[6152] | 1 | /* |
---|
| 2 | Copyright 2006 by Sean Luke |
---|
| 3 | Licensed under the Academic Free License version 3.0 |
---|
| 4 | See the file "LICENSE" for more information |
---|
| 5 | */ |
---|
| 6 | |
---|
| 7 | |
---|
| 8 | package ec.select; |
---|
| 9 | import ec.*; |
---|
| 10 | import ec.util.*; |
---|
| 11 | |
---|
| 12 | /* |
---|
| 13 | * GreedyOverselection.java |
---|
| 14 | * |
---|
| 15 | * Created: Thu Feb 10 17:39:03 2000 |
---|
| 16 | * By: Sean Luke |
---|
| 17 | */ |
---|
| 18 | |
---|
| 19 | /** |
---|
| 20 | * GreedyOverselection is a SelectionMethod which implements Koza-style |
---|
| 21 | * fitness-proportionate greedy overselection. Not appropriate for |
---|
| 22 | * multiobjective fitnesses. |
---|
| 23 | * |
---|
| 24 | * <p> This selection method first |
---|
| 25 | * divides individuals in a population into two groups: the "good" |
---|
| 26 | * ("top") group, and the "bad" ("bottom") group. The best <i>top</i> |
---|
| 27 | * percent of individuals in the population go into the good group. |
---|
| 28 | * The rest go into the "bad" group. With a certain probability (determined |
---|
| 29 | * by the <i>gets</i> setting), an individual will be picked out of the |
---|
| 30 | * "good" group. Once we have determined which group the individual |
---|
| 31 | * will be selected from, the individual is picked using fitness proportionate |
---|
| 32 | * selection in that group, that is, the likelihood he is picked is |
---|
| 33 | * proportionate to his fitness relative to the fitnesses of others in his |
---|
| 34 | * group. |
---|
| 35 | * |
---|
| 36 | * <p> All this is expensive to |
---|
| 37 | * set up and bring down, so it's not appropriate for steady-state evolution. |
---|
| 38 | * If you're not familiar with the relative advantages of |
---|
| 39 | * selection methods and just want a good one, |
---|
| 40 | * use TournamentSelection instead. |
---|
| 41 | * |
---|
| 42 | * <p><b><font color=red> |
---|
| 43 | * Note: Fitnesses must be non-negative. 0 is assumed to be the worst fitness. |
---|
| 44 | * </font></b> |
---|
| 45 | |
---|
| 46 | <p><b>Typical Number of Individuals Produced Per <tt>produce(...)</tt> call</b><br> |
---|
| 47 | Always 1. |
---|
| 48 | |
---|
| 49 | <p><b>Parameters</b><br> |
---|
| 50 | <table> |
---|
| 51 | <tr><td valign=top><i>base.</i><tt>top</tt><br> |
---|
| 52 | <font size=-1>0.0 <= float <= 1.0</font></td> |
---|
| 53 | <td valign=top>(the percentage of the population going into the "good" (top) group)</td></tr> |
---|
| 54 | <tr><td valign=top><i>base.</i><tt>gets</tt><br> |
---|
| 55 | <font size=-1>0.0 <= float <= 1.0</font></td> |
---|
| 56 | <td valign=top>(the likelihood that an individual will be picked from the "good" group)</td></tr> |
---|
| 57 | </table> |
---|
| 58 | |
---|
| 59 | <p><b>Default Base</b><br> |
---|
| 60 | select.greedy |
---|
| 61 | * @author Sean Luke |
---|
| 62 | * @version 1.0 |
---|
| 63 | */ |
---|
| 64 | |
---|
| 65 | public class GreedyOverselection extends SelectionMethod |
---|
| 66 | { |
---|
| 67 | public float[] sortedFitOver; |
---|
| 68 | public float[] sortedFitUnder; |
---|
| 69 | /** Sorted population -- since I *have* to use an int-sized |
---|
| 70 | individual (short gives me only 16K), |
---|
| 71 | I might as well just have pointers to the |
---|
| 72 | population itself. :-( */ |
---|
| 73 | public int[] sortedPop; |
---|
| 74 | |
---|
| 75 | public static final String P_GREEDY = "greedy"; |
---|
| 76 | public static final String P_TOP = "top"; |
---|
| 77 | public static final String P_GETS = "gets"; |
---|
| 78 | |
---|
| 79 | public float top_n_percent; |
---|
| 80 | public float gets_n_percent; |
---|
| 81 | |
---|
| 82 | public Parameter defaultBase() |
---|
| 83 | { |
---|
| 84 | return SelectDefaults.base().push(P_GREEDY); |
---|
| 85 | } |
---|
| 86 | |
---|
| 87 | public void setup(final EvolutionState state, final Parameter base) |
---|
| 88 | { |
---|
| 89 | super.setup(state,base); |
---|
| 90 | |
---|
| 91 | Parameter def = defaultBase(); |
---|
| 92 | |
---|
| 93 | top_n_percent = |
---|
| 94 | state.parameters.getFloatWithMax(base.push(P_TOP),def.push(P_TOP),0.0,1.0); |
---|
| 95 | if (top_n_percent < 0.0) |
---|
| 96 | state.output.fatal("Top-n-percent must be between 0.0 and 1.0", base.push(P_TOP),def.push(P_TOP)); |
---|
| 97 | |
---|
| 98 | gets_n_percent = |
---|
| 99 | state.parameters.getFloatWithMax(base.push(P_GETS),def.push(P_GETS),0.0,1.0); |
