Reimplement output constraint scaling

src/ert/run_models/everest_run_model.py

@@ -34,6 +34,7 @@
 from everest.everest_storage import EverestStorage, OptimalResult
 from everest.optimizer.everest2ropt import everest2ropt
 from everest.optimizer.opt_model_transforms import (
+    ConstraintScaler,
     ObjectiveScaler,
     get_optimization_domain_transforms,
 )
@@ -223,20 +224,39 @@ def _init_transforms(self, variables: NDArray[np.float64]) -> OptModelTransforms
         transforms = get_optimization_domain_transforms(
             self._everest_config.controls,
             self._everest_config.objective_functions,
+            self._everest_config.output_constraints,
             realization_weights,
         )
         # If required, initialize auto-scaling:
         assert isinstance(transforms.objectives, ObjectiveScaler)
-        if transforms.objectives.has_auto_scale:
-            objectives, _, _ = self._run_forward_model(
+        assert transforms.nonlinear_constraints is None or isinstance(
+            transforms.nonlinear_constraints, ConstraintScaler
+        )
+        if transforms.objectives.has_auto_scale or (
+            transforms.nonlinear_constraints
+            and transforms.nonlinear_constraints.has_auto_scale
+        ):
+            # Run the forward model once to find the objective/constraint values
+            # to compute the scales. This will add an ensemble/batch in the
+            # storage that is not part of the optimization run. However, the
+            # results may be used in the optimization via the caching mechanism.
+            objectives, constraints, _ = self._run_forward_model(
                 np.repeat(np.expand_dims(variables, axis=0), nreal, axis=0),
                 realizations,
             )
-            transforms.objectives.calculate_auto_scales(objectives)
+            if transforms.objectives.has_auto_scale:
+                transforms.objectives.calculate_auto_scales(objectives)
+            if (
+                transforms.nonlinear_constraints
+                and transforms.nonlinear_constraints.has_auto_scale
+            ):
+                assert constraints is not None
+                transforms.nonlinear_constraints.calculate_auto_scales(constraints)
         return transforms
 
     def _create_optimizer(self) -> BasicOptimizer:
-        # Initialize the optimization model transforms:
+        # Initialize the optimization model transforms. This may run one initial
+        # ensemble for auto-scaling purposes:
         transforms = self._init_transforms(
             np.asarray(
                 FlattenedControls(self._everest_config.controls).initial_guesses,

src/everest/optimizer/everest2ropt.py

@@ -99,7 +99,6 @@ def _parse_objectives(objective_functions: list[ObjectiveFunctionConfig], ropt_c
 
 
 def _parse_input_constraints(
-    controls: FlattenedControls,
     input_constraints: list[InputConstraintConfig] | None,
     formatted_control_names: list[str],
     formatted_control_names_dotdash: list[str],
@@ -160,17 +159,13 @@ def _parse_output_constraints(
         return
 
     rhs_values: list[float] = []
-    scales: list[float] = []
-    auto_scale: list[bool] = []
     types: list[ConstraintType] = []
 
     def _add_output_constraint(
         rhs_value: float | None, constraint_type: ConstraintType, suffix=None
     ):
         if rhs_value is not None:
             rhs_values.append(rhs_value)
-            scales.append(constr.scale if constr.scale is not None else 1.0)
-            auto_scale.append(constr.auto_scale or False)
             types.append(constraint_type)
 
     for constr in output_constraints:
@@ -181,25 +176,16 @@ def _add_output_constraint(
             raise RuntimeError(
                 "output constraint error: target cannot be combined with bounds"
             )
+        _add_output_constraint(target, ConstraintType.EQ)
         _add_output_constraint(
-            target,
-            ConstraintType.EQ,
+            upper_bound, ConstraintType.LE, None if lower_bound is None else "upper"
         )
         _add_output_constraint(
-            upper_bound,
-            ConstraintType.LE,
-            None if lower_bound is None else "upper",
-        )
-        _add_output_constraint(
-            lower_bound,
-            ConstraintType.GE,
-            None if upper_bound is None else "lower",
+            lower_bound, ConstraintType.GE, None if upper_bound is None else "lower"
         )
 
     ropt_config["nonlinear_constraints"] = {
         "rhs_values": rhs_values,
-        "scales": scales,
-        "auto_scale": auto_scale,
         "types": types,
     }
 
