[1] | 1 | package geniusweb.blingbling;
|
---|
| 2 |
|
---|
| 3 | //old one not used.
|
---|
| 4 | import java.util.ArrayList;
|
---|
| 5 | import java.util.Collections;
|
---|
| 6 | import java.util.Comparator;
|
---|
| 7 | import java.util.HashMap;
|
---|
| 8 | import java.util.HashSet;
|
---|
| 9 | import java.util.LinkedList;
|
---|
| 10 | import java.util.List;
|
---|
| 11 | import java.util.Set;
|
---|
| 12 | import java.util.logging.Level;
|
---|
| 13 |
|
---|
| 14 | import org.neuroph.core.Layer;
|
---|
| 15 | import org.neuroph.core.NeuralNetwork;
|
---|
| 16 | import org.neuroph.core.Neuron;
|
---|
| 17 | import org.neuroph.core.data.DataSet;
|
---|
| 18 | import org.neuroph.core.data.DataSetRow;
|
---|
| 19 | import org.neuroph.core.learning.LearningRule;
|
---|
| 20 | import org.neuroph.nnet.learning.BackPropagation;
|
---|
| 21 | import org.neuroph.util.ConnectionFactory;
|
---|
| 22 |
|
---|
| 23 | import geniusweb.profile.DefaultPartialOrdering;
|
---|
| 24 | import geniusweb.profile.Profile;
|
---|
| 25 | import geniusweb.profile.utilityspace.UtilitySpace;
|
---|
| 26 | import geniusweb.issuevalue.Bid;
|
---|
| 27 | import geniusweb.issuevalue.Domain;
|
---|
| 28 | import geniusweb.issuevalue.Value;
|
---|
| 29 |
|
---|
| 30 | //ranknet
|
---|
| 31 | import geniusweb.blingbling.Ranknet.NeuralRankNet;
|
---|
| 32 | import geniusweb.blingbling.Ranknet.*;
|
---|
| 33 |
|
---|
| 34 |
|
---|
| 35 | import tudelft.utilities.logging.Reporter;
|
---|
| 36 |
|
---|
| 37 |
|
---|
| 38 | public class MyUtilitySpace {
|
---|
| 39 | //my estimate utility space class.
|
---|
| 40 | private Domain domain;
|
---|
| 41 | private List<Bid> bidlist = new ArrayList<>();
|
---|
| 42 | private List<Bid> sortedbidlist = new LinkedList<>();
|
---|
| 43 | private Bid reservationbid;
|
---|
| 44 | private Bid maxBid;
|
---|
| 45 | private Bid minBid;
|
---|
| 46 | private HashMap<String, HashMap<Value, Integer>> valuefrequency = new HashMap<String, HashMap<Value, Integer>>();//new?or null?
|
---|
| 47 | private NeuralNetwork<LearningRule> ann = new NeuralNetwork<LearningRule>();// the network
|
---|
| 48 | private HashMap<String, Integer> issueToNeuron = new HashMap<String, Integer>(); //track the neuron position of issue
|
---|
| 49 | private HashMap<String, HashMap<Value, Integer>> valueToNeuron = new HashMap<String, HashMap<Value, Integer>>();
|
---|
| 50 | int maxIter = 500; //Integer.parseInt(args[0]);
|
---|
| 51 | double maxError = 0.01; // Double.parseDouble(args[1]);
|
---|
| 52 | double learningRate = 0.0005 ; // Double.parseDouble(args[2]);
|
---|
| 53 |
|
---|
| 54 | DefaultPartialOrdering prof;
|
---|
| 55 | List<List<Integer>> betterlist;
|
---|
| 56 |
|
---|
| 57 |
|
---|
| 58 | public MyUtilitySpace(Profile profile, Reporter reporter) {
|
---|
| 59 |
|
---|
| 60 | DefaultPartialOrdering prof = (DefaultPartialOrdering) profile;
|
---|
| 61 |
|
---|
| 62 | this.domain = prof.getDomain();
|
---|
| 63 | this.bidlist = prof.getBids();
|
---|
| 64 | this.reservationbid = prof.getReservationBid();
|
---|
| 65 | this.sortedbidlist = setSortedBids(prof);
|
---|
| 66 | this.maxBid = sortedbidlist.get(sortedbidlist.size()-1);
|
---|
| 67 | this.minBid = sortedbidlist.get(0);
|
---|
| 68 | this.betterlist = prof.getBetter();//for learning to rank method.
|
---|
| 69 |
|
---|
| 70 |
|
---|
| 71 | initValueFrequency();
|
---|
| 72 | setvaluefrequency(prof.getBids());//for elicit compare.
