org.chocosolver.solver.search.loop.move.MoveBinaryHBFS 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.search.loop.move;
import org.chocosolver.solver.Model;
import org.chocosolver.solver.ResolutionPolicy;
import org.chocosolver.solver.Solver;
import org.chocosolver.solver.objective.IObjectiveManager;
import org.chocosolver.solver.search.limits.BacktrackCounter;
import org.chocosolver.solver.search.strategy.decision.Decision;
import org.chocosolver.solver.search.strategy.decision.DecisionPath;
import org.chocosolver.solver.search.strategy.strategy.AbstractStrategy;
import org.chocosolver.solver.variables.IntVar;
import org.chocosolver.util.tools.ArrayUtils;
import java.util.ArrayList;
import java.util.Collections;
import java.util.List;
import java.util.PriorityQueue;
/**
* A move dedicated to run an Hybrid Best-First Search[1] (HBFS) with binary decisions.
*
* [1]:D. Allouche, S. de Givry, G. Katsirelos, T. Schiex, M. Zytnicki,
* Anytime Hybrid Best-First Search with Tree Decomposition for Weighted CSP, CP-2015.
*
* It restarts anytime a backtrack limit is reached and a new open right branch needs to be selected.
*
* Created by cprudhom on 02/11/2015.
* Project: choco.
* @author Charles Prud'homme
* @since 02/11/2015.
*/
public class MoveBinaryHBFS extends MoveBinaryDFS {
/**
* limited number of backtracks for each DFS try
*/
private BacktrackCounter dfslimit;
/**
* limit of bracktracks for the next DFS try.
*/
private long Z;
/**
* as limit is globally maintained, limit += Z at each try.
*/
private long limit;
/**
* maximum number of backtracks to not exceed when updating node recomputation parameters.
*/
private long N;
/**
* for node recomputation.
*/
private long nodesRecompute;
/**
* lower bound to limit the rate of redundantly propagated decisions.
*/
private double a;
/**
* upper bound to limit the rate of redundantly propagated decisions.
*/
private double b;
/**
* The current objective manager, to deal with best bounds.
*/
private IObjectiveManager objectiveManager;
/**
* Indicates if the current resolution policy is minimization.
*/
private boolean isMinimization;
/**
* list of open right branches.
*/
private PriorityQueue opens;
/**
* Current open right branch.
*/
private Decision[] copen;
/**
* Used to find the first unknown open right branch
*/
private List _unkopen;
/**
* Current decision in copen
*/
private int current;
/**
* The owner model.
*/
private Model mModel;
/**
* Create a move dedicated to run an Hybrid Best-First Search[1] (HBFS) with binary decisions.
* @param model a model
* @param strategy the search strategy to use
* @param a lower bound to limit the rate of redundantly propagated decisions.
* @param b upper bound to limit the rate of redundantly propagated decisions.
* @param N maximum number of backtracks to not exceed when updating node recomputation parameters.
*/
public MoveBinaryHBFS(Model model, AbstractStrategy strategy, double a, double b, long N) {
super(strategy);
this.mModel = model;
this.dfslimit = new BacktrackCounter(model, N);
this.opens = new PriorityQueue<>();
this.copen = new Decision[0];
this.current = 0;
this.Z = 1;
this.limit = Z;
this.N = N;
this.a = a;
this.b = b;
this._unkopen = new ArrayList<>();
}
@Override
public boolean init() {
boolean init = super.init();
this.objectiveManager = mModel.getSolver().getObjectiveManager();
if (objectiveManager.getPolicy() == ResolutionPolicy.SATISFACTION) {
throw new UnsupportedOperationException("HBFS is not adapted to satisfaction problems.");
}
isMinimization = objectiveManager.getPolicy() == ResolutionPolicy.MINIMIZE;
return init;
}
@Override
public boolean extend(Solver solver) {
boolean extend;
// as we observe the number of backtracks, no limit can be reached on extend()
if (current < copen.length) {
solver.getDecisionPath().pushDecision(copen[current++]);
solver.getEnvironment().worldPush();
extend = true;
} else /*cut will checker with propagation */ {
extend = super.extend(solver);
}
return extend;
}
@Override
public boolean repair(Solver solver) {
boolean repair;
if (!dfslimit.isMet(limit)) {
current = copen.length;
repair = super.repair(solver);
} else {
extractOpenRightBranches(solver);
repair = true;
}
return repair;
}
/**
* This methods extracts and stores all open right branches for future exploration
* @param solver reference to the solver
*/
protected void extractOpenRightBranches(Solver solver) {
// update parameters for restarts
if (nodesRecompute > 0) {
double ratio = nodesRecompute * 1.d / solver.getNodeCount();
if (ratio > b && Z <= N) {
Z *= 2;
} else if (ratio < a && Z >= 2) {
Z /= 2;
}
}
limit += Z;
// then start the extraction of open right branches
int i = compareSubpath(solver);
if(i < _unkopen.size()) {
extractOB(solver, i);
}
// finally, get the best ORB to keep up the search
Open next = opens.poll();
while (next != null && !isValid(next.currentBound())) {
next = opens.poll();
}
if (next != null) {
copen = next.toArray();
// the decision in 0 is the last taken, then the array us reversed
ArrayUtils.reverse(copen);
current = 0;
nodesRecompute = solver.getNodeCount() + copen.length;
} else{
// to be sure not to use the previous path
current = copen.length;
}
// then do the restart
solver.restart();
}
/**
* Copy the current decision path in _unkopen, for comparison with copen.
