source: trunk/EventBenchCore/src/de/ugoe/cs/eventbench/models/FirstOrderMarkovModel.java @ 101

package de.ugoe.cs.eventbench.models;

import java.util.ArrayList;
import java.util.List;
import java.util.Random;

import de.ugoe.cs.eventbench.data.Event;
import de.ugoe.cs.util.StringTools;
import de.ugoe.cs.util.console.Console;
import edu.uci.ics.jung.graph.Graph;
import edu.uci.ics.jung.graph.SparseMultigraph;
import edu.uci.ics.jung.graph.util.EdgeType;

import Jama.Matrix;

/**
 * <p>
 * Implements first-order Markov models. The implementation is based on
 * {@link HighOrderMarkovModel} and restricts the Markov order to 1. In
 * comparison to {@link HighOrderMarkovModel}, more calculations are possible
 * with first-order models, e.g., the calculation of the entropy (
 * {@link #calcEntropy()}).
 * </p>
 * 
 * @author Steffen Herbold
 * @version 1.0
 */
public class FirstOrderMarkovModel extends HighOrderMarkovModel implements
                IDotCompatible {

        /**
         * <p>
         * Id for object serialization.
         * </p>
         */
        private static final long serialVersionUID = 1L;

        /**
         * <p>
         * Maximum number of iterations when calculating the stationary distribution
         * as the limit of multiplying the transmission matrix with itself.
         * </p>
         */
        final static int MAX_STATDIST_ITERATIONS = 1000;

        /**
         * <p>
         * Constructor. Creates a new FirstOrderMarkovModel.
         * </p>
         * 
         * @param r
         *            random number generator used by probabilistic methods of the
         *            class
         */
        public FirstOrderMarkovModel(Random r) {
                super(1, r);
        }

        /**
         * <p>
         * Generates the transmission matrix of the Markov model.
         * </p>
         * 
         * @return transmission matrix
         */
        private Matrix getTransmissionMatrix() {
                List<Event<?>> knownSymbols = new ArrayList<Event<?>>(
                                trie.getKnownSymbols());
                int numStates = knownSymbols.size();
                Matrix transmissionMatrix = new Matrix(numStates, numStates);

                for (int i = 0; i < numStates; i++) {
                        Event<?> currentSymbol = knownSymbols.get(i);
                        List<Event<?>> context = new ArrayList<Event<?>>();
                        context.add(currentSymbol);
                        for (int j = 0; j < numStates; j++) {
                                Event<?> follower = knownSymbols.get(j);
                                double prob = getProbability(context, follower);
                                transmissionMatrix.set(i, j, prob);
                        }
                }
                return transmissionMatrix;
        }

        /**
         * <p>
         * Calculates the entropy of the model. To make it possible that the model
         * is stationary, a transition from {@link Event#ENDEVENT} to
         * {@link Event#STARTEVENT} is added.
         * </p>
         * 
         * @return entropy of the model or NaN if it could not be calculated
         */
        public double calcEntropy() {
                Matrix transmissionMatrix = getTransmissionMatrix();
                List<Event<?>> knownSymbols = new ArrayList<Event<?>>(
                                trie.getKnownSymbols());
                int numStates = knownSymbols.size();

                int startStateIndex = knownSymbols.indexOf(Event.STARTEVENT);
                int endStateIndex = knownSymbols.indexOf(Event.ENDEVENT);
                if (startStateIndex == -1) {
                        Console.printerrln("Error calculating entropy. Initial state of markov chain not found.");
                        return Double.NaN;
                }
                if (endStateIndex == -1) {
                        Console.printerrln("Error calculating entropy. End state of markov chain not found.");
                        return Double.NaN;
                }
                transmissionMatrix.set(endStateIndex, startStateIndex, 1);

