Point2f GoalPostDetector::findTheShift ()
{
CannyThreshold();
createPaddedImg();
getPointsAlongTheGoalBar();
CMAES<float> evo;
float *arFunvals, *xfinal, *const*pop;
// Initialize everything
const int dim = 2;
float xstart[dim];
for(int i=0; i<dim; i++) xstart[i] = 0.0;
float stddev[dim];
for(int i=0; i<dim; i++) stddev[i] = 0.05;
Parameters<float> parameters;
parameters.init(dim, xstart, stddev);
arFunvals = evo.init(parameters);
// Iterate until stop criterion holds
while(!evo.testForTermination())
{
// Generate lambda new search points, sample population
pop = evo.samplePopulation();
// condition: solution dx and dy should not be bigger than squareSize
for (int i = 0; i < evo.get(CMAES<float>::PopSize); ++i)
while (abs(pop[i][0])*scale>=windowSearchSize||abs(pop[i][1])*scale>=windowSearchSize)
evo.reSampleSingle(i);
// evaluate the new search points using objectiveFunction from above
for (int i = 0; i < evo.get(CMAES<float>::Lambda); ++i)
{
arFunvals[i]= objectiveFunction(pop[i]);
}
evo.updateDistribution(arFunvals);
}
return Point2f(evo.getNew(CMAES<float>::XMean)[0]*scale,
evo.getNew(CMAES<float>::XMean)[1]*scale);
}
/**
* The optimization loop.
*/
int main(int, char**) {
CMAES<double> evo;
double *arFunvals, *const*pop, *xfinal;
// Initialize everything
const int dim = 22;
double xstart[dim];
for(int i=0; i<dim; i++) xstart[i] = 0.5;
double stddev[dim];
for(int i=0; i<dim; i++) stddev[i] = 0.3;
Parameters<double> parameters;
// TODO Adjust parameters here
parameters.init(dim, xstart, stddev);
arFunvals = evo.init(parameters);
std::cout << evo.sayHello() << std::endl;
// Iterate until stop criterion holds
while(!evo.testForTermination())
{
// Generate lambda new search points, sample population
pop = evo.samplePopulation(); // Do not change content of pop
/* Here you may resample each solution point pop[i] until it
becomes feasible, e.g. for box constraints (variable
boundaries). function is_feasible(...) needs to be
user-defined.
Assumptions: the feasible domain is convex, the optimum is
not on (or very close to) the domain boundary, initialX is
feasible and initialStandardDeviations are sufficiently small
to prevent quasi-infinite looping.
*/
/* for (i = 0; i < evo.get(CMAES<double>::PopSize); ++i)
while (!is_feasible(pop[i]))
evo.reSampleSingle(i);
*/
// evaluate the new search points using fitfun from above
for (int i = 0; i < evo.get(CMAES<double>::Lambda); ++i)
arFunvals[i] = fitfun(pop[i], (int) evo.get(CMAES<double>::Dimension));
// update the search distribution used for sampleDistribution()
evo.updateDistribution(arFunvals);
}
std::cout << "Stop:" << std::endl << evo.getStopMessage();
evo.writeToFile(CMAES<double>::WKResume, "resumeevo1.dat"); // write resumable state of CMA-ES
// get best estimator for the optimum, xmean
xfinal = evo.getNew(CMAES<double>::XMean); // "XBestEver" might be used as well
// do something with final solution and finally release memory
delete[] xfinal;
return 0;
}