|
---|
| 73 | buildNetwork(prof.getDomain());
|
---|
| 74 | initweight();
|
---|
| 75 | DataSet ds = setDataset(this.sortedbidlist);
|
---|
| 76 | trainNetwork(ds);
|
---|
| 77 | }
|
---|
| 78 |
|
---|
| 79 |
|
---|
| 80 | public void buildNetwork(Domain domain) {
|
---|
| 81 | //neural network model.
|
---|
| 82 | Set<String> issueset = domain.getIssues();//get all the issues.
|
---|
| 83 | Layer inputlayer = new Layer(); // set the input layer.
|
---|
| 84 | Layer weightlayer = new Layer(); // set the weight layer.
|
---|
| 85 | int weightind = 0; // the ind tracks the issue position in weight layer
|
---|
| 86 | int valueind = 0; // the ind tracks the value position in input layer
|
---|
| 87 | for (String issue: issueset) {//iterate the issues
|
---|
| 88 | weightlayer.addNeuron(weightind, new Neuron()); // add one neuron to the weight layer as
|
---|
| 89 | issueToNeuron.put(issue, weightind);//put the index to the issue
|
---|
| 90 | weightind ++;//update the weight index
|
---|
| 91 |
|
---|
| 92 | HashMap<Value, Integer> temp = new HashMap<Value, Integer>();
|
---|
| 93 | for (Value value: domain.getValues(issue)) {
|
---|
| 94 | inputlayer.addNeuron(valueind, new Neuron());// add neuron for every value.
|
---|
| 95 | temp.put(value, valueind);
|
---|
| 96 | valueToNeuron.put(issue, temp);
|
---|
| 97 | valueind ++; //update the value index(position)
|
---|
| 98 | }
|
---|
| 99 | }
|
---|
| 100 | ann.addLayer(0, inputlayer);//set layer position
|
---|
| 101 | ann.addLayer(1, weightlayer);
|
---|
| 102 | Layer outputlayer = new Layer();
|
---|
| 103 | outputlayer.addNeuron(0, new Neuron());//add one neuron to the output layer.
|
---|
| 104 | ann.addLayer(2, outputlayer);
|
---|
| 105 | ann.setInputNeurons(inputlayer.getNeurons());
|
---|
| 106 | ann.setOutputNeurons(outputlayer.getNeurons());
|
---|
| 107 | }
|
---|
| 108 |
|
---|
| 109 | public void initweight() {
|
---|
| 110 | //init the NN weight via a specific method. And create connections.
|
---|
| 111 | Set<String> issueset = domain.getIssues();// get all the issues
|
---|
| 112 | for (String issue: issueset) {
|
---|
| 113 | for (Value value: domain.getValues(issue)) {
|
---|
| 114 | //create connections from value neurons to issue neurons
|
---|
| 115 | if(valuefrequency.get(issue).get(value) == 0) {
|
---|
| 116 | ann.createConnection(ann.getLayerAt(0).getNeuronAt(valueToNeuron.get(issue).get(value)),
|
---|
| 117 | ann.getLayerAt(1).getNeuronAt(issueToNeuron.get(issue)),
|
---|
| 118 | 0.0);// if one value never shows in the partial data, then we assign 0 to the weight of this value.
|
---|
| 119 | continue;
|
---|
| 120 | }
|
---|
| 121 | ann.createConnection(ann.getLayerAt(0).getNeuronAt(valueToNeuron.get(issue).get(value)),
|
---|
| 122 | ann.getLayerAt(1).getNeuronAt(issueToNeuron.get(issue)),
|
---|
| 123 | 0.5);// otherwise we assign 0.5 to the weight. If there is another proper way to init this weight?
|
---|
| 124 | }
|
---|
| 125 | ann.createConnection(ann.getLayerAt(1).getNeuronAt(issueToNeuron.get(issue)),
|
---|
| 126 | ann.getLayerAt(2).getNeuronAt(0),
|
---|
| 127 | 1.0/issueset.size());//equal issueweight according to number of issues. Is there a way to control the sum of the weight to be 1.0?
|
---|
| 128 | }
|
---|
| 129 | }
|
---|
| 130 |
|
---|
| 131 | public void initValueFrequency() {
|
---|
| 132 | // create an all-0 hashmap
|
---|
| 133 | for (String issue: domain.getIssues()) {
|
---|
| 134 | HashMap<Value, Integer> vmap = new HashMap<Value, Integer>();
|
---|
| 135 | for (Value value: domain.getValues(issue)) {
|
---|
| 136 | vmap.put(value, 0);
|
---|
| 137 | }
|
---|
| 138 | this.valuefrequency.put(issue, vmap);
|
---|
| 139 | }
|
---|
| 140 | }
|
---|
| 141 |
|
---|
| 142 | public void setvaluefrequency(List<Bid> inbidlist) {
|
---|
| 143 | //init and update the valuefrequency.