* Then, it compares each decision, from the top to the bottom, to find the first difference.
* This is required to avoid adding the same decision sub-path more than once
* @param solver the search loop
* @return the index of the decision, in _unkopen, that stops the loop
*/
private int compareSubpath(Solver solver) {
_unkopen.clear();
DecisionPath decisionPath = solver.getDecisionPath();
int pos = decisionPath.size() - 1;
Decision decision = decisionPath.getDecision(pos);
while (decision.getPosition() != topDecisionPosition) {
_unkopen.add(decision);
decision = decisionPath.getDecision(--pos);
}
Collections.reverse(_unkopen);
//
int i = 0;
int I = Math.min(_unkopen.size(), copen.length);
while(i < I && copen[i].isEquivalentTo(_unkopen.get(i))){
i++;
}
return i;
}
/**
* Extract the open right branches from the current path until it reaches the i^th decision of _unkopen
* @param solver the search loop
* @param i the index of the decision, in _unkopen, that stops the loop
*/
private void extractOB(Solver solver, int i) {
// Decision stopAt = solver.getDecisionPath().getDecision(_unkopen.get(i).getPosition()-1);
int stopAt = _unkopen.get(i).getPosition()-1;
// then, goes up in the search tree, and detect open nodes
solver.getEnvironment().worldPop();
DecisionPath dp = solver.getDecisionPath();
int bound;
Decision decision = dp.getLastDecision();
while (decision.getPosition() != stopAt) {
bound = isMinimization ?
objectiveManager.getObjective().getLB() :
objectiveManager.getObjective().getUB();
if (decision.hasNext() && isValid(bound)) {
opens.add(new Open(decision, dp, bound, isMinimization));
}
dp.synchronize();
decision = dp.getLastDecision();
solver.getEnvironment().worldPop();
}
}
/**
* If the bound of an O.R.B exceed the best known so far, it returns false.
* @param bound the current bound of an O.R.B.
* @return true if bound is valid wrt the best known so far.
*/
private boolean isValid(int bound) {
return isMinimization ?
bound < objectiveManager.getBestUB().intValue() :
bound > objectiveManager.getBestLB().intValue();
}
/**
* A class to represent an open right branch, from which the search can be kept up.
*/
private class Open implements Comparable {
/**
* List of open decisions
*/
private List path;
/**
* store the current lower bound of the decision path for minimization
*/
private int currentBound;
/**
* 1 for minimization, -1 for maximization
*/
private byte minimization;
/**
* Create an open right branch for HBFS
*
* @param decision an open decision in decisionPath
* @param decisionPath the current decision path
* @param currentBound current lower (resp. upper) bound of the objective value for mimimization (resp. maximization)
* @param minimization set to true for minimization
*/
public Open(Decision decision, DecisionPath decisionPath, int currentBound, boolean minimization) {
this.path = new ArrayList<>();
while (decision.getPosition() != topDecisionPosition) {
Decision d = decision.duplicate();
while (decision.triesLeft() != d.triesLeft() - 1) {
d.buildNext();
}
path.add(d);
decision = decisionPath.getDecision(decision.getPosition() -1);
}
this.currentBound = currentBound;
this.minimization = (byte) (minimization ? 1 : -1);
}
/**
* Return the current decision path that can be extended
* @return an array of decisions
*/
public Decision[] toArray() {
return path.toArray(new Decision[0]);
}
/**
* Return the current bound of this open right branch
* @return the current bound of this
*/
public int currentBound() {
return currentBound;
}
@Override
public int compareTo(Open o) {
// the minimum lower bound
int clb = minimization * (currentBound - o.currentBound);
if (clb == 0) {
// the maximum depth
return (o.path.size() - path.size());
} else {
return clb;
}
}
@Override
public String toString() {
StringBuilder st = new StringBuilder();
st.append('[').append(currentBound).append(']');
for(int i = path.size() - 1;i > -1 ; i--){
st.append(path.get(i)).append(',');
}
return st.toString();
}
}
}