|
| 1 | +package aima.core.probability; |
| 2 | + |
| 3 | +import aima.core.agent.Action; |
| 4 | +import aima.core.agent.Agent; |
| 5 | +import aima.core.agent.Percept; |
| 6 | +import aima.core.probability.bayes.BayesInference; |
| 7 | +import aima.core.probability.domain.FiniteDomain; |
| 8 | +import aima.core.probability.proposition.AssignmentProposition; |
| 9 | + |
| 10 | +import java.util.ArrayList; |
| 11 | +import java.util.Collections; |
| 12 | +import java.util.List; |
| 13 | + |
| 14 | +/** |
| 15 | + * Artificial Intelligence A Modern Approach (3rd Edition): Figure 16.9, page 632.<br> |
| 16 | + * </br> |
| 17 | + * <pre> |
| 18 | + * |
| 19 | + * function INFORMATION-GATHERING-AGENT(percept) returns an action |
| 20 | + * persistent: D, a decision network |
| 21 | + * |
| 22 | + * integrate percept into D |
| 23 | + * j ← the value that maximizes VPI(Ej) / Cost(Ej) |
| 24 | + * if VPI(Ej) > Cost(Ej) |
| 25 | + * return REQUEST(Ej) |
| 26 | + * else return the best action from D |
| 27 | + * |
| 28 | + * </pre> |
| 29 | + * <p> |
| 30 | + * Figure ?? Design of a simple information-gathering agent. |
| 31 | + * The agent works by repeatedly selecting the observation with |
| 32 | + * the highest information value, until the cost of the next |
| 33 | + * observation is greater than its expected benefit. |
| 34 | + * |
| 35 | + * @author samagra |
| 36 | + */ |
| 37 | +public abstract class InformationGatheringAgent implements Agent { |
| 38 | + |
| 39 | + // To carry out conditional probability calculations |
| 40 | + private BayesInference inferenceMethod; |
| 41 | + // persistent: D, a decision network |
| 42 | + private DecisionNetwork decisionNetwork; |
| 43 | + // To store the information collected till now |
| 44 | + private List<AssignmentProposition> observedEvidence; |
| 45 | + // To store the scope of information that can be collected |
| 46 | + private List<RandomVariable> randomVars; |
| 47 | + |
| 48 | + /** |
| 49 | + * Constructor for the agent. |
| 50 | + * |
| 51 | + * @param decisionNetwork The decision network which represents the problem |
| 52 | + * for which the information is to be collected |
| 53 | + * @param inferenceMethod To carry out various conditional probability calculations |
| 54 | + * @param initialEvidence The information which is available beforehand to the agent. |
| 55 | + */ |
| 56 | + public InformationGatheringAgent(DecisionNetwork decisionNetwork, |
| 57 | + BayesInference inferenceMethod, |
| 58 | + List<AssignmentProposition> initialEvidence) { |
| 59 | + this.decisionNetwork = decisionNetwork; |
| 60 | + this.inferenceMethod = inferenceMethod; |
| 61 | + this.observedEvidence = initialEvidence; |
| 62 | + this.randomVars = this.decisionNetwork.getNetwork().getVariablesInTopologicalOrder(); |
| 63 | + } |
| 64 | + |
| 65 | + public InformationGatheringAgent(DecisionNetwork decisionNetwork, |
| 66 | + BayesInference inferenceMethod) { |
| 67 | + this(decisionNetwork, inferenceMethod, new ArrayList<>()); |
| 68 | + } |
| 69 | + |
| 70 | + /** |
| 71 | + * function INFORMATION-GATHERING-AGENT(percept) returns an action |
| 72 | + * |
| 73 | + * @param percept The current percept of a sequence perceived by the Agent. |
| 74 | + * @return action to be executed by the agent |
| 75 | + */ |
| 76 | + @Override |
| 77 | + public Action execute(Percept percept) { |
| 78 | + // integrate percept into D |
| 79 | + observedEvidence = integratePercept(observedEvidence, percept); |
| 80 | + |