                // Calculate stationary distribution by raising the power of the
                // transmission matrix.
                // The rank of the matrix should fall to 1 and each two should be the
                // vector of the stationory distribution.
                int iter = 0;
                int rank = transmissionMatrix.rank();
                Matrix stationaryMatrix = (Matrix) transmissionMatrix.clone();
                while (iter < MAX_STATDIST_ITERATIONS && rank > 1) {
                        stationaryMatrix = stationaryMatrix.times(stationaryMatrix);
                        rank = stationaryMatrix.rank();
                        iter++;
                }

                if (rank != 1) {
                        Console.traceln("rank: " + rank);
                        Console.printerrln("Unable to calculate stationary distribution.");
                        return Double.NaN;
                }

                double entropy = 0.0;
                for (int i = 0; i < numStates; i++) {
                        for (int j = 0; j < numStates; j++) {
                                if (transmissionMatrix.get(i, j) != 0) {
                                        double tmp = stationaryMatrix.get(i, 0);
                                        tmp *= transmissionMatrix.get(i, j);
                                        tmp *= Math.log(transmissionMatrix.get(i, j)) / Math.log(2);
                                        entropy -= tmp;
                                }
                        }
                }
                return entropy;
        }

        /**
         * <p>
         * The dot represenation of {@link FirstOrderMarkovModel}s is its graph
         * representation with the states as nodes and directed edges weighted with
         * transition probabilities.
         * </p>
         * 
         * @see de.ugoe.cs.eventbench.models.IDotCompatible#getDotRepresentation()
         */
        @Override
        public String getDotRepresentation() {
                StringBuilder stringBuilder = new StringBuilder();
                stringBuilder.append("digraph model {" + StringTools.ENDLINE);

                List<Event<?>> knownSymbols = new ArrayList<Event<?>>(
                                trie.getKnownSymbols());

                for (Event<?> symbol : knownSymbols) {
                        final String thisSaneId = symbol.getShortId().replace("\"", "\\\"")
                                        .replaceAll("[\r\n]", "");
                        stringBuilder.append(" " + symbol.hashCode() + " [label=\""
                                        + thisSaneId + "\"];" + StringTools.ENDLINE);
                        List<Event<?>> context = new ArrayList<Event<?>>();
                        context.add(symbol);
                        List<Event<?>> followers = trie.getFollowingSymbols(context);
                        for (Event<?> follower : followers) {
                                stringBuilder.append(" " + symbol.hashCode() + " -> "
                                                + follower.hashCode() + " ");
                                stringBuilder.append("[label=\""
                                                + getProbability(context, follower) + "\"];"
                                                + StringTools.ENDLINE);
                        }
                }
                stringBuilder.append('}' + StringTools.ENDLINE);
                return stringBuilder.toString();
        }

        /**
         * <p>
         * Returns a {@link Graph} represenation of the model with the states as
         * nodes and directed edges weighted with transition probabilities.
         * </p>
         * 
         * @return {@link Graph} of the model
         */
        public Graph<String, MarkovEdge> getGraph() {
                Graph<String, MarkovEdge> graph = new SparseMultigraph<String, MarkovEdge>();

                List<Event<?>> knownSymbols = new ArrayList<Event<?>>(
                                trie.getKnownSymbols());

                for (Event<?> symbol : knownSymbols) {
                        String from = symbol.getShortId();
                        List<Event<?>> context = new ArrayList<Event<?>>();
                        context.add(symbol);

                        List<Event<?>> followers = trie.getFollowingSymbols(context);

                        for (Event<?> follower : followers) {
                                String to = follower.getShortId();
                                MarkovEdge prob = new MarkovEdge(getProbability(context,
                                                follower));
                                graph.addEdge(prob, from, to, EdgeType.DIRECTED);
                        }
                }
                return graph;
        }

        /**
         * Inner class used for the {@link Graph} representation of the model.
         * 
         * @author Steffen Herbold
         * @version 1.0
         */
        static public class MarkovEdge {
                /**
                 * <p>
                 * Weight of the edge, i.e., its transition probability.
                 * </p>
                 */
                double weight;

                /**
                 * <p>
                 * Constructor. Creates a new MarkovEdge.
                 * </p>
                 * 
                 * @param weight
                 *            weight of the edge, i.e., its transition probability
                 */
                MarkovEdge(double weight) {
                        this.weight = weight;
                }

                /**
                 * <p>
                 * The weight of the edge as {@link String}.
                 * </p>
                 */
                public String toString() {
                        return "" + weight;
                }
        }

}