Merge pull request #2091 from numbbo/release

Bug fix of best_value computation on the constrained linear functions

code-experiments/build/python/example_experiment2.py

@@ -86,12 +86,13 @@ def random_search(f, lbounds, ubounds, evals):
 
 suite_name = "bbob"  # see cocoex.known_suite_names
 budget_multiplier = 2  # times dimension, increase to 10, 100, ...
-suite_filter_options = (""  # without filtering a suite has instance_indices 1-15
+suite_filter_options = (""  # without filtering, a suite has instance_indices 1-15
                         # "dimensions: 2,3,5,10,20 "  # skip dimension 40
                         # "instance_indices: 1-5 "  # relative to suite instances
-                        # "year:2019 "  # select instances by year
                        )
 # for more suite filter options see http://numbbo.github.io/coco-doc/C/#suite-parameters
+suite_year_option = ""  # "year: 2022"  # determine instances by year, not all years work for all suites :-(
+
 batches = 1  # number of batches, batch=3/32 works to set both, current_batch and batches
 current_batch = 1  # only current_batch modulo batches is relevant
 output_folder = ''
@@ -113,7 +114,7 @@ def random_search(f, lbounds, ubounds, evals):
     output_folder += "_batch%03dof%d" % (current_batch, batches)
 
 ### prepare
-suite = cocoex.Suite(suite_name, "", suite_filter_options)
+suite = cocoex.Suite(suite_name, suite_year_option, suite_filter_options)
 observer = cocoex.Observer(suite_name, "result_folder: " + output_folder)
 minimal_print = cocoex.utilities.MiniPrint()
 stoppings = defaultdict(list)  # dict of lists, key is the problem index

code-experiments/build/python/python/solvers.py

@@ -16,8 +16,9 @@ def random_search(fun, lbounds, ubounds, budget):
         # about five times faster than "for k in range(budget):..."
         X = lbounds + (ubounds - lbounds) * np.random.rand(chunk, dim)
         if fun.number_of_constraints > 0:
-            C = [fun.constraint(x) for x in X]  # call constraints
-            F = [fun(x) for i, x in enumerate(X) if np.all(C[i] <= 0)]
+            FC = [(fun(x), fun.constraint(x)) for x in X]  # call  objective and constraints
+            F = [fc[0] for fc in FC if all(fc[1] <= 0)]
+            budget -= chunk  # one more to account for constraint evals
         else:
             F = [fun(x) for x in X]
         if fun.number_of_objectives == 1:

code-experiments/src/suite_cons_bbob_problems.c

@@ -449,6 +449,12 @@ static coco_problem_t *f_linear_slope_c_linear_cons_bbob_problem_allocate(const
    */
   for (i = 0; i < dimension; ++i)
     problem->best_parameter[i] = 0.0;
+
+  /* Apply a translation to the whole problem so that the constrained 
+   * minimum is no longer at the origin.
+   */
+  problem = transform_vars_shift(problem, xopt, 1);
+
   assert(problem->evaluate_function != NULL);
   problem->evaluate_function(problem, problem->best_parameter, problem->best_value);
 
@@ -457,12 +463,7 @@ static coco_problem_t *f_linear_slope_c_linear_cons_bbob_problem_allocate(const
    */
   problem->evaluations = 0;
   problem->evaluations_constraints = 0;
-
-  /* Apply a translation to the whole problem so that the constrained 
-   * minimum is no longer at the origin.
-   */
-  problem = transform_vars_shift(problem, xopt, 1);
-
+
   /* Construct problem type */
   coco_problem_set_type(problem, "%s_%s", problem_type_temp, 
   problem_c->problem_type);

code-experiments/src/transform_vars_shift.c

@@ -107,6 +107,9 @@ static void transform_vars_shift_free(void *thing) {
 
 /**
  * @brief Creates the transformation.
+ * 
+ * CAVEAT: when shifting the constraint only, the best_value of best_parameter
+ *         will get in an inconsistent state.
  */
 static coco_problem_t *transform_vars_shift(coco_problem_t *inner_problem,
                                             const double *offset,