|
---|
| 144 | for (Bid bid: inbidlist) {
|
---|
| 145 | for (String issue: bid.getIssues()) {
|
---|
| 146 | Value v = bid.getValue(issue);
|
---|
| 147 | HashMap<Value, Integer> temp = valuefrequency.get(issue);
|
---|
| 148 | int cnt = temp.get(v);
|
---|
| 149 | temp.put(v, cnt+1);
|
---|
| 150 | valuefrequency.put(issue, temp);
|
---|
| 151 | }
|
---|
| 152 | }
|
---|
| 153 | }
|
---|
| 154 |
|
---|
| 155 | public HashMap<String, List<Value>> getmostinformative(){
|
---|
| 156 | HashMap<String, List<Value>> infovalue = new HashMap<String, List<Value>>();
|
---|
| 157 | for (String issue : domain.getIssues()) {
|
---|
| 158 | List<Value> elicitvalueset = new ArrayList<Value>();
|
---|
| 159 | int minfreq = 0;
|
---|
| 160 | for (Value value: domain.getValues(issue)) {
|
---|
| 161 | int freq = valuefrequency.get(issue).get(value);
|
---|
| 162 | if (elicitvalueset.isEmpty()) {
|
---|
| 163 | elicitvalueset.add(value);
|
---|
| 164 | minfreq = freq;
|
---|
| 165 | }else {
|
---|
| 166 | if (freq<minfreq) {
|
---|
| 167 | elicitvalueset.clear();
|
---|
| 168 | elicitvalueset.add(value);
|
---|
| 169 | }else if(freq == minfreq) {
|
---|
| 170 | elicitvalueset.add(value);
|
---|
| 171 | }
|
---|
| 172 | }
|
---|
| 173 | }
|
---|
| 174 | infovalue.put(issue, elicitvalueset);
|
---|
| 175 |
|
---|
| 176 | }
|
---|
| 177 | return infovalue;
|
---|
| 178 | }
|
---|
| 179 |
|
---|
| 180 | public DataSet setDataset(List<Bid> inbidlist) {
|
---|
| 181 | //Construct the training dataset.
|
---|
| 182 | // input sorted bidlist. Utility from low to high. Trained using assigned utility.
|
---|
| 183 | int inputsize = ann.getInputsCount();
|
---|
| 184 | int outputsize = ann.getOutputsCount();
|
---|
| 185 | DataSet ds = new DataSet(inputsize, outputsize);
|
---|
| 186 | int datasize = inbidlist.size();// the size
|
---|
| 187 | double cnt = 0;//
|
---|
| 188 | for (Bid bid: inbidlist) {
|
---|
| 189 | double[] input = new double[inputsize];
|
---|
| 190 | double[] output = new double[outputsize];
|
---|
| 191 | Set<Integer> indset = new HashSet<Integer>();
|
---|
| 192 | for (String issue: domain.getIssues()) {
|
---|
| 193 | Value v = bid.getValue(issue);//get bid's value of issue.
|
---|
| 194 | indset.add(valueToNeuron.get(issue).get(v));// keep a set of the position of the bid values.
|
---|
| 195 | //ValuetoNeural is a map maps every possible value into the position of input value.
|
---|
| 196 | }
|
---|
| 197 |
|
---|
| 198 | for (int ind =0; ind<inputsize; ind++) { //construct input vector.
|
---|
| 199 | if (indset.contains(ind)) {
|
---|
| 200 | input[ind] = 1.0; // assign 1.0 to the bid's value position
|
---|
| 201 | }else {
|
---|
| 202 | input[ind] =0.0;// assign 0 to others
|
---|
| 203 | }
|
---|
| 204 | }
|
---|
| 205 | output[0] = 0.9*(cnt/datasize)+0.1;//the uniform distribution between [0.1, 1], the minimum set to 0.1?
|
---|
| 206 | DataSetRow dsr = new DataSetRow(input, output);// combine the input vector and the output values.
|
---|
| 207 | ds.add(dsr);//add this row to the dataset
|
---|
| 208 | cnt = cnt+1.0; //update the cnt by +1.
|
---|
| 209 | }
|
---|
| 210 | return ds;
|
---|
| 211 | }
|
---|
| 212 |
|
---|
| 213 | public void trainNetwork(DataSet ds) {
|
---|
| 214 | //update the NN via the given compared bids. retrain the
|
---|
| 215 | // is there space to improve? maybe a better learning rule?
|
---|
| 216 | BackPropagation bp = new BackPropagation();
|
---|
| 217 | bp.setMaxIterations(maxIter);
|
---|
| 218 | //bp.setMaxError(maxError);
|
---|
| 219 | bp.setLearningRate(learningRate);
|
---|
| 220 | ann.setLearningRule(bp);
|
---|
| 221 | ann.learn(ds);
|
---|
| 222 | }
|
---|
| 223 |
|
---|
| 224 | private List<Bid> setSortedBids(Profile profile) {
|
---|
| 225 | //returns a sortedlist here. from low utility to high utility.
