maxent:最大熵模型的java实现，GIS训练

先给出实现的代码，以后再进行一些理论的概述

训练样本

Outdoor Sunny Happy
Outdoor Sunny Happy Dry
Outdoor Sunny Happy Humid
Outdoor Sunny Sad Dry
Outdoor Sunny Sad Humid
Outdoor Cloudy Happy Humid
Outdoor Cloudy Happy Humid
Outdoor Cloudy Sad Humid
Outdoor Cloudy Sad Humid
Indoor Rainy Happy Humid
Indoor Rainy Happy Dry
Indoor Rainy Sad Dry
Indoor Rainy Sad Humid
Indoor Cloudy Sad Humid
Indoor Cloudy Sad Humid

public class Maxent {

	List<Instance> instanceList = new ArrayList<Instance>();
	List<Feature> featureList = new ArrayList<Feature>();
	List<Integer> featureCountList = new ArrayList<Integer>();
	List<String> labels = new ArrayList<String>();
	float[] weight = null;
	float[] lastweight = null;
	float[] empiricalE = null;
	float[] modelE = null;
	int M = 0;

	public static void main(String[] args) throws IOException {

		String path = System.getProperty("user.dir") + "/config/maxent";
		Maxent maxent = new Maxent();
		maxent.loadData(path);
		maxent.train(200);
		List<String> fieldList = new ArrayList<String>();
		fieldList.add("Sunny");
		float[] prob = maxent.predict(fieldList);
		System.out.println(Arrays.toString(prob));
	}

	/**
	 * 加载数据，并且将几个变量赋值
	 * featureList：特征函数的list
	 * featureCountList:与特征函数一一对应的，特征函数出现的次数
	 * instanceList:样本数据list
	 * labels:类别list
	 * @param path
	 * @throws IOException
	 */
	public void loadData(String path) throws IOException {

		BufferedReader br = new BufferedReader(new FileReader(new File(path)));
		String line = br.readLine();
		while(line != null) {
			String[] segs = line.split("\\s");
			String label = segs[0];
			List<String> fieldList = new ArrayList<String>();
			for(int i = 1; i < segs.length; i++) {
				fieldList.add(segs[i]);
				Feature feature = new Feature(label, segs[i]);
				int index = featureList.indexOf(feature);
				if(index == -1){
					featureList.add(feature);
					featureCountList.add(1);
				} else {
					featureCountList.set(index, featureCountList.get(index) + 1);
				}
			}
			if(fieldList.size() > M) M = fieldList.size();
			Instance instance = new Instance(label,fieldList);
			instanceList.add(instance);
			if(labels.indexOf(label) == -1) labels.add(label);
			line = br.readLine();
		}
	}

	public void train(int maxIt) {

		initParams();
		for(int i = 0; i < maxIt; i++) {
			System.out.println(i);
			modeE();
			for(int w = 0; w < weight.length; w++) {
				lastweight[w] = weight[w];
				weight[w] += 1.0 / M * Math.log(empiricalE[w] / modelE[w]);
			}
			System.out.println(Arrays.toString(weight));
			if(checkExist(lastweight, weight)) break;
		}
	}

	public float[] predict(List<String> fieldList) {

		float[] prob = calProb(fieldList);
		return prob;
	}

	public boolean checkExist(float[] w1, float[] w2) {

		for(int i = 0; i < w1.length; i++) {
			if(Math.abs(w1[i] - w2[i]) >= 0.01)
				return false;
		}
		return true;
	}

	/**
	 * 初始化一些变量
	 * 计算特征函数的经验期望，就是特征函数出现的次数/样本数
	 */
	public void initParams() {

		int size = featureList.size();
		weight = new float[size];
		lastweight = new float[size];
		empiricalE = new float[size];
		modelE = new float[size];

		for(int i = 0; i < size; i++) {
			empiricalE[i] = (float)featureCountList.get(i) / instanceList.size();
		}
	}

	/**
	 * 计算模型期望，即在当前的特征函数的权重下，计算特征函数的模型期望值。
	 */
	public void modeE() {

		modelE = new float[modelE.length];
		for(int i = 0; i < instanceList.size(); i++) {
			List<String> fieldList = instanceList.get(i).fieldList;
			//计算当前样本X对应所有类别的概率
			float[] pro = calProb(fieldList);
			for(int j = 0; j < fieldList.size(); j++) {
				for(int k = 0; k < labels.size(); k++) {
					Feature feature = new Feature(labels.get(k), fieldList.get(j));
					int index = featureList.indexOf(feature);
					if(index != -1)
						modelE[index] += pro[k] * (1.0 / instanceList.size());
				}
			}
		}
	}

	//计算p(y|x),此时的x指的是instance里的field
	public float[] calProb(List<String> fieldList) {

		float[] p = new float[labels.size()];
		float sum = 0;
		for(int i = 0; i < labels.size(); i++) {
			float weightSum = 0;
			for(String field : fieldList) {
				Feature feature = new Feature(labels.get(i), field);
				int index = featureList.indexOf(feature);
				if(index != -1)
					weightSum += weight[index];
			}
			p[i] = (float) Math.exp(weightSum);
			sum += p[i];
		}

		for(int i = 0; i < p.length; i++) {
			p[i] /= sum;
		}
		return p;
	}

	class Instance {

		String label;
		List<String> fieldList = new ArrayList<String>();
		public Instance(String label, List<String> fieldList) {
			this.label = label;
			this.fieldList = fieldList;
		}
	}

	class Feature{

		String label;
		String value;
		public Feature(String label, String value) {
			this.label = label;
			this.value = value;
		}
		@Override
		public boolean equals(Object obj) {
			Feature feature = (Feature) obj;
			if(this.label.equals(feature.label) && this.value.equals(feature.value))
				return true;
			return false;
		}
		@Override
		public String toString() {
			return "[" + label + ", " + value + "]";
		}

	}
}

maxent:最大熵模型的java实现，GIS训练