[feature] wrapping cost function to return best value explored

src/orquestra/opt/optimizers/scipy_optimizer.py

@@ -14,6 +14,77 @@
 )
 from ..history.recorder import RecorderFactory
 from ..history.recorder import recorder as _recorder
+from .pso.continuous_pso_optimizer import _get_bounds_like_array
+
+
+class _CostFunctionWithBestValueType(type):
+    def __instancecheck__(cls, obj: object) -> bool:
+        if (
+            isinstance(obj, Callable)
+            and hasattr(obj, "best_value")
+            and hasattr(obj, "best_params")
+        ):
+            return True
+        return False
+
+
+class _CostFunctionWithBestValue(metaclass=_CostFunctionWithBestValueType):
+    best_value: float = np.inf
+    best_params: np.ndarray = np.empty(1)
+
+    def __init__(
+        self,
+        cost_function: Union[CallableWithGradient, Callable],
+        constraints: Sequence[Dict[str, Callable]],
+        bounds: Union[
+            scipy.optimize.Bounds,
+            Sequence[Tuple[float, float]],
+            None,
+        ] = None,
+    ) -> None:
+        self.cost_function = cost_function
+        # Inherit all attributes from the cost function
+        for attr in dir(cost_function):
+            if not attr.startswith("__"):
+                setattr(self, attr, getattr(cost_function, attr))
+        self.best_params.fill(np.nan)
+        self.constraints = constraints
+        self.bounds = _get_bounds_like_array(bounds) if bounds is not None else None
+
+    def __call__(self, parameters: np.ndarray) -> float:
+        value = self.cost_function(parameters)
+        if (
+            value < self.best_value
+            and self._are_params_bounded(parameters)
+            and self._are_params_constrained(parameters)
+        ):
+            self.best_value = value
+            self.best_params = parameters
+        return value
+
+    def _are_params_bounded(self, params: np.ndarray) -> bool:
+        if self.bounds:
+            return bool(
+                np.all(
+                    np.bitwise_and(params >= self.bounds[0], params <= self.bounds[1])
+                )
+            )
+        return True
+
+    def _are_params_constrained(self, params: np.ndarray) -> bool:
+        if self.constraints:
+            for constraint in self.constraints:
+                constraint_args = constraint.get("args", ())
+                assert isinstance(
+                    constraint_args, tuple
+                ), "If you pass args to the constraint, they must be a tuple."
+                if constraint["type"] == "eq":
+                    if not np.isclose(constraint["fun"](params, *constraint_args), 0):
+                        return False
+                elif constraint["type"] == "ineq":
+                    if constraint["fun"](params, *constraint_args) < 0:
+                        return False
+        return True
 
 
 class ScipyOptimizer(Optimizer):
@@ -50,6 +121,11 @@ def __init__(
         self.constraints = [] if constraints is None else constraints
         self.bounds = bounds
 
+    def _preprocess_cost_function(
+        self, cost_function: Union[CallableWithGradient, Callable]
+    ) -> _CostFunctionWithBestValue:
+        return _CostFunctionWithBestValue(cost_function, self.constraints, self.bounds)
+
     def _minimize(
         self,
         cost_function: Union[CallableWithGradient, Callable],
@@ -66,7 +142,7 @@ def _minimize(
                 evaluations should be recorded.
 
         """
-
+        assert isinstance(cost_function, _CostFunctionWithBestValue)
         jacobian = None
         if isinstance(cost_function, CallableWithGradient) and callable(
             getattr(cost_function, "gradient")
@@ -82,8 +158,8 @@ def _minimize(
             bounds=self.bounds,
             jac=jacobian,
         )
-        opt_value = result.fun
-        opt_params = result.x
+        opt_value = cost_function.best_value
+        opt_params = cost_function.best_params
 
         nit = result.get("nit", None)
         nfev = result.get("nfev", None)

tests/orquestra/opt/optimizers/scipy_optimizer_test.py

@@ -20,6 +20,7 @@
         {"method": "L-BFGS-B"},
         {"method": "Nelder-Mead"},
         {"method": "SLSQP"},
+        {"method": "Powell"},
         {"method": "COBYLA", "options": {"maxiter": 50000, "tol": 1e-7}},
     ]
 )
@@ -89,7 +90,7 @@ def test_SLSQP_with_inequality_constraints(self):
         cost_function = FunctionWithGradient(
             rosenbrock_function, finite_differences_gradient(rosenbrock_function)
         )
-        constraints = {"type": "ineq", "fun": lambda x: x[0] + x[1] - 3}
+        constraints = ({"type": "ineq", "fun": lambda x: x[0] + x[1] - 3},)
         optimizer = ScipyOptimizer(method="SLSQP")
         initial_params = np.array([0, 0])
 
@@ -107,3 +108,30 @@ def test_SLSQP_with_inequality_constraints(self):
             results_with_constraints.opt_value, abs=1e-1
         )
         assert results_with_constraints.opt_params.sum() >= 3
+
+    # https://github.com/scipy/scipy/issues/17673 reports a function and that is
+    # minimised with an initial point that has a given cost function value which is
+    # lower than the optimised value by the Powell optimizer. This test is to ensure
+    # that our solution over scipy is not affected by this issue.
+    @pytest.mark.parametrize("method", ["Powell", "L-BFGS-B", "Nelder-Mead", "SLSQP"])
+    def test_Langermann_function(self, method):
+        # Definition of the function
+        langer_c = np.array([6, 1, 4, 4, 8])
+        langer_A = np.array(
+            [[4, 6, 3, 5], [8, 7, 9, 9], [2, 7, 8, 8], [9, 2, 6, 9], [5, 4, 1, 4]]
+        )
+
+        def langermann(parameters: np.ndarray) -> float:
+            # This function is bounded in the [0, 10] box
+            sum_ = np.sum(np.subtract(langer_A, parameters) ** 2, axis=1)
+            vec = np.exp(-1 / np.pi * sum_) * np.cos(np.pi * sum_)
+            result = np.dot(langer_c, vec)
+            return result
+
+        # Initial point
+        x0 = np.array([4.65116802, 4.42985893, 1.74720157, 4.29727392])
+        y0 = langermann(x0)
+
+        optimiser = ScipyOptimizer(method=method, bounds=[(0, 10)] * 4)
+        result = optimiser.minimize(langermann, x0)
+        assert result.opt_value <= y0