001 package aima.probability.reasoning;
002
003 import java.util.Arrays;
004 import java.util.List;
005
006 import aima.probability.RandomVariable;
007 import aima.util.Matrix;
008
009 /**
010 * @author Ravi Mohan
011 *
012 */
013
014 public class HiddenMarkovModel {
015
016 SensorModel sensorModel;
017
018 TransitionModel transitionModel;
019
020 private RandomVariable priorDistribution;
021
022 public HiddenMarkovModel(RandomVariable priorDistribution,
023 TransitionModel tm, SensorModel sm) {
024 this.priorDistribution = priorDistribution;
025 this.transitionModel = tm;
026 this.sensorModel = sm;
027 }
028
029 public RandomVariable prior() {
030 return priorDistribution;
031 }
032
033 public RandomVariable predict(RandomVariable aBelief, String action) {
034 RandomVariable newBelief = aBelief.duplicate();
035
036 Matrix beliefMatrix = aBelief.asMatrix();
037 Matrix transitionMatrix = transitionModel.asMatrix(action);
038 Matrix predicted = transitionMatrix.transpose().times(beliefMatrix);
039 newBelief.updateFrom(predicted);
040 return newBelief;
041 }
042
043 public RandomVariable perceptionUpdate(RandomVariable aBelief,
044 String perception) {
045 RandomVariable newBelief = aBelief.duplicate();
046
047 // one way - use matrices
048 Matrix beliefMatrix = aBelief.asMatrix();
049 Matrix o_matrix = sensorModel.asMatrix(perception);
050 Matrix updated = o_matrix.times(beliefMatrix);
051 newBelief.updateFrom(updated);
052 newBelief.normalize();
053 return newBelief;
054
055 // alternate way of doing this. clearer in intent.
056 // for (String state : aBelief.states()){
057 // double probabilityOfPerception= sensorModel.get(state,perception);
058 // newBelief.setProbabilityOf(state,probabilityOfPerception *
059 // aBelief.getProbabilityOf(state));
060 // }
061 }
062
063 public RandomVariable forward(RandomVariable aBelief, String action,
064 String perception) {
065
066 return perceptionUpdate(predict(aBelief, action), perception);
067 }
068
069 public RandomVariable forward(RandomVariable aBelief, String perception) {
070
071 return forward(aBelief, HmmConstants.DO_NOTHING, perception);
072 }
073
074 public RandomVariable calculate_next_backward_message(
075 RandomVariable forwardBelief,
076 RandomVariable present_backward_message, String perception) {
077 RandomVariable result = present_backward_message.duplicate();
078 // System.out.println("fb :-calculating new backward message");
079 // System.out.println("fb :-diagonal matrix from sens model = ");
080 Matrix oMatrix = sensorModel.asMatrix(perception);
081 // System.out.println(oMatrix);
082 Matrix transitionMatrix = transitionModel.asMatrix();// action
083 // should
084 // be
085 // passed
086 // in
087 // here?
088 // System.out.println("fb :-present backward message = "
089 // +present_backward_message);
090 Matrix backwardMatrix = transitionMatrix.times(oMatrix
091 .times(present_backward_message.asMatrix()));
092 Matrix resultMatrix = backwardMatrix.arrayTimes(forwardBelief
093 .asMatrix());
094 result.updateFrom(resultMatrix);
095 result.normalize();
096 // System.out.println("fb :-normalized new backward message = "
097 // +result);
098 return result;
099 }
100
101 public List<RandomVariable> forward_backward(List<String> perceptions) {
102 RandomVariable forwardMessages[] = new RandomVariable[perceptions
103 .size() + 1];
104 RandomVariable backwardMessage = priorDistribution.createUnitBelief();
105 RandomVariable smoothedBeliefs[] = new RandomVariable[perceptions
106 .size() + 1];
107
108 forwardMessages[0] = priorDistribution;
109 smoothedBeliefs[0] = null;
110
111 // populate forward messages
112 for (int i = 0; i < perceptions.size(); i++) { // N.B i starts at 1,
113 // not zero
114 forwardMessages[i + 1] = forward(forwardMessages[i], perceptions
115 .get(i));
116 }
117 for (int i = perceptions.size(); i > 0; i--) {
118 RandomVariable smoothed = priorDistribution.duplicate();
119 smoothed.updateFrom(forwardMessages[i].asMatrix().arrayTimes(
120 backwardMessage.asMatrix()));
121 smoothed.normalize();
122 smoothedBeliefs[i] = smoothed;
123 backwardMessage = calculate_next_backward_message(
124 forwardMessages[i], backwardMessage, perceptions.get(i - 1));
125 }
126
127 return Arrays.asList(smoothedBeliefs);
128 }
129
130 public SensorModel sensorModel() {
131
132 return sensorModel;
133 }
134
135 public TransitionModel transitionModel() {
136
137 return transitionModel;
138 }
139
140 }