Fix MIP dual value behaviour

optlang/cplex_interface.py

@@ -132,6 +132,12 @@
     [(val, key) for key, val in six.iteritems(_CPLEX_VTYPE_TO_VTYPE)]
 )
 
+_CPLEX_MIP_TYPES_TO_CONTINUOUS = {
+    cplex.Cplex.problem_type.MILP: cplex.Cplex.problem_type.LP,
+    cplex.Cplex.problem_type.MIQP: cplex.Cplex.problem_type.QP,
+    cplex.Cplex.problem_type.MIQCP: cplex.Cplex.problem_type.QCP
+}
+
 
 def _constraint_lb_and_ub_to_cplex_sense_rhs_and_range_value(lb, ub):
     """Helper function used by Constraint and Model"""
@@ -175,6 +181,12 @@ def type(self, value):
                     ", ".join(_VTYPE_TO_CPLEX_VTYPE.keys())
                 )
             self.problem.problem.variables.set_types(self.name, cplex_kind)
+            if value == "continuous" and not self.problem.is_integer:
+                print("hmm")
+                cplex_type = self.problem.problem.get_problem_type()
+                if cplex_type in _CPLEX_MIP_TYPES_TO_CONTINUOUS:
+                    print("yes")
+                    self.problem.problem.set_problem_type(_CPLEX_MIP_TYPES_TO_CONTINUOUS[cplex_type])
         super(Variable, Variable).type.fset(self, value)
 
     def _get_primal(self):
@@ -184,8 +196,8 @@ def _get_primal(self):
     @property
     def dual(self):
         if self.problem is not None:
-            if self.problem.problem.get_problem_type() == self.problem.problem.problem_type.MILP:  # cplex cannot determine reduced costs for MILP problems ...
-                return None
+            if self.problem.is_integer:
+                raise ValueError("Dual values are not well-defined for integer problems")
             return self.problem.problem.solution.get_reduced_costs(self.name)
         else:
             return None
@@ -252,22 +264,23 @@ def primal(self):
     @property
     def dual(self):
         if self.problem is not None:
+            if self.problem.is_integer:
+                raise ValueError("Dual values are not well-defined for integer problems")
             return self.problem.problem.solution.get_dual_values(self.name)
         else:
             return None
 
     @interface.Constraint.name.setter
     def name(self, value):
-        old_name = getattr(self, 'name', None)
-        self._name = value
         if getattr(self, 'problem', None) is not None:
             if self.indicator_variable is not None:
                 raise NotImplementedError(
                     "Unfortunately, the CPLEX python bindings don't support changing an indicator constraint's name"
                 )
             else:
-                self.problem.problem.linear_constraints.set_names(old_name, value)
-            self.problem.constraints.update_key(old_name)
+                self.problem.problem.linear_constraints.set_names(self.name, value)
+            self.problem.constraints.update_key(self.name)
+        self._name = value
 
     @interface.Constraint.lb.setter
     def lb(self, value):
@@ -705,8 +718,11 @@ def primal_values(self):
 
     @property
     def reduced_costs(self):
-        return collections.OrderedDict(
-            zip((variable.name for variable in self.variables), self.problem.solution.get_reduced_costs()))
+        if self.is_integer:
+            raise ValueError("Dual values are not well-defined for integer problems")
+        else:
+            return collections.OrderedDict(
+                zip((variable.name for variable in self.variables), self.problem.solution.get_reduced_costs()))
 
     @property
     def constraint_values(self):
@@ -716,8 +732,15 @@ def constraint_values(self):
 