@@ -350,12 +336,9 @@ def everest2ropt(
     """
     ropt_config: dict[str, Any] = {}
 
-    flattened_controls = FlattenedControls(ever_config.controls)
-
-    _parse_controls(flattened_controls, ropt_config)
+    _parse_controls(FlattenedControls(ever_config.controls), ropt_config)
     _parse_objectives(ever_config.objective_functions, ropt_config)
     _parse_input_constraints(
-        flattened_controls,
         ever_config.input_constraints,
         ever_config.formatted_control_names,
         ever_config.formatted_control_names_dotdash,

src/everest/optimizer/opt_model_transforms.py

@@ -2,10 +2,19 @@
 
 import numpy as np
 from numpy.typing import NDArray
+from ropt.enums import ConstraintType
 from ropt.transforms import OptModelTransforms
-from ropt.transforms.base import ObjectiveTransform, VariableTransform
-
-from everest.config import ControlConfig, ObjectiveFunctionConfig
+from ropt.transforms.base import (
+    NonLinearConstraintTransform,
+    ObjectiveTransform,
+    VariableTransform,
+)
+
+from everest.config import (
+    ControlConfig,
+    ObjectiveFunctionConfig,
+    OutputConstraintConfig,
+)
 from everest.config.utils import FlattenedControls
 
 
@@ -79,9 +88,58 @@ def has_auto_scale(self) -> bool:
         return bool(np.any(self._auto_scales))
 
 
+class ConstraintScaler(NonLinearConstraintTransform):
+    def __init__(
+        self, scales: list[float], auto_scales: list[bool], weights: list[float]
+    ) -> None:
+        self._scales = np.asarray(scales, dtype=np.float64)
+        self._auto_scales = np.asarray(auto_scales, dtype=np.bool_)
+        self._weights = np.asarray(weights, dtype=np.float64)
+
+    def transform_rhs_values(
+        self, rhs_values: NDArray[np.float64], types: NDArray[np.ubyte]
+    ) -> tuple[NDArray[np.float64], NDArray[np.ubyte]]:
+        def flip_type(constraint_type: ConstraintType) -> ConstraintType:
+            match constraint_type:
+                case ConstraintType.GE:
+                    return ConstraintType.LE
+                case ConstraintType.LE:
+                    return ConstraintType.GE
+                case _:
+                    return constraint_type
+
+        rhs_values = rhs_values / self._scales  # noqa: PLR6104
+        # Flip inequality types if self._scales < 0 in the division above:
+        types = np.fromiter(
+            (
+                flip_type(type_) if scale < 0 else type_
+                for type_, scale in zip(types, self._scales, strict=False)
+            ),
+            np.ubyte,
+        )
+        return rhs_values, types
+
+    def forward(self, constraints: NDArray[np.float64]) -> NDArray[np.float64]:
+        return constraints / self._scales
+
+    def backward(self, constraints: NDArray[np.float64]) -> NDArray[np.float64]:
+        return constraints * self._scales
+
+    def calculate_auto_scales(self, constraints: NDArray[np.float64]) -> None:
+        auto_scales = np.abs(
+            np.nansum(constraints * self._weights[:, np.newaxis], axis=0)
+        )
+        self._scales[self._auto_scales] *= auto_scales[self._auto_scales]
+
+    @property
+    def has_auto_scale(self) -> bool:
+        return bool(np.any(self._auto_scales))
+
+
 def get_optimization_domain_transforms(
     controls: list[ControlConfig],
     objectives: list[ObjectiveFunctionConfig],
+    constraints: list[OutputConstraintConfig] | None,
     weights: list[float],
 ) -> OptModelTransforms:
     flattened_controls = FlattenedControls(controls)
@@ -107,4 +165,19 @@ def get_optimization_domain_transforms(
             ],
             weights,
         ),
+        nonlinear_constraints=(
+            ConstraintScaler(
+                [
+                    1.0 if constraint.scale is None else constraint.scale
+                    for constraint in constraints
+                ],
+                [
+                    False if constraint.auto_scale is None else constraint.auto_scale
+                    for constraint in constraints
+                ],
+                weights,
+            )
+            if constraints
+            else None
+        ),
     )

tests/everest/snapshots/test_ropt_initialization/test_everest2ropt_snapshot/config_advanced.yml/ropt_config.json

