cstrs_self_adaptive::cstrs_self_adaptive(const base &problem):
base_meta(
problem,
problem.get_dimension(),
problem.get_i_dimension(),
problem.get_f_dimension(),
0,
0,
std::vector<double>()),
m_apply_penalty_1(false),
m_scaling_factor(0.0),
m_c_scaling(problem.get_c_dimension(),0.0),
m_f_hat_down(problem.get_f_dimension(),0.0),
m_f_hat_up(problem.get_f_dimension(),0.0),
m_f_hat_round(problem.get_f_dimension(),0.0),
m_i_hat_down(0.0),
m_i_hat_up(0.0),
m_i_hat_round(0.0),
m_map_fitness(),
m_map_constraint(),
m_decision_vector_hash()
{
population pop(*m_original_problem,0);
if(m_original_problem->get_c_dimension() <= 0){
pagmo_throw(value_error,"The original problem has no constraints.");
}
// check that the dimension of the problem is 1
if (m_original_problem->get_f_dimension() != 1) {
pagmo_throw(value_error,"The original fitness dimension of the problem must be one, multi objective problems can't be handled with self adaptive meta problem.");
}
update_penalty_coeff(pop);
}
/**
* Constructor using std::vector (for python exposition purposes)
*
* @param[in] p base::problem to be rotated
* @param[in] rotation std::vector<std::vector<double> > expressing the problem rotation
*
* @see problem::base constructors.
*/
rotated::rotated(const base &p,
const std::vector<std::vector<double> > &rotation):
base_meta(
p,
p.get_dimension(),
p.get_i_dimension(),
p.get_f_dimension(),
p.get_c_dimension(),
p.get_ic_dimension(),
p.get_c_tol()),
m_Rotate(),m_normalize_translation(), m_normalize_scale()
{
if(!(rotation.size()==get_dimension())){
pagmo_throw(value_error,"The input matrix dimensions seem incorrect");
}
if(p.get_i_dimension()>0){
pagmo_throw(value_error,"Input problem has an integer dimension. Cannot rotate it.");
}
m_Rotate.resize(rotation.size(),rotation.size());
for (base::size_type i = 0; i < rotation.size(); ++i) {
if(!(rotation.size()==rotation[i].size())){
pagmo_throw(value_error,"The input matrix seems not to be square");
}
for (base::size_type j = 0; j < rotation[i].size(); ++j) {
m_Rotate(i,j) = rotation[i][j];
}
}
m_InvRotate = m_Rotate.transpose();
Eigen::MatrixXd check = m_InvRotate * m_Rotate;
if(!check.isIdentity(1e-5)){
pagmo_throw(value_error,"The input matrix seems not to be orthonormal (to a tolerance of 1e-5)");
}
configure_new_bounds();
}
robust::robust(const base & p, unsigned int trials, const double param_rho, unsigned int seed):
base_stochastic((int)p.get_dimension(),
p.get_i_dimension(),
p.get_f_dimension(),
p.get_c_dimension(),
p.get_ic_dimension(),
p.get_c_tol(), seed),
m_original_problem(p.clone()),
m_normal_dist(0, 1),
m_uniform_dist(0, 1),
m_trials(trials),
m_rho(param_rho)
{
if(param_rho < 0){
pagmo_throw(value_error, "Rho should be greater than 0");
}
set_bounds(p.get_lb(),p.get_ub());
}
noisy::noisy(const base & p, unsigned int trials, const double param_first, const double param_second, noise_type distribution, unsigned int seed):
base_stochastic((int)p.get_dimension(),
p.get_i_dimension(),
p.get_f_dimension(),
p.get_c_dimension(),
p.get_ic_dimension(),
p.get_c_tol(), seed),
m_original_problem(p.clone()),
m_trials(trials),
m_normal_dist(0.0,1.0),
m_uniform_dist(0.0,1.0),
m_decision_vector_hash(),
m_param_first(param_first),
m_param_second(param_second),
m_noise_type(distribution)
{
if(distribution == UNIFORM && param_first > param_second){
pagmo_throw(value_error, "Bounds specified for the uniform noise are not valid.");
}
set_bounds(p.get_lb(),p.get_ub());
}
rotated::rotated(const base &p, const Eigen::MatrixXd &rotation ):
base_meta(
p,
p.get_dimension(),
p.get_i_dimension(),
p.get_f_dimension(),
p.get_c_dimension(),
p.get_ic_dimension(),
p.get_c_tol()),
m_Rotate(rotation), m_normalize_translation(), m_normalize_scale()
