Adjust subsolver module to components interface

src/estimagic/optimization/pounders_auxiliary.py

@@ -290,10 +290,11 @@ def solve_subproblem(
             "gtol_abs_conjugate_gradient": gtol_abs_conjugate_gradient,
             "gtol_rel_conjugate_gradient": gtol_rel_conjugate_gradient,
         }
-        result = bntr(main_model, lower_bounds, upper_bounds, **options)
+        result = bntr(main_model, lower_bounds, upper_bounds, x_candidate=x0, **options)
     elif solver == "gqtpar":
         result = gqtpar(
             main_model,
+            x_candidate=x0,
             k_easy=k_easy,
             k_hard=k_hard,
             maxiter=maxiter,

src/estimagic/optimization/subsolvers/bntr.py

@@ -26,6 +26,7 @@ def bntr(
     model,
     lower_bounds,
     upper_bounds,
+    x_candidate,
     *,
     conjugate_gradient_method,
     maxiter,
@@ -61,6 +62,7 @@ def bntr(
             for the parameter vector x.
         upper_bounds (np.ndarray): 1d array of shape (n,) with upper bounds
             for the parameter vector x.
+        x_candidate (np.ndarray): Initial guess for the solution of the subproblem.
         conjugate_gradient_method (str): Method for computing the conjugate gradient
             step. Available conjugate gradient methods are:
                 - "cg"
@@ -105,8 +107,6 @@ def bntr(
         "default_radius": 100.00,
     }
 
-    x_candidate = np.zeros_like(model.linear_terms)
-
     (
         x_candidate,
         f_candidate,

src/estimagic/optimization/subsolvers/bntr_fast.py

@@ -18,6 +18,7 @@ def bntr_fast(
     model,
     lower_bounds,
     upper_bounds,
+    x_candidate,
     *,
     conjugate_gradient_method,
     maxiter,
@@ -56,6 +57,7 @@ def bntr_fast(
             for the parameter vector x.
         upper_bounds (np.ndarray): 1d array of shape (n,) with upper bounds
             for the parameter vector x.
+        x_candidate (np.ndarray): Initial guess for the solution of the subproblem.
         conjugate_gradient_method (str): Method for computing the conjugate gradient
             step. Available conjugate gradient methods are:
                 - "cg"
@@ -99,6 +101,7 @@ def bntr_fast(
         model_hessian=model_hessian,
         lower_bounds=lower_bounds,
         upper_bounds=upper_bounds,
+        x_candidate=x_candidate,
         conjugate_gradient_method=conjugate_gradient_method,
         maxiter=maxiter,
         maxiter_gradient_descent=maxiter_gradient_descent,
@@ -126,6 +129,7 @@ def _bntr_fast_jitted(
     model_hessian,
     lower_bounds,
     upper_bounds,
+    x_candidate,
     conjugate_gradient_method,
     maxiter,
     maxiter_gradient_descent,
@@ -163,6 +167,7 @@ def _bntr_fast_jitted(
             for the parameter vector x.
         upper_bounds (np.ndarray): 1d array of shape (n,) with upper bounds
             for the parameter vector x.
+        x_candidate (np.ndarray): Initial guess for the solution of the subproblem.
         conjugate_gradient_method (str): Method for computing the conjugate gradient
             step. Available conjugate gradient methods are:
                 - "cg"
@@ -209,6 +214,7 @@ def _bntr_fast_jitted(
         model_hessian,
         lower_bounds,
         upper_bounds,
+        x_candidate,
         maxiter_gradient_descent,
         gtol_abs,
         gtol_rel,
@@ -341,6 +347,7 @@ def _take_preliminary_gradient_descent_step_and_check_for_solution(
     model_hessian,
     lower_bounds,
     upper_bounds,
+    x_candidate,
     maxiter_gradient_descent,
     gtol_abs,
     gtol_rel,
@@ -357,6 +364,7 @@ def _take_preliminary_gradient_descent_step_and_check_for_solution(
             for the parameter vector x.
         upper_bounds (np.ndarray): 1d array of shape (n,) with upper bounds
             for the parameter vector x.
+        x_candidate (np.ndarray): Initial guess for the solution of the subproblem.
         maxiter_gradient_descent (int): Maximum number of iterations in performing
             gradient descent step
         gtol_abs (float): Convergence tolerance for the absolute gradient norm.
@@ -384,8 +392,6 @@ def _take_preliminary_gradient_descent_step_and_check_for_solution(
     converged = False
     convergence_reason = 0
 
-    x_candidate = np.zeros(len(model_gradient))
-
     criterion_candidate = _evaluate_model_criterion(
         x_candidate, model_gradient, model_hessian
     )

src/estimagic/optimization/subsolvers/gqtpar.py

@@ -19,7 +19,7 @@ class DampingFactors(NamedTuple):
     upper_bound: Union[float, None] = None
 
 
-def gqtpar(model, *, k_easy=0.1, k_hard=0.2, maxiter=200):
+def gqtpar(model, x_candidate, *, k_easy=0.1, k_hard=0.2, maxiter=200):
     """Solve the quadratic trust-region subproblem via nearly exact iterative method.
 
