/** * Runs gradient decent on the hypothesis * * @param tSet the training set to be used */ private void gradientDecent(TrainingExample[] tSet) { double h, val, newVal; int answer; double lm = LogisticRegression.this.learningRate / tSet.length; for (int i = 0; i < numFeatures; ++i) { val = 0; for (TrainingExample t : tSet) { answer = t.getAnswer(); h = predict(t.getInput()); if (i == 0) val += (h - answer); else val += (h - answer) * t.getInput()[i - 1]; } newVal = parameter.get(i, 0) * (1 - lm * LogisticRegression.this.getRegularizationParam()); newVal -= lm * val; parameter.set(i, 0, newVal); } }
/** * Runs gradient decent to tune the parameters of each hypothesis. * * @param iterations the number of times to run gradient decent */ public void tune(int iterations) { for (Hypothesis h : hypothesis) { // construct a new training set using One vs. Rest // if the training example has the same value as the // hypothesis then set the answer to 1 // otherwise set the answer to 0. TrainingExample[] tSet = new TrainingExample[trainingSet.length]; int answer; int i = 0; for (TrainingExample t : trainingSet) { if (t.getAnswer() == h.getClassification()) answer = 1; else answer = 0; tSet[i] = new TrainingExample(t.getInput(), answer); ++i; } for (i = 0; i < iterations; ++i) { h.gradientDecent(tSet); } } }
/** * Calculates the cost of the <code>trainingSet</code>. * * @param hyp the hypothesis to use in calculating the cost. * @return the cost associated with the hypothesis. */ public double defaultCostFunction(Hypothesis hyp) { double error = 0; double h; int answer; for (TrainingExample t : trainingSet) { try { h = (Double) hyp.predict(t.getInput()); } catch (Exception e) { e.printStackTrace(); continue; } answer = t.getAnswer(); error -= answer * log(h) + (1 - answer) * log(1 - h); } double regError = 0; for (int i = 0; i < hyp.getNumFeatures(); ++i) { regError += pow(hyp.getParameter(i), 2); } error += regError / regularizationParam; return error / (2 * trainingSet.length); }