a stop criterion before finding all solutions,
* - or all solutions have been found.
*
*
*
*
* Note that all variables will be recorded
*
* @return a list that contained the found solutions.
*/
public Stream streamSolutions() {
Spliterator it = new Spliterator() {
@Override
public boolean tryAdvance(Consumer action) {
if (solve()) {
action.accept(new Solution(getBestModel()).record());
return true;
}
return false;
}
@Override
public Spliterator trySplit() {
return null;
}
@Override
public long estimateSize() {
return Long.MAX_VALUE;
}
@Override
public int characteristics() {
return Spliterator.ORDERED | Spliterator.DISTINCT | Spliterator.NONNULL | Spliterator.CONCURRENT;
}
};
return StreamSupport.stream(it, false);
}
////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
/////////////////////////////////////// INTERNAL METHODS //////////////////////////////////////////////////////
////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
@SuppressWarnings("unchecked")
public void prepare(){
isPrepared = true;
check();
for(int i=0;i getSolverTerminated().get());
s.plugMonitor((IMonitorSolution) () -> updateFromSolution(s.getModel()));
if(searchAutoConf){
configureModel(i);
}
}
}
private synchronized void updateFromSolution(Model m){
if (m.getResolutionPolicy() == ResolutionPolicy.SATISFACTION) {
finder = m;
getSolutionFound().set(true);
}else{
int solverVal = ((IntVar)m.getObjective()).getValue();
int bestVal = m.getSolver().getObjectiveManager().getBestSolutionValue().intValue();
if(m.getResolutionPolicy()==ResolutionPolicy.MAXIMIZE){
assert solverVal<=bestVal:solverVal+">"+bestVal;
}else
assert
m.getResolutionPolicy() != ResolutionPolicy.MINIMIZE || solverVal >= bestVal :solverVal+"<"+bestVal;
if(solverVal == bestVal){
getSolutionFound().set(true);
finder = m;
if (m.getResolutionPolicy() == ResolutionPolicy.MAXIMIZE) {
models.forEach(s1 -> s1.getSolver().getObjectiveManager().updateBestLB(bestVal));
}else {
models.forEach(s1 -> s1.getSolver().getObjectiveManager().updateBestUB(bestVal));
}
}
}
}
private void configureModel(int workerID) {
Model worker = getModels().get(workerID);
Solver solver = worker.getSolver();
ResolutionPolicy policy = worker.getResolutionPolicy();
// compute decision variables
Variable[] varsX;
if (solver.getSearch() != null && solver.getSearch().getVariables().length > 0) {
varsX = solver.getSearch().getVariables();
}else{
varsX = worker.getVars();
}
IntVar[] ivars = new IntVar[varsX.length];
SetVar[] svars = new SetVar[varsX.length];
RealVar[] rvars = new RealVar[varsX.length];
int ki=0,ks=0,kr=0;
for (Variable aVarsX : varsX) {
if ((aVarsX.getTypeAndKind() & Variable.INT) > 0) {
ivars[ki++] = (IntVar) aVarsX;
} else if ((aVarsX.getTypeAndKind() & Variable.SET) > 0) {
svars[ks++] = (SetVar) aVarsX;
} else if ((aVarsX.getTypeAndKind() & Variable.REAL) > 0) {
rvars[kr++] = (RealVar) aVarsX;
} else {
throw new UnsupportedOperationException("unrecognized variable kind " + aVarsX);
}
}
ivars = Arrays.copyOf(ivars, ki);
svars = Arrays.copyOf(svars, ks);
rvars = Arrays.copyOf(rvars, kr);
// set heuristic
boolean opt = policy != ResolutionPolicy.SATISFACTION;
switch (workerID) {
case 0:
