001 package ifs.heuristics;
002
003 import ifs.extension.*;
004 import ifs.model.*;
005 import ifs.solution.*;
006 import ifs.solver.*;
007 import ifs.util.*;
008
009 import java.util.*;
010
011 /**
012 * General implementation of value selection criterion.
013 * <br><br>
014 * Value selection criterion is based on weighted sum of various criteria. It also allows random walk technique and
015 * tabu search.
016 * <br>
017 * Parameters:
018 * <br>
019 * <table border='1'><tr><th>Parameter</th><th>Type</th><th>Comment</th></tr>
020 * <tr><td>General.MPP</td><td>{@link Boolean}</td><td>if true, MPP is being solved</td></tr>
021 * <tr><td>Value.MPPLimit</td><td>{@link Integer}</td><td>MPP: limitation of the number of allowed perturbations. If a solution within this limit is gound, it is decreased.</td></tr>
022 * <tr><td>Value.InitialSelectionProb</td><td>{@link Double}</td><td>MPP: probability of selection of the initial value</td></tr>
023 * <tr><td>Value.RandomWalkProb</td><td>{@link Double}</td><td>Random Walk: probability of selection of a value randomly among all the values</td></tr>
024 * <tr><td>Value.Tabu</td><td>{@link Integer}</td><td>Tabu Search: length of the tabu-list</td></tr>
025 * <tr><td>Value.GoodSelectionProb</td><td>{@link Double}</td><td>In case of {@link MacPropagation}, with this probability (1.0 means always), the selection is made only among good values (not removed from the domain).</td></tr>
026 * </table>
027 * <br>
028 * Following weights are used in the weighted sum (computed for all values). The value with the lowest weighted sum is selected.
029 * If there are more than one of such values, one of them is selected randomly.
030 * <br>
031 * <table border='1'><tr><th>Parameter</th><th>Type</th><th>Comment</th></tr>
032 * <tr><td>Value.WeightDeltaInitialAssignments</td><td>{@link Double}</td><td>MPP: Difference in the number of assigned initial values if the value is assigned to the variable (weighted by this Value.WeightDeltaInitialAssignments): -1 if the value is initial, 0 otherwise, increased by the number of initial values assigned to variables with hard conflicts with the value</td></tr>
033 * <tr><td>Value.WeightWeightedConflicts</td><td>{@link Double}</td><td>When {@link ConflictStatistics} is used: weighted number of conflicting variables</td></tr>
034 * <tr><td>Value.WeightPotentialConflicts</td><td>{@link Double}</td><td>When {@link ConflictStatistics} is used: weighted number of potentially conflicting variables</td></tr>
035 * <tr><td>Value.WeightConflicts</td><td>{@link Double}</td><td>Number of conflicting variables {@link Model#conflictValues(Value)}.</td></tr>
036 * <tr><td>Value.WeightNrAssignments</td><td>{@link Double}</td><td>Number of previous assignments of the value</td></tr>
037 * <tr><td>Value.WeightValue</td><td>{@link Double}</td><td>Value {@link Value#toInt()}</td></tr>
038 * </table>
039 *
040 * @see VariableSelection
041 * @see Solver
042 *
043 * @author <a href="mailto:muller@ktiml.mff.cuni.cz">Tomáš Müller</a>
044 * @version 1.0
045 **/
046
047 public class GeneralValueSelection implements ValueSelection {
048 private static org.apache.log4j.Logger sLogger = org.apache.log4j.Logger.getLogger(GeneralValueSelection.class);
049 private double iRandomWalkProb = 0.0;
050 private double iInitialSelectionProb = 0.0;
051 private double iGoodSelectionProb = 0.0;
052 private int iMPPLimit = -1;
053
054 private double iWeightDeltaInitialAssignment = 0.0;
055 private double iWeightPotentialConflicts = 0.0;
056 private double iWeightWeightedCoflicts = 0.0;
057 private double iWeightCoflicts = 1.0;
058 private double iWeightNrAssignments = 0.5;
059 private double iWeightValue = 0.0;
060
061 private int iTabuSize = 0;
062 private ArrayList iTabu = null;
063 private int iTabuPos = 0;
064
065 private boolean iMPP = false;
066 private ConflictStatistics iStat = null;
067 private MacPropagation iProp = null;
068 private ViolatedInitials iViolatedInitials = null;
069
070 public GeneralValueSelection() {}
071
072 /** Constructor
073 * @param properties input configuration
074 */
075 public GeneralValueSelection(DataProperties properties) {
076 iMPP = properties.getPropertyBoolean("General.MPP", false);
077 if (iMPP) {
078 iMPPLimit = properties.getPropertyInt("Value.MPPLimit", -1);
079 iInitialSelectionProb = properties.getPropertyDouble("Value.InitialSelectionProb", 0.75);
080 iWeightDeltaInitialAssignment = properties.getPropertyDouble("Value.WeightDeltaInitialAssignments", 0.0);
081 }
082 iGoodSelectionProb = properties.getPropertyDouble("Value.GoodSelectionProb", 0.00);
083 iWeightWeightedCoflicts = properties.getPropertyDouble("Value.WeightWeightedConflicts", 1.0);
084 iWeightPotentialConflicts = properties.getPropertyDouble("Value.WeightPotentialConflicts", 0.0);
085