     @property
     def shadow_prices(self):
-        return collections.OrderedDict(
-            zip((constraint.name for constraint in self.constraints), self.problem.solution.get_dual_values()))
+        if self.is_integer:
+            raise ValueError("Dual values are not well-defined for integer problems")
+        else:
+            return collections.OrderedDict(
+                zip((constraint.name for constraint in self.constraints), self.problem.solution.get_dual_values()))
+
+    @property
+    def is_integer(self):
+        return (self.problem.variables.get_num_integer() + self.problem.variables.get_num_binary()) > 0
 
     def to_lp(self):
         self.update()

optlang/glpk_interface.py

@@ -132,6 +132,8 @@ def _get_primal(self):
     @property
     def dual(self):
         if self.problem:
+            if self.problem.is_integer:
+                raise ValueError("Dual values are not well-defined for integer problems")
             return glp_get_col_dual(self.problem.problem, self._index)
         else:
             return None
@@ -228,6 +230,8 @@ def primal(self):
     @property
     def dual(self):
         if self.problem is not None:
+            if self.problem.is_integer:
+                raise ValueError("Dual values are not well-defined for integer problems")
             return glp_get_row_dual(self.problem.problem, self._index)
         else:
             return None
@@ -593,14 +597,11 @@ def primal_values(self):
 
     @property
     def reduced_costs(self):
-        reduced_costs = collections.OrderedDict()
-        is_mip = self._glpk_is_mip()
-        for index, variable in enumerate(self.variables):
-            if is_mip:
-                value = None
-            else:
-                value = glp_get_col_dual(self.problem, index + 1)
-            reduced_costs[variable.name] = value
+        if self.is_integer:
+            raise ValueError("Dual values are not well-defined for integer problems")
+        reduced_costs = collections.OrderedDict(
+            (var.name, glp_get_col_dual(self.problem, index + 1)) for index, var in enumerate(self.variables)
+        )
         return reduced_costs
 
     @property
@@ -617,14 +618,11 @@ def constraint_values(self):
 
     @property
     def shadow_prices(self):
-        is_mip = self._glpk_is_mip()
-        shadow_prices = collections.OrderedDict()
-        for index, constraint in enumerate(self.constraints):
-            if is_mip:
-                value = None
-            else:
-                value = glp_get_row_dual(self.problem, index + 1)
-            shadow_prices[constraint.name] = value
+        if self.is_integer:
+            raise ValueError("Dual values are not well-defined for integer problems")
+        shadow_prices = collections.OrderedDict(
+            (constraint.name, glp_get_row_dual(self.problem, index + 1)) for index, constraint in enumerate(self.constraints)
+        )
         return shadow_prices
 
     def to_lp(self):
@@ -776,6 +774,10 @@ def _glpk_set_col_bounds(self, variable):
     def _glpk_is_mip(self):
         return glp_get_num_int(self.problem) > 0
 
+    @property
+    def is_integer(self):
+        return self._glpk_is_mip()
+
     def _glpk_set_row_bounds(self, constraint):
         if constraint.lb is None and constraint.ub is None:
             # 0.'s are ignored

optlang/gurobi_interface.py

@@ -166,6 +166,8 @@ def primal(self):
     @property
     def dual(self):
         if self.problem:
+            if self.problem.is_integer:
+                raise ValueError("Dual values are not well-defined for integer problems")
             return self._internal_variable.getAttr('RC')
         else:
             return None
@@ -253,6 +255,8 @@ def primal(self):
     @property
     def dual(self):
         if self.problem is not None:
+            if self.problem.is_integer:
+                raise ValueError("Dual values are not well-defined for integer problems")
             return self._internal_constraint.Pi
         else:
             return None
@@ -638,6 +642,11 @@ def _remove_constraints(self, constraints):
         for internal_constraint in internal_constraints:
             self.problem.remove(internal_constraint)
 
+    @property
+    def is_integer(self):
+        self.problem.update()
+        return self.problem.NumIntVars > 0
+
 
 if __name__ == '__main__':
     x = Variable('x', lb=0, ub=10)

optlang/interface.py

@@ -1246,6 +1246,10 @@ def shadow_prices(self):
         # Fallback, if nothing faster is available
         return collections.OrderedDict([(constraint.name, constraint.dual) for constraint in self.constraints])
 
