const vector<double>& KernelEstimationInterpolator::interpolate(const vector<double>& point) const
{
const DataFrame& df = *_df;
_result.resize(_depColumns.size());
for (size_t i = 0; i < _result.size(); i++)
{
_result[i] = 0.0;
}
vector<double> simplePoint(_indColumns.size());
for (size_t i = 0; i < _indColumns.size(); i++)
{
simplePoint[i] = point[_indColumns[i]];
}
double n0 = Normal::normal(0, _sigma);
KnnIteratorNd it(_getIndex(), simplePoint);
double wSum = 0.0;
while (it.next() && it.getDistance() < _sigma * 3.0)
{
size_t i = it.getId();
const vector<double>& record = df.getDataVector(i);
// figure out the distance between point and this data vector
double d = 0;
for (size_t j = 0; j < _indColumns.size(); j++)
{
double v = point[_indColumns[j]] - record[_indColumns[j]];
d += v * v;
}
d = sqrt(d);
if (d / _sigma < 3.0)
{
// calculate the weight of this sample.
double w = Normal::normal(d, _sigma) / n0;
wSum += w;
assert(w <= 1.000001);
// calculate the contribution to the predicted value.
for (size_t j = 0; j < _result.size(); j++)
{
_result[j] += (record[_depColumns[j]] * w);
}
}
}
// do less rubber sheeting as we get far away from tie points.
wSum = std::max(wSum, n0);
for (size_t j = 0; j < _result.size(); j++)
{
_result[j] /= wSum;
}
return _result;
}
double KernelEstimationInterpolator::_estimateError(unsigned int index) const
{
const DataFrame& df = *_df;
vector<double> predicted(_depColumns.size(), 0.0);
const vector<double>& uut = df.getDataVector(index);
vector<double> simplePoint(_indColumns.size());
for (size_t i = 0; i < _indColumns.size(); i++)
{
simplePoint[i] = uut[_indColumns[i]];
}
double n0 = Normal::normal(0, _sigma);
KnnIteratorNd it(_getIndex(), simplePoint);
double wSum = 0.0;
while (it.next() && it.getDistance() < _sigma * 3.0)
{
size_t i = it.getId();
if (i == index)
{
continue;
}
const vector<double>& record = df.getDataVector(i);
// figure out the distance between point and this data vector
double d = 0;
for (size_t j = 0; j < _indColumns.size(); j++)
{
double v = uut[_indColumns[j]] - record[_indColumns[j]];
d += v * v;
}
d = sqrt(d);
if (d / _sigma < 3.0)
{
// calculate the weight of this sample.
double w = Normal::normal(d, _sigma) / n0;
wSum += w;
assert(w <= 1.000001);
// calculate the contribution to the predicted value.
for (size_t j = 0; j < predicted.size(); j++)
{
predicted[j] += (record[_depColumns[j]] * w);
}
}
}
// do less rubber sheeting as we get far away from tie points.
wSum = std::max(wSum, n0);
double result = 0.0;
for (size_t j = 0; j < predicted.size(); j++)
{
//cout << "uut[_depColumns[" << j << "]] " << uut[_depColumns[j]] << endl;
//cout << "predicted[" << j << "] " << predicted[j] << endl;
double diff = uut[_depColumns[j]] - (predicted[j] / wSum);
result += diff * diff;
}
result = sqrt(result);
//cout << "wSum: " << wSum << " result: " << result << endl;
return result;
}
const vector<double>& IdwInterpolator::_interpolate(const vector<double>& point, int ignoreId) const
{
const DataFrame& df = *_df;
_result.resize(_depColumns.size());
for (size_t i = 0; i < _result.size(); i++)
{
_result[i] = 0.0;
}
vector<double> simplePoint(_indColumns.size());
for (size_t i = 0; i < _indColumns.size(); i++)
{
simplePoint[i] = point[_indColumns[i]];
}
KnnIteratorNd it(_getIndex(), simplePoint);
double wSum = 0.0;
int samples = 0;
int iterations = 0;
while (it.next() && samples < 50 && iterations <= _maxAllowedPerLoopOptimizationIterations)
{
size_t i = it.getId();
if ((int)i == ignoreId)
{
continue;
}
const vector<double>& record = df.getDataVector(i);
// figure out the distance between point and this data vector
double d = 0;
for (size_t j = 0; j < _indColumns.size(); j++)
{
double v = point[_indColumns[j]] - record[_indColumns[j]];
d += v * v;
}
d = sqrt(d);
// if the distance is zero then the weight is infinite and we don't need to look any further.
if (d == 0)
{
wSum = 1.0;
// Set the contribution equal to this value.
for (size_t j = 0; j < _result.size(); j++)
{
_result[j] += record[_depColumns[j]];
}
break;
}
// calculate the weight of this sample.
double w = _calculateWeight(d);
wSum += w;
// calculate the contribution to the predicted value.
for (size_t j = 0; j < _result.size(); j++)
{
_result[j] += (record[_depColumns[j]] * w);
}
iterations++;
}
if (iterations > _iterations)
{
_iterations = iterations;
}
for (size_t j = 0; j < _result.size(); j++)
{
_result[j] /= wSum;
}
return _result;
}