     This subproblem solver is mainly based on Conn et al. (2000) "Trust region methods"
@@ -50,11 +50,10 @@ def gqtpar(model, *, k_easy=0.1, k_hard=0.2, maxiter=200):
     See pp. 194-197 in :cite:`Conn2000` for a more detailed description.
 
     Args:
-        main_model (NamedTuple): NamedTuple containing the parameters of the
-            main model, i.e.:
+        model (NamedTuple): NamedTuple containing the parameters of the main model, i.e.
             - ``linear_terms``, a np.ndarray of shape (n,) and
             - ``square_terms``, a np.ndarray of shape (n,n).
-        trustregion_radius (float): Trustregion radius, often referred to as delta.
+        x_candidate (np.ndarray): Initial guess for the solution of the subproblem.
         k_easy (float): topping criterion for the "easy" case.
         k_hard (float): Stopping criterion for the "hard" case.
         maxiter (int): Maximum number of iterations to perform. If reached,
@@ -69,8 +68,6 @@ def gqtpar(model, *, k_easy=0.1, k_hard=0.2, maxiter=200):
     """
     hessian_info = HessianInfo()
 
-    x_candidate = np.zeros_like(model.linear_terms)
-
     # Small floating point number signaling that for vectors smaller
     # than that backward substituition is not reliable.
     # See Golub, G. H., Van Loan, C. F. (2013), "Matrix computations", p.165.

src/estimagic/optimization/subsolvers/gqtpar_fast.py

@@ -5,7 +5,7 @@
 from scipy.linalg.lapack import dpotrf as compute_cholesky_factorization
 
 
-def gqtpar_fast(model, *, k_easy=0.1, k_hard=0.2, maxiter=200):
+def gqtpar_fast(model, x_candidate, *, k_easy=0.1, k_hard=0.2, maxiter=200):
     """Solve the quadratic trust-region subproblem via nearly exact iterative method.
 
     This subproblem solver is mainly based on Conn et al. (2000) "Trust region methods"
@@ -36,11 +36,10 @@ def gqtpar_fast(model, *, k_easy=0.1, k_hard=0.2, maxiter=200):
     See pp. 194-197 in :cite:`Conn2000` for a more detailed description.
 
     Args:
-        main_model (NamedTuple): NamedTuple containing the parameters of the
-            main model, i.e.:
+        model (NamedTuple): NamedTuple containing the parameters of the main model, i.e.
             - ``linear_terms``, a np.ndarray of shape (n,) and
             - ``square_terms``, a np.ndarray of shape (n,n).
-        trustregion_radius (float): Trustregion radius, often referred to as delta.
+        x_candidate (np.ndarray): Initial guess for the solution of the subproblem.
         k_easy (float): topping criterion for the "easy" case.
         k_hard (float): Stopping criterion for the "hard" case.
         maxiter (int): Maximum number of iterations to perform. If reached,
@@ -56,7 +55,6 @@ def gqtpar_fast(model, *, k_easy=0.1, k_hard=0.2, maxiter=200):
     hessian_already_factorized = False
     model_gradient = model.linear_terms
     model_hessian = model.square_terms
-    x_candidate = np.zeros(len(model_gradient))
 
     # Small floating point number signaling that for vectors smaller
     # than that backward substituition is not reliable.

src/estimagic/optimization/tranquilo/process_arguments.py

@@ -25,7 +25,6 @@
     get_default_radius_options,
     get_default_sample_size,
     get_default_search_radius_factor,
-    get_default_subsolver,
     update_option_bundle,
 )
 from estimagic.optimization.tranquilo.region import Region
@@ -156,13 +155,9 @@ def process_arguments(
     sample_points = get_sampler(sampler, sampler_options)
 
     solve_subproblem = get_subsolver(
-        solver=get_default_subsolver(
-            bounds=_bounds,
-            cube_subsolver=cube_subsolver,
-            sphere_subsolver=sphere_subsolver,
-        ),
+        cube_solver=cube_subsolver,
+        sphere_solver=sphere_subsolver,
         user_options=subsolver_options,
-        bounds=_bounds,
     )
 
     filter_points = get_sample_filter(