pkg/src/ew.predict_noNA.c

   1 #include <math.h>
   2 #include <stdlib.h>
   3
   4 void ew_predict_noNA(double* X, double* Y, int* n_, int* K_, double* alpha_, int* grad_, double* weight)
   5 {
   6     int K = *K_;
   7     int n = *n_;
   8     double alpha = *alpha_;
   9     int grad = *grad_;
  10
  11     //at least two experts to combine: various inits
  12     double invMaxError = 1. / 50; //TODO: magic number
  13     double logK = log(K);
  14     double initWeight = 1. / K;
  15     for (int i=0; i<K; i++)
  16         weight[i] = initWeight;
  17     double* error = (double*)malloc(K*sizeof(double));
  18     double* cumError = (double*)calloc(K, sizeof(double));
  19
  20     //start main loop
  21     for (int t=0; t<n; t++ < n)
  22     {
  23         if (grad)
  24         {
  25             double hatY = 0.;
  26             for (int i=0; i<K; i++)
  27                 hatY += X[t*K+i] * weight[i];
  28             for (int i=0; i<K; i++)
  29                 error[i] = 2. * (hatY - Y[t]) * X[t*K+i];
  30         }
  31         else
  32         {
  33             for (int i=0; i<K; i++)
  34             {
  35                 double delta = X[t*K+i] - Y[t];
  36                 error[i] = delta * delta;
  37 /*              if ((X[t*K+i] <= 30 && Y[t] > 30) || (X[t*K+i] > 30 && Y[t] <= 30))
  38                     error[i] = 1.0;
  39                 else
  40                     error[i] = 0.0;
  41 */
  42             }
  43         }
  44         for (int i=0; i<K; i++)
  45             cumError[i] += error[i];
  46
  47         if (t < n-1 && !grad)
  48         {
  49             //weight update is useless
  50             continue;
  51         }
  52
  53         //double eta = invMaxError * sqrt(8*logK/(t+1)); //TODO: good formula ?
  54         double eta = invMaxError * 1. / (t+1); //TODO: good formula ?
  55         for (int i=0; i<K; i++)
  56             weight[i] = exp(-eta * cumError[i]);
  57         double sumWeight = 0.0;
  58         for (int i=0; i<K; i++)
  59             sumWeight += weight[i];
  60         for (int i=0; i<K; i++)
  61             weight[i] /= sumWeight;
  62         //redistribute weights if alpha > 0 (all weights are 0 or more, sum > 0)
  63         for (int i=0; i<K; i++)
  64             weight[i] = (1. - alpha) * weight[i] + alpha/K;
  65     }
  66
  67     free(error);
  68     free(cumError);
  69 }