RelaxationComponent: Implement check_dropped

miplearn/components/relaxation.py

@@ -6,36 +6,59 @@ import logging
 import sys
 from copy import deepcopy
 
-import numpy as np
-from miplearn.components import classifier_evaluation_dict
 from tqdm import tqdm
 
 from miplearn import Component
 from miplearn.classifiers.counting import CountingClassifier
+from miplearn.components import classifier_evaluation_dict
+from miplearn.components.lazy_static import LazyConstraint
 
 logger = logging.getLogger(__name__)
 
 
 class RelaxationComponent(Component):
     """
-    A Component which builds a relaxation of the problem by dropping constraints.
+    A Component that tries to build a relaxation that is simultaneously strong and easy to solve.
 
-    Currently, this component drops all integrality constraints, as well as
-    all inequality constraints which are not likely binding in the LP relaxation.
-    In a future version of MIPLearn, this component may decide to keep some
-    integrality constraints it it determines that they have small impact on
-    running time, but large impact on dual bound.
+    Currently, this component performs the following operations:
+        - Drops all integrality constraints
+        - Drops all inequality constraints that are not likely to be binding.
+
+    In future versions of MIPLearn, this component may keep some integrality constraints and perform other operations.
+
+    Parameters
+    ----------
+    classifier : Classifier, optional
+        Classifier used to predict whether each constraint is binding or not. One deep copy of this classifier
+        is made for each constraint category.
+    threshold : float, optional
+        If the probability that a constraint is binding exceeds this threshold, the constraint is dropped from the
+        linear relaxation.
+    slack_tolerance : float, optional
+        If a constraint has slack greater than this threshold, then the constraint is considered loose. By default,
+        this threshold equals a small positive number to compensate for numerical issues.
+    check_dropped : bool, optional
+        If `check_dropped` is true, then, after the problem is solved, the component verifies that all dropped
+        constraints are still satisfied and re-adds the ones that are not.
+    violation_tolerance : float, optional
+        If `check_dropped` is true, a constraint is considered satisfied during the check if its violation is smaller
+        than this tolerance.
     """
 
     def __init__(self,
                  classifier=CountingClassifier(),
                  threshold=0.95,
                  slack_tolerance=1e-5,
+                 check_dropped=False,
+                 violation_tolerance=1e-5,
                  ):
         self.classifiers = {}
         self.classifier_prototype = classifier
         self.threshold = threshold
         self.slack_tolerance = slack_tolerance
+        self.pool = []
+        self.check_dropped = check_dropped
+        self.violation_tolerance = violation_tolerance
 
     def before_solve(self, solver, instance, _):
         logger.info("Relaxing integrality...")
@@ -47,14 +70,14 @@ class RelaxationComponent(Component):
                                 constraint_ids=cids,
                                 return_constraints=True)
         y = self.predict(x)
-        n_removed = 0
         for category in y.keys():
             for i in range(len(y[category])):
                 if y[category][i][0] == 1:
                     cid = constraints[category][i]
-                    solver.internal_solver.extract_constraint(cid)
-                    n_removed += 1
-        logger.info("Removed %d predicted redundant LP constraints" % n_removed)
+                    c = LazyConstraint(cid=cid,
+                                       obj=solver.internal_solver.extract_constraint(cid))
+                    self.pool += [c]
+        logger.info("Extracted %d predicted constraints" % len(self.pool))
 
     def after_solve(self, solver, instance, model, results):
         instance.slacks = solver.internal_solver.get_constraint_slacks()
@@ -148,4 +171,19 @@ class RelaxationComponent(Component):
                         tn += 1
         return classifier_evaluation_dict(tp, tn, fp, fn)
 