|
---|
| 226 | DefaultPartialOrdering prof = (DefaultPartialOrdering) profile;
|
---|
| 227 | List<Bid> bidslist = prof.getBids(); //get all the bid in the partial information.
|
---|
| 228 | // NOTE sort defaults to ascending order.
|
---|
| 229 | Collections.sort(bidslist, new Comparator<Bid>() {
|
---|
| 230 |
|
---|
| 231 | @Override
|
---|
| 232 | public int compare(Bid b1, Bid b2) {
|
---|
| 233 | return prof.isPreferredOrEqual(b1, b2) ? 1 : -1; //
|
---|
| 234 | }
|
---|
| 235 |
|
---|
| 236 | });
|
---|
| 237 |
|
---|
| 238 | return bidslist;
|
---|
| 239 | }
|
---|
| 240 |
|
---|
| 241 | //get and update method
|
---|
| 242 | public void update(Bid bid, List<Bid> worseBids) {//note the frequency also need to be updated
|
---|
| 243 | //relist the bids
|
---|
| 244 | int n = 0;
|
---|
| 245 | while (n < sortedbidlist.size() && worseBids.contains(sortedbidlist.get(n)))
|
---|
| 246 | n++;
|
---|
| 247 | LinkedList<Bid> newbids = new LinkedList<Bid>(sortedbidlist);
|
---|
| 248 | newbids.add(n, bid);
|
---|
| 249 | sortedbidlist = newbids;
|
---|
| 250 |
|
---|
| 251 | DataSet ds = setDataset(sortedbidlist);
|
---|
| 252 | trainNetwork(ds);
|
---|
| 253 |
|
---|
| 254 | bidlist.add(bid);//this is unsorted bid list.
|
---|
| 255 | }
|
---|
| 256 |
|
---|
| 257 | public double getUtility(Bid bid) {
|
---|
| 258 | int inputsize = ann.getInputsCount();
|
---|
| 259 | double[] input = new double[inputsize];
|
---|
| 260 |
|
---|
| 261 | Set<Integer> indset = new HashSet<Integer>();
|
---|
| 262 | for (String issue: domain.getIssues()) {
|
---|
| 263 | Value v = bid.getValue(issue);
|
---|
| 264 | indset.add(valueToNeuron.get(issue).get(v));
|
---|
| 265 | }
|
---|
| 266 |
|
---|
| 267 | for (int ind =0; ind<inputsize; ind++) { //construct input
|
---|
| 268 | if (indset.contains(ind)) {
|
---|
| 269 | input[ind] = 1.0;
|
---|
| 270 | }else {
|
---|
| 271 | input[ind] = 0.0;
|
---|
| 272 | }
|
---|
| 273 | }
|
---|
| 274 | ann.setInput(input);
|
---|
| 275 | ann.calculate();
|
---|
| 276 | return ann.getOutput()[0];//transfer double[] to double.
|
---|
| 277 | }
|
---|
| 278 |
|
---|
| 279 | public Domain getDomain() {
|
---|
| 280 | return this.domain;
|
---|
| 281 | }
|
---|
| 282 |
|
---|
| 283 | public Bid getReservationBid() {
|
---|
| 284 | return this.reservationbid;
|
---|
| 285 | }
|
---|
| 286 |
|
---|
| 287 | public List<Bid> getSortedBids(){
|
---|
| 288 | return this.sortedbidlist;
|
---|
| 289 | }
|
---|
| 290 |
|
---|
| 291 | public Bid getBestBid() {
|
---|
| 292 | return this.maxBid;
|
---|
| 293 | }
|
---|
| 294 |
|
---|
| 295 | public Bid getWorstBid() {
|
---|
| 296 | return this.minBid;
|
---|
| 297 | }
|
---|
| 298 | public NeuralNetwork getann() {
|
---|
| 299 | return this.ann;
|
---|
| 300 | }
|
---|
| 301 |
|
---|
| 302 | // The following is for the learning to rank method. use deeplearning4j.
|
---|
| 303 | public DataSet setPairwiseDataset() {
|
---|
| 304 |
|
---|
| 305 | betterlist= prof.getBetter();
|
---|
| 306 |
|
---|
| 307 | return null;
|
---|
| 308 | }
|
---|
| 309 |
|
---|
| 310 | public void trainPairwise(DataSet ds) {
|
---|
| 311 | //using pairwise information
|
---|
| 312 | }
|
---|
| 313 |
|
---|
| 314 | }
|
---|