Java BoltzmannDistribution.getProbabilities方法代码示例

import burlap.datastructures.BoltzmannDistribution; //导入方法依赖的package包/类
/**
 * Overrides the superclass method to perform a Boltzmann backup operator
 * instead of a Bellman backup operator.
 * Results are stored in this valueFunction's internal map.
 * @param sh the hashed state on which to perform the Boltzmann update.
 * @return the new value
 */
@Override
protected double performBellmanUpdateOn(HashableState sh){

	if(this.tf.isTerminal(sh.s)){
		this.valueFunction.put(sh, 0.);
		return 0.;
	}

	List<QValue> qs = this.getQs(sh.s);
	double [] dqs = new double[qs.size()];
	for(int i = 0; i < qs.size(); i++){
		dqs[i] = qs.get(i).q;
	}
	BoltzmannDistribution bd = new BoltzmannDistribution(dqs, 1./this.boltzBeta);
	double [] dist = bd.getProbabilities();

	double sum = 0.;
	for(int i = 0; i < dqs.length; i++){
		sum += dqs[i] * dist[i];
	}

	this.valueFunction.put(sh, sum);

	return sum;
}
 

import burlap.datastructures.BoltzmannDistribution; //导入方法依赖的package包/类
@Override
public List<ActionProb> policyDistribution(State s) {
	
	HashableState sh = this.hashingFactory.hashState(s);
	PolicyNode node = this.getNode(sh);
	
	double [] prefs = new double[node.preferences.size()];
	for(int i = 0; i < node.preferences.size(); i++){
		prefs[i] = node.preferences.get(i).preference;
	}
	
	BoltzmannDistribution bd = new BoltzmannDistribution(prefs);
	double [] probsArray = bd.getProbabilities();
	
	List <ActionProb> probs = new ArrayList<ActionProb>(probsArray.length);
	for(int i = 0; i < probsArray.length; i++){
		ActionPreference ap = node.preferences.get(i);
		probs.add(new ActionProb(ap.ga, probsArray[i]));
	}
	
	
	return probs;
}
 

import burlap.datastructures.BoltzmannDistribution; //导入方法依赖的package包/类
protected void setV(QAndQGradient qvs){
	double [] qArray = new double[qvs.qs.size()];
	for(int i = 0; i < qvs.qs.size(); i++){
		qArray[i] = qvs.qs.get(i).q;
	}
	BoltzmannDistribution bd = new BoltzmannDistribution(qArray, 1./DifferentiableSparseSampling.this.boltzBeta);
	double [] probs = bd.getProbabilities();
	double sum = 0.;
	for(int i = 0; i < qArray.length; i++){
		sum += qArray[i] * probs[i];
	}
	this.v = sum;
}
 

import burlap.datastructures.BoltzmannDistribution; //导入方法依赖的package包/类
public static double[] getProbabilities(
        int numberOfArms, Function<Integer, Double> expectedReturnOfArm, double temperature) {
    double[] preferences = new double[numberOfArms];
    for(int i=0; i<preferences.length; i++)
        preferences[i] = expectedReturnOfArm.apply(i);
    BoltzmannDistribution distribution = new BoltzmannDistribution(preferences,temperature);

    return distribution.getProbabilities();
}
 

import burlap.datastructures.BoltzmannDistribution; //导入方法依赖的package包/类
@Override
public double apply(double[] qs) {
	BoltzmannDistribution bd = new BoltzmannDistribution(qs, temp);
	double [] dist = bd.getProbabilities();
	double sum = 0.;
	for(int i = 0; i < qs.length; i++){
		sum += qs[i] * dist[i];
	}
	return sum;
}
 

import burlap.datastructures.BoltzmannDistribution; //导入方法依赖的package包/类
/**
    * Computes the gradient of the Boltzmann (softmax) policy wrt some parameters.
    * @param prefs the action-wise preference values passed through the softmax
    * @param grads the gradients of the preference-values with respect to the parameters
    * @param aind the index of the action for which the gradient is being queried
    * @param beta the softmax beta parameter. This parameter is the inverse of the Botlzmann temperature. As beta becomes larger, the policy becomes more deterministic. Should lie in [0, +ifnty].
    * @return the gradient of the policy
    */
public static FunctionGradient computePolicyGradient(double [] prefs, FunctionGradient[] grads, int aind, double beta){

	//first compute policy probs
	BoltzmannDistribution bd = new BoltzmannDistribution(prefs, 1./beta);
	double [] probs = bd.getProbabilities();

	return computePolicyGradient(probs, prefs, grads, aind, beta);

}