[BUG] cmaes & xnes do not honor bounds

[jschueller]

it seems cmaes (and xnes) does not care about the bounds of the problem as the final population ends up way outside eventhough the initial population is inside the bounds:

#include <pagmo/algorithms/cmaes.hpp>
#include <pagmo/population.hpp>
#include <pagmo/problems/rosenbrock.hpp>

using namespace pagmo;

struct problem1 {
    vector_double fitness(const vector_double &x) const
    {
        double x1=x[0];
        double x2=x[1];
        return {x2};
//         return {1.+100.*(x2-x1*x1)*(x2-x1*x1)+(1.-x1)*(1.-x1)};
    }
    std::pair<vector_double, vector_double> get_bounds() const
    {
        return {{-1.5, -1.5}, {1.5, 1.5}};
    }
    vector_double::size_type get_nobj() const
    {
        return 1;
    }
};

int main()
{
  try {
    problem prob{problem1{}};

    population pop1{prob, 0, 0};
    int size = 100;
    std::uniform_real_distribution<double> unif(-1.5, 1.5);
    std::default_random_engine re;
    for (int i=0; i<size; ++i)
    {
      vector_double x(2);
      x[0] = unif(re);
      x[1] = unif(re);
      pop1.push_back(x);
    }

    cmaes user_algo1{10u};
    pop1 = user_algo1.evolve(pop1);
    for (int i=0; i<size; ++i)
    {
      vector_double x = pop1.get_x()[i];
      std::cout << "x1="<<x[0]<<" x2="<<x[1] <<std::endl;
    }

    }
    catch(const std::exception & exc) {
        std::cout << "ex="<<exc.what() <<std::endl;
        return 2;
    }
    return 0;
}

...
x1=-1495.25 x2=-8833.04
x1=-731.242 x2=-4187.9
x1=-1317.33 x2=-8026.68
x1=-5534.71 x2=-34318.1
x1=1402.88 x2=8971.83
x1=970.03 x2=5018.52

[bluescarni]

I am going to ping @darioizzo about this.

But as far as I remember, in pagmo it is unspecified at this time if the bounds are always honoured strictly (i.e., it is algorithm-dependent). Some algorithms benefit from being able to explore the search space slightly outside the bounds, but I do not know if that is the case for cmaes/xnes.

[jschueller]

it seems force_bounds make take the bounds into account but the behavior is inconsistent with the other algorithms if it defaults to false