{
m_InvRotate = m_Rotate.transpose();
Eigen::MatrixXd check = m_InvRotate * m_Rotate;
if(!check.isIdentity(1e-5)){
pagmo_throw(value_error,"The input matrix seems not to be orthonormal (to a tolerance of 1e-5)");
}
if(p.get_i_dimension()>0){
pagmo_throw(value_error,"Input problem has an integer dimension. Cannot rotate it.");
}
configure_new_bounds();
}
/**
* Constructor
*
* @param[in] p base::problem to be decomposed
* @param[in] method decomposition method (WEIGHTS, TCHEBYCHEFF, BI)
* @param[in] weights the weight vector (by default is set to random weights)
* @param[in] z ideal reference point (used in Tchebycheff and Boundary Intersection (BI) methods, by default it is set to 0)
* @param[in] adapt_ideal if true it updates the ideal reference point each time the objective function is called checking if the computed fitness is better (assumes minimization)
* @see For the uniform random generation of weights vector see Appendix 2 in "A. Jaszkiewicz - On the Performance of Multiple-Objective Genetic Local Search
on the 0/1 Knapsack Problem—A Comparative Experiment"
* @see For the different decomposition methods see "Q. Zhang - MOEA/D: A Multiobjective Evolutionary Algorithm Based on Decomposition"
*/
decompose::decompose(const base & p, method_type method, const std::vector<double> & weights, const std::vector<double> & z, const bool adapt_ideal):
base_meta(
p,
p.get_dimension(), // Ambiguous without the cast ...
p.get_i_dimension(),
1, //it transforms the problem into a single-objective problem
p.get_c_dimension(),
p.get_ic_dimension(),
p.get_c_tol()),
m_method(method),
m_weights(weights),
m_z(z),
m_adapt_ideal(adapt_ideal)
{
//0 - Check whether method is implemented
if(m_method != WEIGHTED && m_method != TCHEBYCHEFF && m_method != BI) {
pagmo_throw(value_error,"non existing decomposition method");
}
if (p.get_f_dimension() == 1) {
pagmo_throw(value_error,"decompose works only for multi-objective problems, you are trying to decompose a single objective one.");
}
//1 - Checks whether the weight vector has a dimension, if not, sets its default value
if (m_weights.size() == 0) {
//Initialise a random weight vector
rng_double m_drng = rng_generator::get<rng_double>();
m_weights = std::vector<double>((int)p.get_f_dimension(), 0.0);
double sum = 0;
for(std::vector<double>::size_type i = 0; i<m_weights.size(); ++i) {
m_weights[i] = (1-sum) * (1 - pow(boost::uniform_real<double>(0,1)(m_drng), 1.0 / (m_weights.size() - i - 1)));
sum += m_weights[i];
}
} else {
//1.1 - Checks whether the weight has lenght equal to the fitness size
if (m_weights.size() != p.get_f_dimension()) {
pagmo_throw(value_error,"the weight vector must have length equal to the fitness size");
}
//1.2 - Checks whether the weight vector sums to 1
double sum = 0.0;
for (std::vector<double>::size_type i=0; i<m_weights.size(); ++i) {
sum += m_weights[i];
}
if (fabs(sum-1.0) > 1E-8) {
pagmo_throw(value_error,"the weight vector should sum to 1 with a tolerance of E1-8");
}
//1.4 - Checks that all weights are positive
for (std::vector<double>::size_type i=0; i<m_weights.size(); ++i) {
if (m_weights[i] < 0) {
pagmo_throw(value_error,"the weight vector should contain only positive values");
}
}
}
//2 - Checks whether the reference point has a dimension, if not, sets its default value m_z = (0, ..., 0)
if (m_z.size() == 0) {
m_z = std::vector<double>((int)p.get_f_dimension(), 0.0); //by default m_z = (0, ..., 0)
} else {
//2.1 - Checks whether the reference point has lenght equal to the fitness size
if (m_z.size() != p.get_f_dimension()) {
pagmo_throw(value_error,"the the reference point vector must have equal length to the fitness size");
}
}
}