神经网络自适应学习代码

时间:2013-09-19 12:44:10

标签: java double neural-network nan

我有以下代码

public void alearningrate(){

        //adopt learningrate 
        System.out.println(errorpositive+""+learningrate);
        if ((errorpositive&&errorgradient>0)||!errorpositive&&errorgradient<0){
            learningrate=learningrate*1.5;

        }else{
            learningrate=learningrate*0.75;
        }
        if (learningrate<0.0001){
            learningrate=0.0001;
        }
    }

    public void lastblame(double blame){

        errorpositive=errorgradient>0;
        errorgradient= Math.exp(-value)/Math.pow(1+Math.exp(-value), 2)*blame;
        if (value==Double.NaN){
            System.out.println("nan detected=?");
            errorgradient=1;
        }
        for (int i=0; i<weight.length;i++){
            weight[i]=weight[i]+learningrate*source[i].output()*errorgradient;
        }

        //alearningrate();

    }
    public void hiddenblame(double blame){
        System.out.println(value);
        errorpositive=errorgradient>0;
        errorgradient=Math.exp(-value)/Math.pow(1+Math.exp(-value), 2)*blame;
        if (value==Double.NaN){
            System.out.println("nan detected=?");
            errorgradient=1;
        }
        for (int i=0; i<weight.length;i++){
            weight[i]=weight[i]+learningrate*source[i].output()*errorgradient;
        }
        System.out.println(errorgradient);
        alearningrate();

    }

然而这个causus的错误梯度很快就变成了NaN,但是如果我决定删除alearningrate()那么代码运行正常。我查看了我的函数errorgradient函数,它应该永远不会产生nan,任何人都知道为什么我一直得到NaN?

0 个答案:

没有答案