void CfsSubsetEvaluator::setDataFrame(const DataFrame& dataFrame, TgsProgress* progress)
{
TgsProgress* discretizeProgress = NULL;
TgsProgress* classCorrelationProgress = NULL;
TgsProgress* correlationMatrixProgress = NULL;
if (progress)
{
discretizeProgress = progress->createTgsChild("Discretizing Data Frame", .5);
classCorrelationProgress = progress->createTgsChild("Calculating Class Correlation", .1);
correlationMatrixProgress = progress->createTgsChild("Calculating Correlation Matrix", .4);
}
_df = dataFrame;
DataFrameDiscretizer dfd;
dfd.discretize(_df, discretizeProgress);
_calculateClassCorrelations(classCorrelationProgress);
_calculateCorrelationMatrix(correlationMatrixProgress);
// cout << "class corr: " << _classCorr << endl;
// cout << "corr matrix: " << endl;
// for (unsigned int i = 0; i < _corrMatrix.size(); i++)
// {
// cout << _corrMatrix[i] << endl;
// }
_df.clear();
}
double FeatureScoreFitnessFunction::calculateFitness(const Genome& genome)
{
const CalculatorGenome& cg = dynamic_cast<const CalculatorGenome&>(genome);
_workingCopy->setFactorType(0, DataFrame::Numerical);
for (unsigned int i = 0; i < _workingCopy->getNumDataVectors(); i++)
{
_workingCopy->setDataElement(i, 0, cg.calculateValue(i));
}
DataFrameDiscretizer dfd;
try
{
dfd.discretize(*_workingCopy);
}
catch (const Tgs::Exception&)
{
// probably because of no valid values.
return 0.0;
}
double score = _fsc->calculateScore(*_workingCopy, -1, *_workingCopy, 0);
int nodeCount = cg.countNodes();
if (nodeCount > 10)
{
score *= pow(.95, nodeCount - 10);
}
return score;
}
void ConsistencySubsetEvaluator::setDataFrame(const DataFrame& dataFrame, TgsProgress* progress)
{
_dataFrame = dataFrame;
DataFrameDiscretizer dfd;
dfd.discretize(_dataFrame, progress);
_enumCnt = 0;
for (unsigned int i = 0; i < _dataFrame.getNumDataVectors(); i++)
{
string className = _dataFrame.getTrainingLabel(i);
if (_enumMap.find(className) == _enumMap.end())
{
_enumMap[className] = _enumCnt++;
}
}
_putDataIntoBins();
}
double CfsFitnessFunction::calculateFitness(const Genome& genome)
{
const CalculatorGenome& cg = dynamic_cast<const CalculatorGenome&>(genome);
_workingCopy->setFactorType(_workingFactor, DataFrame::Numerical);
double start = Time::getTime();
//cout << cg.toString() << endl;
// one tile per 3 minute
const double MIN_EVAL_PER_SECOND = 250000 / 180;
for (unsigned int i = 0; i < _workingCopy->getNumDataVectors(); i++)
{
double elapsed = Time::getTime() - start;
if (elapsed >= 0.5)
{
double speed = i / elapsed;
cout << "speed " << speed << "\r";
if (speed < MIN_EVAL_PER_SECOND || elapsed >= 10)
{
//cout << endl << "Too slow: " << cg.toString() << endl;
return -1e9;
}
cout.flush();
}
double v;
v = cg.calculateValue(_workingUids[i]);
_workingCopy->setDataElement(i, _workingFactor, v);
}
DataFrameDiscretizer dfd;
try
{
dfd.discretize(*_workingCopy);
}
catch (const DataFrameDiscretizer::AllNullsException&)
{
// because of no valid values.
return -1e9;
}
CfsSubsetEvaluator cse;
cse.setDataFrame(*_workingCopy);
std::vector<int> v;
v.resize(_workingCopy->getNumFactors());
for (unsigned int i = 0; i < _workingCopy->getNumFactors(); i++)
{
v[i] = i;
}
double score = cse.evaluateSubset(v) - _baseScore;
int nodeCount = cg.countNodes();
// penalize for large trees
const int MAX_DESIRABLE_SIZE = 10;
if (nodeCount > MAX_DESIRABLE_SIZE)
{
double m = 1 - pow(nodeCount - MAX_DESIRABLE_SIZE, 1.4) * .05;
//cout << nodeCount << ", " << m << ", " << score << ", " << score * m << endl;
if (m <= 0)
{
score = -numeric_limits<double>::infinity();
}
else if (score > 0.0)
{
score *= m;
}
else
{
score /= m;
}
}
return score;
}