// DWD + fast restart + LC (+ B2V)
solver.setSearch(
lastConflict(
VarH.DOMWDEG.make(solver, ivars, ValH.BEST, opt)
)
);
Restarts.LUBY.declare(solver, 500, 0.d,5000);
manager.add(worker);
break;
case 1:
solver.setSearch(
lastConflict(
VarH.CHS.make(solver, ivars, ValH.MIN, opt)
)
);
Restarts.LUBY.declare(solver, 500, 0.d, 5000);
manager.add(worker);
break;
case 2:
// input order + LC
solver.setSearch(
lastConflict(
VarH.INPUT.make(solver, ivars, ValH.MIN, opt)
)
);
manager.add(worker);
break;
case 3:
if(!opt) {
solver.setSearch(
lastConflict(
VarH.DOMWDEGR.make(solver, ivars, ValH.MIN, opt)
)
);
Restarts.LUBY.declare(solver, 500, 0.d, 5000);
break;
}else{
// input order + LC + LNS
solver.setSearch(
lastConflict(
VarH.INPUT.make(solver, ivars, ValH.MIN, opt)
)
);
solver.setLNS(INeighborFactory.blackBox(ivars), new FailCounter(solver.getModel(), 1000));
}
manager.add(worker);
break;
case 4:
// ABS + fast restart + LC
solver.setSearch(
lastConflict(
VarH.ABS.make(solver, ivars, ValH.DEFAULT, opt)
)
);
Restarts.LUBY.declare(solver, 500, 0.d, 5000);
break;
case 5:
// DWD + fast restart + COS
solver.setSearch(
Search.conflictOrderingSearch(
VarH.DOMWDEG.make(solver, ivars, ValH.MIN, opt)
)
);
Restarts.LUBY.declare(solver, 500, 0.d, 5000);
solver.setSearch(lastConflict(solver.getSearch()));
break;
case 6:
// DWD + fast restart + LC (+ B2V)
if (!opt) {
solver.setSearch(VarH.DOMWDEG.make(solver, ivars, ValH.MIN, false));
} else {
solver.setSearch(
lastConflict(
VarH.DOMWDEGR.make(solver, ivars, ValH.MIN, opt)
)
);
Restarts.LUBY.declare(solver, 500, 0.d, 5000);
}
break;
case 7:
solver.setSearch(
lastConflict(
VarH.CHS.make(solver, ivars, ValH.MIN, opt)
)
);
Restarts.LUBY.declare(solver, 40, 0.d, 5000);
manager.add(worker);
break;
default:
// random search (various seeds) + LNS if optim
solver.setSearch(lastConflict(randomSearch(ivars,workerID)));
if(policy!=ResolutionPolicy.SATISFACTION){
solver.setLNS(INeighborFactory.blackBox(ivars), new FailCounter(solver.getModel(), 1000));
}
solver.plugMonitor(new NogoodFromRestarts(worker, manager));
solver.setRestarts(count -> solver.getFailCount() >= count, new LubyCutoffStrategy(500), 5000);
break;
}
// complete with set default search
if(ks>0) {
solver.setSearch(solver.getSearch(),setVarSearch(svars));
}
// complete with real default search
if(kr>0) {
solver.setSearch(solver.getSearch(),realVarSearch(rvars));
}
}
private void check(){
if (models.size() == 0) {
throw new SolverException("No model found in the ParallelPortfolio.");
}
if(models.get(0).getResolutionPolicy() != ResolutionPolicy.SATISFACTION) {
Variable objective = models.get(0).getObjective();
if (objective == null) {
throw new UnsupportedOperationException("No objective has been defined");
}
if ((objective.getTypeAndKind() & Variable.REAL) != 0) {
for(Constraint c : models.get(0).getCstrs()){
if(c instanceof RealConstraint){
throw new UnsupportedOperationException("" +
"Ibex is not multithread safe, ParallelPortfolio cannot be used");
}
}
}
}
}
private synchronized AtomicBoolean getSolverTerminated(){
return solverTerminated;
}
private synchronized AtomicBoolean getSolutionFound(){
return solutionFound;
}
}