public UpdateParams UpdateWeights(double[] _weights, int _numNetworks, double[] _prevWeightsDelta, double _learningRate, double _momentum, double[] _gradients, double[] _inputs)
{
UpdateParams _up = new UpdateParams();
int _index = 0;
int _index2 = 0;
for (int i = 0; i < _weights.Length; i += _numNetworks) // 0..2 (3)
{
for (int j = 0; j < _numNetworks; ++j) // 0..3 (4)
{
double delta = _learningRate * _gradients[j] * _inputs[_index2]; // compute the new delta = "eta * hGrad * input"
_weights[_index] += delta; // update
_weights[_index] += _momentum * _prevWeightsDelta[_index]; // add mom_momentumentum using previous delta. on first pass old value will be 0.0 but that's OK.
_prevWeightsDelta[_index++] = delta; // save the delta for next time
}
_index2++;
}
_up._weights = _weights;
_up._delta = _prevWeightsDelta;
return(_up);
}
public UpdateParams UpdateBiases(double[] _biases, double[] _prevBiasesDelta, double _learningRate, double _momentum, double[] _gradients)
{
UpdateParams _up = new UpdateParams();
for (int i = 0; i < _biases.Length; ++i)
{
double delta = _learningRate * _gradients[i] * 1.0; // the 1.0 is the constant input for any bias; could leave out
_biases[i] += delta;
_biases[i] += _momentum * _prevBiasesDelta[i];
_prevBiasesDelta[i] = delta; // save delta
}
_up._weights = _biases;
_up._delta = _prevBiasesDelta;
return(_up);
}