+    @property
+    def is_integer(self):
+        return any(var.type in ("integer", "binary") for var in self.variables)
+
     def __str__(self):
         if hasattr(self, "to_lp"):
             return self.to_lp()

optlang/tests/abstract_test_cases.py

@@ -26,6 +26,7 @@
 import copy
 import os
 import sympy
+from functools import partial
 
 __test__ = False
 
@@ -697,6 +698,82 @@ def test_constraint_set_linear_coefficients(self):
         self.assertEqual(model.optimize(), optlang.interface.OPTIMAL)
         self.assertAlmostEqual(x.primal, 0.5)
 
+    def test_is_integer(self):
+        model = self.model
+        self.assertFalse(model.is_integer)
+
+        model.variables[0].type = "integer"
+        self.assertTrue(model.is_integer)
+
+        model.variables[0].type = "continuous"
+        model.variables[1].type = "binary"
+        self.assertTrue(model.is_integer)
+
+        model.variables[1].type = "continuous"
+        self.assertFalse(model.is_integer)
+
+    def test_integer_variable_dual(self):
+        model = self.interface.Model()
+        x = self.interface.Variable("x", lb=0)
+        y = self.interface.Variable("y", lb=0)
+        c = self.interface.Constraint(x + y, ub=1)
+        model.add(c)
+        model.objective = self.interface.Objective(x)
+
+        model.optimize()
+        self.assertEqual(y.dual, -1)
+
+        x.type = "integer"
+        model.optimize()
+        self.assertRaises(ValueError, partial(getattr, y, "dual"))
+
+        x.type = "continuous"
+        model.optimize()
+        self.assertEqual(y.dual, -1)
+
+    def test_integer_constraint_dual(self):
+        model = self.interface.Model()
+        x = self.interface.Variable("x")
+        c = self.interface.Constraint(x, ub=1)
+        model.add(c)
+        model.objective = self.interface.Objective(x)
+
+        model.optimize()
+        self.assertEqual(c.dual, 1)
+
+        x.type = "integer"
+        model.optimize()
+        self.assertRaises(ValueError, partial(getattr, c, "dual"))
+
+        x.type = "continuous"
+        model.optimize()
+        self.assertEqual(c.dual, 1)
+
+    def test_integer_batch_duals(self):
+        model = self.interface.Model()
+        x = self.interface.Variable("x")
+        c = self.interface.Constraint(x, ub=1)
+        model.add(c)
+        model.objective = self.interface.Objective(x)
+        model.optimize()
+
+        self.assertEqual(model.reduced_costs[x.name], 0)
+        self.assertEqual(model.shadow_prices[c.name], 1)
+
+        x.type = "integer"
+        model.optimize()
+
+        with self.assertRaises(ValueError):
+            model.reduced_costs
+        with self.assertRaises(ValueError):
+            model.shadow_prices
+
+        x.type = "continuous"
+        model.optimize()
+
+        self.assertEqual(model.reduced_costs[x.name], 0)
+        self.assertEqual(model.shadow_prices[c.name], 1)
+
 
 @six.add_metaclass(abc.ABCMeta)
 class AbstractConfigurationTestCase(unittest.TestCase):

optlang/tests/test_scipy_interface.py

@@ -288,3 +288,15 @@ def test_scipy_coefficient_dict(self):
         model.add(c)
         model.objective = obj
         self.assertEqual(model.optimize(), optlang.interface.OPTIMAL)
+
+    def test_is_integer(self):
+        self.skipTest("No integers with scipy")
+
+    def test_integer_variable_dual(self):
+        self.skipTest("No duals with scipy")
+
+    def test_integer_constraint_dual(self):
+        self.skipTest("No duals with scipy")
+
+    def test_integer_batch_duals(self):
+        self.skipTest("No duals with scipy")