StochasticLearner.cpp

//
//  StochasticLearner.cpp
//  Vasco
//
//  Created by Jamil Dhanani on 3/20/2014.
//  Copyright (c) 2014 Jamil Dhanani. All rights reserved.
//

#include "StochasticLearner.h"

#define STOCHASTIC_TOLERANCE 0.00001

#pragma mark - Constructor

StochasticLearner::StochasticLearner(int numParams, vector<SupervisedData*> data, float learningRate) : Learner(numParams, data, learningRate) {
  // No specific initialization needed for derived class
}

#pragma mark - Hypothesis

double StochasticLearner::getHypothesisForData(Data *data) {
  // LOGISTIC REGRESSION
  // h_{\theta}(x) = \frac{1}{1+e^{-\theta^{T}x}}

  double hyp = 0;
  for (int i = 0; i < _M; i++) {
    hyp += data->features()[i] * _parameterValues[i];
  }

  if (_hypothesis)
    hyp = _hypothesis(hyp);
  else { // Assume sigmoid function if no activation function found
    hyp = 1 + exp(-1 * hyp); // Denominator
    hyp = 1/hyp;
  }

  return hyp;
}

#pragma mark - Update Rule

void StochasticLearner::step_stochastic(int i, int j, float target) {
  // Where i is the parameter to update, j is the training example
  // Fixed learning rate for now, but may change that later depending on performance

  Data* data = _data.at(j);

  float hypothesis = getHypothesisForData(data);
  float xi = data->features()[i];

  _parameterValues[i] += _learningRate * (target - hypothesis) * xi;
}

void StochasticLearner::update() {
  updateUntilConvergence();
}

void StochasticLearner::updateUntilConvergence() {
  int examples = (int)_data.size();

  float oldParams[_M];
  for (int i = 0; i < _M; i++)
    oldParams[i] = _parameterValues[i];
  bool nochange[_M];
  for (int i = 0; i < _M; i++)
    nochange[i] = 0;

  int finished = 0;
  do {
    LOG("NEW ITERATION\n", NULL);

    // Iterate through training set
    for (int j = 0; j < examples; j++) {
      SupervisedData *example = _data.at(j);
      float target = example->supervisedValue();
      // Update each parameter
      for (int i = 0; i < _M; i++) {
        if (!nochange[i])
          step_stochastic(i, j, target);
      }
    }

    // Check against old parameters
    finished = 0;
    for (int i = 0; i < _M; i++) {
      double percdiff = oldParams[i] ? fabs((_parameterValues[i] - oldParams[i])/oldParams[i]) : fabs(_parameterValues[i] - oldParams[i]);
      if (percdiff < STOCHASTIC_TOLERANCE) {
        //cout << "Done " << i << endl;
        nochange[i] = true;
        finished++;
      }
      else {
        LOG("Not done: %d\tDiff: %f%%\n", i, percdiff * 100);
      }
      oldParams[i] = _parameterValues[i];
    }

  } while (finished < _M);
}