-
+    def iteration_cb(self, solver, instance, model):
+        if not self.check_dropped:
+            return False
+        logger.debug("Checking that dropped constraints are satisfied...")
+        constraints_to_add = []
+        for c in self.pool:
+            if not solver.internal_solver.is_constraint_satisfied(c.obj, self.violation_tolerance):
+                constraints_to_add.append(c)
+        for c in constraints_to_add:
+            self.pool.remove(c)
+            solver.internal_solver.add_constraint(c.obj)
+        if len(constraints_to_add) > 0:
+            logger.info("%8d constraints %8d in the pool" % (len(constraints_to_add), len(self.pool)))
+            return True
+        else:
+            return False

miplearn/components/tests/test_relaxation.py

@@ -11,7 +11,7 @@ from miplearn import (RelaxationComponent,
 from miplearn.classifiers import Classifier
 
 
-def test_usage_with_solver():
+def _setup():
     solver = Mock(spec=LearningSolver)
 
     internal = solver.internal_solver = Mock(spec=InternalSolver)
@@ -22,6 +22,8 @@ def test_usage_with_solver():
         "c3": 0.0,
         "c4": 1.4,
     })
+    internal.extract_constraint = Mock(side_effect=lambda cid: "<%s>" % cid)
+    internal.is_constraint_satisfied = Mock(return_value=False)
 
     instance = Mock(spec=Instance)
     instance.get_constraint_features = Mock(side_effect=lambda cid: {
@@ -36,21 +38,29 @@ def test_usage_with_solver():
         "c4": "type-b",
     }[cid])
 
-    component = RelaxationComponent()
-    component.classifiers = {
+    classifiers = {
         "type-a": Mock(spec=Classifier),
         "type-b": Mock(spec=Classifier),
     }
-    component.classifiers["type-a"].predict_proba = \
+    classifiers["type-a"].predict_proba = \
         Mock(return_value=[
             [0.20, 0.80],
             [0.05, 0.95],
         ])
-    component.classifiers["type-b"].predict_proba = \
+    classifiers["type-b"].predict_proba = \
         Mock(return_value=[
             [0.02, 0.98],
         ])
 
+    return solver, internal, instance, classifiers
+
+
+def test_usage():
+    solver, internal, instance, classifiers = _setup()
+
+    component = RelaxationComponent()
+    component.classifiers = classifiers
+
     # LearningSolver calls before_solve
     component.before_solve(solver, instance, None)
 
@@ -98,6 +108,44 @@ def test_usage_with_solver():
     }
 
 
+def test_usage_with_check_dropped():
+    solver, internal, instance, classifiers = _setup()
+
+    component = RelaxationComponent(check_dropped=True,
+                                    violation_tolerance=1e-3)
+    component.classifiers = classifiers
+
+    # LearningSolver call before_solve
+    component.before_solve(solver, instance, None)
+
+    # Assert constraints are extracted
+    assert internal.extract_constraint.call_count == 2
+    internal.extract_constraint.assert_has_calls([
+        call("c3"), call("c4"),
+    ])
+
+    # LearningSolver calls iteration_cb (first time)
+    should_repeat = component.iteration_cb(solver, instance, None)
+
+    # Should ask LearningSolver to repeat
+    assert should_repeat
+
+    # Should ask solver if removed constraints are satisfied (mock always returns false)
+    internal.is_constraint_satisfied.assert_has_calls([
+        call("<c3>", 1e-3),
+        call("<c4>", 1e-3),
+    ])
+
+    # Should add constraints back to LP relaxation
+    internal.add_constraint.assert_has_calls([
+        call("<c3>"), call("<c4>")
+    ])
+
+    # LearningSolver calls iteration_cb (second time)
+    should_repeat = component.iteration_cb(solver, instance, None)
+    assert not should_repeat
+
+
 def test_x_y_fit_predict_evaluate():
     instances = [Mock(spec=Instance), Mock(spec=Instance)]
     component = RelaxationComponent(slack_tolerance=0.05,
@@ -182,7 +230,3 @@ def test_x_y_fit_predict_evaluate():
     assert ev["True negative"] == 1
     assert ev["False positive"] == 1
     assert ev["False negative"] == 0
-
-
-
-