086 iRandomWalkProb = properties.getPropertyDouble("Value.RandomWalkProb", 0.0);
087 iWeightCoflicts = properties.getPropertyDouble("Value.WeightConflicts", 1.0);
088 iWeightNrAssignments = properties.getPropertyDouble("Value.WeightNrAssignments", 0.5);
089 iWeightValue = properties.getPropertyDouble("Value.WeightValue", 0.0);
090 iTabuSize = properties.getPropertyInt("Value.Tabu", 0);
091 if (iTabuSize > 0)
092 iTabu = new ArrayList(iTabuSize);
093 }
094
095 /** Initialization */
096 public void init(Solver solver) {
097 for (Enumeration i = solver.getExtensions().elements(); i.hasMoreElements();) {
098 Extension extension = (Extension)i.nextElement();
099 if (extension instanceof ConflictStatistics)
100 iStat = (ConflictStatistics)extension;
101 if (extension instanceof MacPropagation)
102 iProp = (MacPropagation)extension;
103 if (extension instanceof ViolatedInitials)
104 iViolatedInitials = (ViolatedInitials)extension;
105 }
106 }
107
108 /** Value selecion */
109 public Value selectValue(Solution solution, Variable selectedVariable) {
110 if (iMPP) {
111 if (selectedVariable.getInitialAssignment() != null) {
112 if (solution.getModel().unassignedVariables().isEmpty()) {
113 if (solution.getModel().perturbVariables().size() <= iMPPLimit)
114 iMPPLimit = solution.getModel().perturbVariables().size() - 1;
115 }
116 if (iMPPLimit >= 0 && solution.getModel().perturbVariables().size() > iMPPLimit)
117 return selectedVariable.getInitialAssignment();
118 if (selectedVariable.getInitialAssignment() != null && ToolBox.random() <= iInitialSelectionProb)
119 return selectedVariable.getInitialAssignment();
120 }
121 }
122
123 Vector values = selectedVariable.values();
124 if (ToolBox.random() <= iRandomWalkProb) return (Value)ToolBox.random(values);
125 if (iProp != null && selectedVariable.getAssignment() == null && ToolBox.random() <= iGoodSelectionProb) {
126 Collection goodValues = iProp.goodValues(selectedVariable);
127 if (!goodValues.isEmpty()) values = new FastVector(goodValues);
128 }
129 if (values.size() == 1)
130 return (Value)values.firstElement();
131
132 Vector bestValues = null;
133 double bestWeightedSum = 0;
134
135 for (Enumeration i1 = values.elements(); i1.hasMoreElements();) {
136 Value value = (Value)i1.nextElement();
137 if (iTabu != null && iTabu.contains(value)) continue;
138 if (selectedVariable.getAssignment() != null && selectedVariable.getAssignment().equals(value))
139 continue;
140
141 Collection conf = solution.getModel().conflictValues(value);
142 double weightedConflicts = (iStat == null || iWeightCoflicts == 0.0 ? 0.0 : iStat.countRemovals(solution.getIteration(), conf, value));
143 double potentialConflicts = (iStat == null || iWeightPotentialConflicts == 0.0 ? 0.0 : iStat.countPotentialConflicts(solution.getIteration(), value, 3));
144
145 long deltaInitialAssignments = 0;
146 if (iMPP && iWeightDeltaInitialAssignment != 0.0) {
147 if (iViolatedInitials != null) {
148 Set violations = iViolatedInitials.getViolatedInitials(value);
149 if (violations != null) {
150 for (Iterator it1 = violations.iterator(); it1.hasNext();) {
151 Value aValue = (Value)it1.next();
152 if (aValue.variable().getAssignment() == null || aValue.variable().getAssignment().equals( aValue))
153 deltaInitialAssignments += 2;
154 }
155 }
156 }
157 for (Iterator it1 = conf.iterator(); it1.hasNext();) {
158 Value aValue = (Value)it1.next();
159 if (aValue.variable().getInitialAssignment() != null)
160 deltaInitialAssignments--;
161 }
162 if (selectedVariable.getInitialAssignment() != null && !selectedVariable.getInitialAssignment().equals(value)) {
163 deltaInitialAssignments++;
164 }
165 if (iMPPLimit >= 0 && (solution.getModel().perturbVariables().size() + deltaInitialAssignments) > iMPPLimit)
166 continue;
167 }
168
169 double weightedSum =
170 (iWeightDeltaInitialAssignment * deltaInitialAssignments)
171 + (iWeightPotentialConflicts * potentialConflicts)
172 + (iWeightWeightedCoflicts * weightedConflicts)
173 + (iWeightCoflicts * conf.size())
174 + (iWeightNrAssignments * value.countAssignments())
175 + (iWeightValue * value.toInt());
176
177 if (bestValues == null || bestWeightedSum > weightedSum) {
178 bestWeightedSum = weightedSum;
179 if (bestValues == null) bestValues = new FastVector();
180 else bestValues.clear();
181 bestValues.addElement(value);
182 } else {
183 if (bestWeightedSum == weightedSum)
184 bestValues.addElement(value);
185 }
186 }
187
188 Value selectedValue = (Value)(bestValues==null?null:ToolBox.random(bestValues));
189 if (selectedValue == null) selectedValue = (Value)ToolBox.random(values);
190 if (iTabu != null) {
191 if (iTabu.size() == iTabuPos)
192 iTabu.add(selectedValue);
193 else
194 iTabu.set(iTabuPos, selectedValue);
195 iTabuPos = (iTabuPos + 1) % iTabuSize;
196 }
197 return (bestValues == null ? null : selectedValue);
198 }
199
200 }