/**
 * 
 */
package ann.bp;

import java.util.Random;


/**
 * @author tian.yj
 *
 */
public class BPAnn {

	private double eta ;
	private int hn ;
	private int on ;
	private double b = 1.0 ;
	private double momentum = 1.0 ;
	
	private double[][] hw ;
	private double[][] ow ;
	
	private double[] ho ;
	private double[] oo ;
	
	private double[] oe ;
	private double[] he ;
	
	public double hes ;
	public double oes ;
	
	public BPAnn(double eta , int in , int hn , int on) {
		this.eta = eta ;
		this.hn = hn ;
		this.on = on ;
		this.hw = new double[ hn ][ in + 1 ] ;
		this.ow = new double[ on ][ hn + 1 ] ;
		this.oe = new double[ on ] ;
		this.he = new double[ hn ] ;
		random( hw ) ;
		random( ow ) ;
	}
	
	private void random(double[][] m ){
		Random r = new Random(System.currentTimeMillis()) ;
		for( int i = 0 ; i < m.length ; i++ ) {
			for( int j = 0 ; j < m[ i ].length ; j++ ) {
				m[ i ][ j ] = r.nextDouble() * 2 - 1 ;
			}
		}
	}
	
	public double[] test( double[] x ) {
		// 隐藏层的输出
		double[] ho = new double[ hn ];
		double[] oo = new double[ on ] ;
		forward( x , ho , hw ) ;
		forward( ho , oo , ow ) ;
		this.ho = ho ;
		this.oo = oo ;
		return oo ;
	}
	
	private void forward(double[] in , double[] out , double[][] ws){
		// j 第 j 个 单元的 输入权重向量
		for( int j = 0 ; j < ws.length ; j++ ) {
			// ws[ j ][ 0 ] j 的偏置权重
			double sum = b * ws[ j ][ 0 ] ;
			for( int i = 0 ; i < in.length ; i++ ) {
				sum += in[ i ] * ws[ j ][ i + 1 ] ;
			}
			// out[j] 第 j 输出
			sum = sigmoid( sum ) ;
			out[ j ] = sum ;
		}
	}
	
	public void train(double[] in , double[] t){
		test( in ) ;
		calculateError(in, t) ;
		adjust( ho , oe , ow ) ;
		adjust( in , he , hw ) ;
	}
	
	public void calculateError(double[] in , double[] t) {
		double sum = 0 ;
		//调整输出层权重
		for( int i = 0 ; i < oe.length ; i++ ) {
			double o = oo[ i ] ;
			double d = t[ i ] ;
			double e = o * ( 1 - o ) * ( d - o ) ;
			oe[ i ] = e ;
			sum += Math.abs( e ) ;
		}
		oes = sum ;
		sum = 0 ;
		//调整隐藏层权重
		for( int i = 0 ; i < he.length ; i++ ) {
			double o = ho[ i ] ;
			double s = 0 ;
			for( int j = 0 ; j < oe.length ; j++ ) {
				s += ow[ j ][ i + 1 ] * oe[ j ] ;
			}
			double e = o * ( 1 - o ) * s ;
			he[ i ] = e ;
			sum += Math.abs( e ) ;
		}
		hes = sum ;
	}
	
	/**
	 * @param in
	 * @param delta 误差
	 * @param ws
	 */
	private void adjust(double[] in , double[] delta , double[][] ws) {
		// i : 第 i 个输出单元
		for( int i = 0 ; i < delta.length ; i++ ) {
			// j ：分量权重
			for( int j = 0 ; j <= in.length ; j++ ) {
				// xi : i 单元的 输入向量 的 j 分量
				double xi = j == 0 ? b : in[ j - 1 ] ;
				ws[ i ][ j ] = eta * xi * delta[ i ] + momentum * ws[ i ][ j ] ;
			}
		}
	}
	
	private double sigmoid(double x) {
		return 1.0 / ( 1.0 + Math.exp( - x ) ) ;
	}
}

BP（反向传播）神经网络 java 实现