@@ -51,10 +51,10 @@
     "function_estimators": null,
     "realization_filters": null,
     "rhs_values": [
-      0.1
+      1.0
     ],
     "scales": [
-      0.1
+      1.0
     ],
     "types": [
       2

tests/everest/test_math_func.py

@@ -1,6 +1,7 @@
 import os
 from pathlib import Path
 
+import numpy as np
 import pytest
 import yaml
 
@@ -181,3 +182,41 @@ def test_math_func_auto_scaled_objectives(copy_math_func_test_data_to_tmp):
     total = -(expected_p * 0.5 + expected_q * 0.25) / (0.5 + 0.25)
 
     assert total == optim
+
+
+@pytest.mark.integration_test
+def test_math_func_auto_scaled_constraints(copy_math_func_test_data_to_tmp):
+    config = EverestConfig.load_file("config_advanced.yml")
+    config_dict = config.model_dump(exclude_none=True)
+
+    # control number of batches, no need for full convergence:
+    config_dict["optimization"]["convergence_tolerance"] = 1e-10
+    config_dict["optimization"]["max_batch_num"] = 3
+
+    # Run with auto_scaling:
+    config_dict["environment"]["output_folder"] = "output_auto_scale"
+    config_dict["output_constraints"][0]["auto_scale"] = True
+    config_dict["output_constraints"][0]["scale"] = 1.0
+    config = EverestConfig.model_validate(config_dict)
+    run_model = EverestRunModel.create(config)
+    evaluator_server_config = EvaluatorServerConfig()
+    run_model.run_experiment(evaluator_server_config)
+    result1 = run_model.result
+
+    # Run the equivalent without auto-scaling:
+    config_dict["environment"]["output_folder"] = "output_manual_scale"
+    config_dict["output_constraints"][0]["auto_scale"] = False
+    config_dict["output_constraints"][0]["scale"] = 0.25  # x(0)
+    # We need one batch less, no auto-scaling:
+    config_dict["optimization"]["max_batch_num"] -= 1
+    config = EverestConfig.model_validate(config_dict)
+    run_model = EverestRunModel.create(config)
+    evaluator_server_config = EvaluatorServerConfig()
+    run_model.run_experiment(evaluator_server_config)
+    result2 = run_model.result
+
+    assert result1.total_objective == pytest.approx(result2.total_objective)
+    assert np.allclose(
+        np.fromiter(result1.controls.values(), dtype=np.float64),
+        np.fromiter(result2.controls.values(), dtype=np.float64),
+    )

tests/everest/test_ropt_initialization.py

@@ -48,6 +48,7 @@ def test_everest2ropt_controls_auto_scale():
         transforms=get_optimization_domain_transforms(
             config.controls,
             config.objective_functions,
+            config.output_constraints,
             config.model.realizations_weights,
         ),
     )
@@ -65,6 +66,7 @@ def test_everest2ropt_variables_auto_scale():
         transforms=get_optimization_domain_transforms(
             config.controls,
             config.objective_functions,
+            config.output_constraints,
             config.model.realizations_weights,
         ),
     )
@@ -136,6 +138,7 @@ def test_everest2ropt_controls_input_constraint_auto_scale():
         transforms=get_optimization_domain_transforms(
             config.controls,
             config.objective_functions,
+            config.output_constraints,
             config.model.realizations_weights,
         ),
     )
@@ -280,6 +283,7 @@ def test_everest2ropt_snapshot(case, snapshot):
         transforms=get_optimization_domain_transforms(
             config.controls,
             config.objective_functions,
+            config.output_constraints,
             config.model.realizations_weights,
         ),
     ).model_dump()