---|
| 100 | if (gets_n_percent < 0.0) |
---|
| 101 | state.output.fatal("Gets-n-percent must be between 0.0 and 1.0", base.push(P_GETS),def.push(P_GETS)); |
---|
| 102 | |
---|
| 103 | } |
---|
| 104 | |
---|
| 105 | |
---|
| 106 | // don't need clone etc. -- I'll never clone with my arrays intact |
---|
| 107 | |
---|
| 108 | public void prepareToProduce(final EvolutionState s, |
---|
| 109 | final int subpopulation, |
---|
| 110 | final int thread) |
---|
| 111 | { |
---|
| 112 | // load sortedPop integers |
---|
| 113 | final Individual[] i = s.population.subpops[subpopulation].individuals; |
---|
| 114 | |
---|
| 115 | sortedPop = new int[i.length]; |
---|
| 116 | for(int x=0;x<sortedPop.length;x++) sortedPop[x] = x; |
---|
| 117 | |
---|
| 118 | // sort sortedPop in increasing fitness order |
---|
| 119 | QuickSort.qsort(sortedPop, |
---|
| 120 | new SortComparatorL() |
---|
| 121 | { |
---|
| 122 | public boolean lt(long a, long b) |
---|
| 123 | { |
---|
| 124 | return ((Individual)(i[(int)b])).fitness.betterThan( |
---|
| 125 | ((Individual)(i[(int)a])).fitness); |
---|
| 126 | } |
---|
| 127 | |
---|
| 128 | public boolean gt(long a, long b) |
---|
| 129 | { |
---|
| 130 | return ((Individual)(i[(int)a])).fitness.betterThan( |
---|
| 131 | ((Individual)(i[(int)b])).fitness); |
---|
| 132 | } |
---|
| 133 | }); |
---|
| 134 | |
---|
| 135 | // determine my boundary -- must be at least 1 and must leave 1 over |
---|
| 136 | int boundary = (int)(sortedPop.length * top_n_percent); |
---|
| 137 | if (boundary == 0) boundary = 1; |
---|
| 138 | if (boundary == sortedPop.length) boundary = sortedPop.length-1; |
---|
| 139 | if (boundary == 0) // uh oh |
---|
| 140 | s.output.fatal("Greedy Overselection can only be done with a population of size 2 or more (offending subpopulation #" + subpopulation + ")"); |
---|
| 141 | |
---|
| 142 | // load sortedFitOver |
---|
| 143 | sortedFitOver = new float[boundary]; |
---|
| 144 | int y=0; |
---|
| 145 | for(int x=sortedPop.length-boundary;x<sortedPop.length;x++) |
---|
| 146 | { |
---|
| 147 | sortedFitOver[y] = (i[sortedPop[x]]).fitness.fitness(); |
---|
| 148 | if (sortedFitOver[y] < 0) // uh oh |
---|
| 149 | s.output.fatal("Discovered a negative fitness value. Greedy Overselection requires that all fitness values be non-negative (offending subpopulation #" + subpopulation + ")"); |
---|
| 150 | y++; |
---|
| 151 | } |
---|
| 152 | |
---|
| 153 | // load sortedFitUnder |
---|
| 154 | sortedFitUnder = new float[sortedPop.length-boundary]; |
---|
| 155 | y=0; |
---|
| 156 | for(int x=0;x<sortedPop.length-boundary;x++) |
---|
| 157 | { |
---|
| 158 | sortedFitUnder[y] = (i[sortedPop[x]]).fitness.fitness(); |
---|
| 159 | if (sortedFitUnder[y] < 0) // uh oh |
---|
| 160 | s.output.fatal("Discovered a negative fitness value. Greedy Overselection requires that all fitness values be non-negative (offending subpopulation #" + subpopulation + ")"); |
---|
| 161 | y++; |
---|
| 162 | } |
---|
| 163 | |
---|
| 164 | // organize the distributions. All zeros in fitness is fine |
---|
| 165 | RandomChoice.organizeDistribution(sortedFitUnder, true); |
---|
| 166 | RandomChoice.organizeDistribution(sortedFitOver, true); |
---|
| 167 | } |
---|
| 168 | |
---|
| 169 | public int produce(final int subpopulation, |
---|
| 170 | final EvolutionState state, |
---|
| 171 | final int thread) |
---|
| 172 | { |
---|
| 173 | // pick a coin toss |
---|
| 174 | if (state.random[thread].nextBoolean(gets_n_percent)) |
---|
| 175 | // over -- sortedFitUnder.length to sortedPop.length |
---|
| 176 | return sortedPop[ |
---|
| 177 | sortedFitUnder.length + RandomChoice.pickFromDistribution( |
---|
| 178 | sortedFitOver,state.random[thread].nextFloat())]; |
---|
| 179 | else |
---|
| 180 | // under -- 0 to sortedFitUnder.length |
---|
| 181 | return sortedPop[RandomChoice.pickFromDistribution( |
---|
| 182 | sortedFitUnder,state.random[thread].nextFloat())]; |
---|
| 183 | } |
---|
| 184 | |
---|
| 185 | public void finishProducing(final EvolutionState s, |
---|
| 186 | final int subpopulation, |
---|
| 187 | final int thread) |
---|
| 188 | { |
---|
| 189 | // release the distributions so we can quickly |
---|
| 190 | // garbage-collect them if necessary |
---|
| 191 | sortedFitUnder = null; |
---|
| 192 | sortedFitOver = null; |
---|
| 193 | sortedPop = null; |
---|
| 194 | } |
---|
| 195 | } |
---|