void RpropMinus::step(ObjectiveFunctionType const& objectiveFunction) {
for (size_t i = 0; i < m_parameterSize; i++)
{
double p = m_best.point(i);
if (m_derivative(i) * m_oldDerivative(i) > 0)
{
m_delta(i) = std::min(m_maxDelta, m_increaseFactor * m_delta(i));
}
else if (m_derivative(i) * m_oldDerivative(i) < 0)
{
m_delta(i) = std::max(m_minDelta, m_decreaseFactor * m_delta(i));
}
m_best.point(i) -= m_delta(i) * boost::math::sign(m_derivative(i));
if (! objectiveFunction.isFeasible(m_best.point))
{
m_best.point(i) = p;
m_delta(i) *= m_decreaseFactor;
m_oldDerivative(i) = 0.0;
}
else
{
m_oldDerivative(i) = m_derivative(i);
}
}
//evaluate the new point
m_best.value = objectiveFunction.evalDerivative(m_best.point,m_derivative);
}
void NoisyRprop::step(const ObjectiveFunctionType& objectiveFunction)
{
size_t pc = m_best.point.size();
ObjectiveFunctionType::FirstOrderDerivative dedw;
// compute noisy error function value and derivative
m_best.value = objectiveFunction.evalDerivative(m_best.point,dedw);
// loop over all coordinates
m_gradient += dedw;
for (size_t p=0; p<pc; p++)
{
m_current[p]++;
if (m_current[p] == m_numberOfAverages[p])
{
// basic Rprop algorithm
double value = m_best.point(p);
if (m_gradient(p) < 0.0)
{
m_best.point(p)+=m_delta(p);
m_position[p]++;
m_toTheRight[p]++;
m_rightsum(p) -= m_gradient(p);
}
else if (m_gradient(p) > 0.0)
{
m_best.point(p)-=m_delta(p);
m_position[p]--;
m_leftsum(p) += m_gradient(p);
}
// collect leftmost and rightmost of 100 position samples
if (m_iteration[p] == m_sampleIteration[p])
{
if (m_position[p] > m_rightmost[p]) m_rightmost[p] = m_position[p];
if (m_position[p] < m_leftmost[p]) m_leftmost[p] = m_position[p];
m_sample[p]++;
m_sampleIteration[p] = m_sample[p] * m_episodeLength[p] / 100;
if (m_sampleIteration[p] <= m_iteration[p]) m_sampleIteration[p] = m_iteration[p] + 1;
}
if (!objectiveFunction.isFeasible(m_best.point))
{
m_best.point(p)=value;
}
// strategy adaptation
m_iteration[p]++;
if (m_iteration[p] == m_episodeLength[p])
{
// Compute the normalization of the toTheRight statistics
// and compare it to its expected normal approximation
unsigned int N = m_episodeLength[p];
unsigned int newN = N;
double normalized = std::fabs((m_toTheRight[p] - 0.5 * N) / std::sqrt((double)N));
if (normalized > 1.64) // 90% quantile
{
// increase the step size
m_delta(p) *= 2.0;
newN /= 2;
}
else
{
if (normalized < 0.25) // 20% quantile
{
// decrease the step size
m_delta(p) /= 2.0;
newN *= 2;
}
// Compute the normalization of the spread statistics
// and compare it to its expected distribution
// compute approximate quantiles of the spread statistics
int spread = m_rightmost[p] - m_leftmost[p];
normalized = spread / std::sqrt((double)N);
if (normalized < 0.95) // 10% quantile
{
// decrease the number of iterations
newN /= 2;
}
else if (normalized > 1.77) // 75% quantile
{
// increase the number of iterations
newN *= 2;
}
// Compute the normalization of the asymmetry statistics
// and compare it to its expected distribution
normalized = std::fabs(m_rightsum[p] / m_toTheRight[p] - m_leftsum[p] / (N - m_toTheRight[p])) / (m_rightsum[p] + m_leftsum[p]) * N * std::sqrt((double)N);
if (normalized < 0.29) // 15% quantile
{
// decrease averaging
if (m_numberOfAverages[p] > 1)
{
m_numberOfAverages[p] /= 2;
newN *= 2;
}
}
else if (normalized > 2.5) // 90% quantile
{
// increase averaging
m_numberOfAverages[p] *= 2;
newN /= 2;
}
}
if (newN < m_minEpisodeLength) newN = m_minEpisodeLength;
m_episodeLength[p] = newN;
// reset episode
m_iteration[p] = 0;
m_sample[p] = 0;
m_sampleIteration[p] = m_minEpisodeLength / 100;
m_position[p] = 0;
m_leftmost[p] = 0;
m_rightmost[p] = 0;
m_toTheRight[p] = 0;
m_leftsum(p) = 0.0;
m_rightsum(p) = 0.0;
}
m_current[p] = 0;
m_gradient(p) = 0.0;
}
}
}