/// Performs one pattern-presentation of stochastic gradient descent, and dynamically tunes the
// learning rate
void doSomeTraining() {
// Present a sample of beliefs and corresponding contentment for training
int index = rand.nextInt(trainSize);
model.regularize(learningRate, 0.000001);
model.trainIncremental(samples.row(index), contentment.row(index), learningRate);
}
/// Marshals this model to a JSON DOM.
JSONObject marshal() {
JSONObject obj = new JSONObject();
obj.put("model", model.marshal());
obj.put("samples", samples.marshal());
obj.put("contentment", contentment.marshal());
obj.put("trainPos", trainPos);
obj.put("trainSize", trainSize);
obj.put("trainIters", trainIters);
obj.put("learningRate", learningRate);
return obj;
}
/// Refines this model based on feedback from the teacher
void trainIncremental(double[] sample_beliefs, double sample_contentment) {
// Buffer the samples
double[] dest = samples.row(trainPos);
if (sample_beliefs.length != dest.length) throw new IllegalArgumentException("size mismatch");
for (int i = 0; i < dest.length; i++) dest[i] = sample_beliefs[i];
contentment.row(trainPos)[0] = sample_contentment;
trainPos++;
trainSize = Math.max(trainSize, trainPos);
if (trainPos >= samples.rows()) trainPos = 0;
// Do a few iterations of stochastic gradient descent
int iters = Math.min(trainIters, trainSize);
for (int i = 0; i < iters; i++) doSomeTraining();
}