| 81 | + // j ← the value that maximizes VPI(Ej) / Cost(Ej) |
| 82 | + List<Double> vpiPerUnitCosts = this.vpiPerUnitCost(this.randomVars); |
| 83 | + int j = vpiPerUnitCosts.indexOf(Collections.max(vpiPerUnitCosts)); |
| 84 | + RandomVariable randomVar = this.randomVars.get(j); |
| 85 | + |
| 86 | + // if VPI(Ej) > Cost(Ej) |
| 87 | + if (getVpi(randomVar) > getCost(randomVar)) { |
| 88 | + // return REQUEST(Ej) |
| 89 | + return this.request(randomVar); |
| 90 | + } |
| 91 | + // else return the best action from D |
| 92 | + return ((Action) decisionNetwork.getBestAction()); |
| 93 | + } |
| 94 | + |
| 95 | + /** |
| 96 | + * We assume that the result of |
| 97 | + * the action Request (Ej ) is that the next percept provides the value of Ej . |
| 98 | + * |
| 99 | + * @param randomVar The random variable for which the information is needed. |
| 100 | + * @return The action which leads to the agent to the value of Ej. |
| 101 | + */ |
| 102 | + protected abstract Action request(RandomVariable randomVar); |
| 103 | + |
| 104 | + /** |
| 105 | + * Calculates the vpi (value of perfect information) per unit cost |
| 106 | + * for all the random variables. |
| 107 | + * |
| 108 | + * @param variablesInTopologicalOrder The variables for which information is required. |
| 109 | + * @return A list of vpi values. |
| 110 | + */ |
| 111 | + private List<Double> vpiPerUnitCost(List<RandomVariable> variablesInTopologicalOrder) { |
| 112 | + List<Double> vpiPerUnitCost = new ArrayList<>(); |
| 113 | + for (RandomVariable var : |
| 114 | + variablesInTopologicalOrder) { |
| 115 | + vpiPerUnitCost.add(getVpi(var) / getCost(var)); |
| 116 | + } |
| 117 | + return vpiPerUnitCost; |
| 118 | + } |
| 119 | + |
| 120 | + /** |
| 121 | + * Calculates the cost of obtaining information for |
| 122 | + * a particular variable. |
| 123 | + * |
| 124 | + * @param var |
| 125 | + * @return |
| 126 | + */ |
| 127 | + abstract double getCost(RandomVariable var); |
| 128 | + |
| 129 | + /** |
| 130 | + * Calculates VPI for a particular random variable. |
| 131 | + * |
| 132 | + * @param var |
| 133 | + * @return |
| 134 | + */ |
| 135 | + double getVpi(RandomVariable var) { |
| 136 | + double vpi = 0; |
| 137 | + CategoricalDistribution distribution = inferenceMethod.ask((new RandomVariable[]{var}), |
| 138 | + ((AssignmentProposition[]) observedEvidence.toArray()), decisionNetwork.getNetwork()); |
| 139 | + for (Object value : |
| 140 | + ((FiniteDomain) var.getDomain()).getPossibleValues()) { |
| 141 | + double posterierProb = distribution.getValue(value); |
| 142 | + List<AssignmentProposition> modifiedEvidence = new ArrayList<>(observedEvidence); |
| 143 | + modifiedEvidence.add(new AssignmentProposition(var, value)); |
| 144 | + double expectedUtilityForParticularValue = decisionNetwork.getExpectedUtility(var, |
| 145 | + modifiedEvidence); |
| 146 | + vpi += posterierProb * expectedUtilityForParticularValue; |
| 147 | + } |
| 148 | + vpi -= decisionNetwork.getExpectedUtility(var, observedEvidence); |
| 149 | + return vpi; |
| 150 | + } |
| 151 | + |
| 152 | + /** |
| 153 | + * Extracts the information from the percepts and adds ot to our observed evidence. |
| 154 | + * |
| 155 | + * @param observedEvidence |
| 156 | + * @param percept |
| 157 | + * @return |
| 158 | + */ |
| 159 | + abstract List<AssignmentProposition> integratePercept(List<AssignmentProposition> observedEvidence, Percept percept); |
| 160 | + |
| 161 | + |
| 162 | +} |
0 commit comments