Improvements to the simulated annealing algorithm (#23)

anti_clustering/_base.py

@@ -37,7 +37,6 @@ def run(
         num_groups: int,
         destination_column: str,
     ) -> pd.DataFrame:
-        # pylint: disable = R0913
         """
         Run anti clustering algorithm on dataset.
         :param df: The dataset to run anti-clustering on.

anti_clustering/simulated_annealing_heuristic.py

@@ -12,7 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 """
-A simulated annealing approach to solving the anti-clustering problem.
+A simulated annealing with restarts approach to solving the anti-clustering problem.
 """
 
 import math
@@ -23,7 +23,7 @@
 
 class SimulatedAnnealingHeuristicAntiClustering(ClusterSwapHeuristic):
     """
-    A simulated annealing approach to solving the anti-clustering problem.
+    A simulated annealing with restarts approach to solving the anti-clustering problem.
     """
 
     def __init__(
@@ -32,13 +32,15 @@ def __init__(
         random_seed: int = None,
         alpha: float = 0.9,
         iterations: int = 2000,
-        starting_temperature: float = 10,
+        starting_temperature: float = 100,
+        restarts: int = 9,
     ):
         # pylint: disable = R0913
         super().__init__(verbose=verbose, random_seed=random_seed)
         self.alpha = alpha
         self.iterations = iterations
         self.starting_temperature = starting_temperature
+        self.restarts = restarts
 
     def _solve(self, distance_matrix: npt.NDArray[float], num_groups: int) -> npt.NDArray[bool]:
         # Start with random cluster assignment
@@ -47,34 +49,48 @@ def _solve(self, distance_matrix: npt.NDArray[float], num_groups: int) -> npt.ND
         if self.verbose:
             print("Solving")
 
-        temperature = self.starting_temperature
-        # Initial objective value
-        objective = self._calculate_objective(cluster_assignment, distance_matrix)
-        for iteration in range(self.iterations):
-            if self.verbose and iteration % 5 == 0:
-                print(f"Iteration {iteration + 1} of {self.iterations}")
+        candidate_solutions = []
 
-            # Select random element
-            i = self.rnd.randint(0, len(distance_matrix) - 1)
-            # Get possible swaps
-            possible_exchanges = self._get_exchanges(cluster_assignment, i)
-            if len(possible_exchanges) == 0:
-                continue
-            # Select random possible swap.
-            j = possible_exchanges[self.rnd.randint(0, len(possible_exchanges) - 1)]
+        for restart in range(self.restarts):
+            temperature = self.starting_temperature
+            # Initial objective value
+            objective = self._calculate_objective(cluster_assignment, distance_matrix)
+            for iteration in range(self.iterations):
+                if self.verbose and iteration % 5 == 0:
+                    print(f"Iteration {iteration + 1} of {self.iterations}")
 
-            new_cluster_assignment = self._swap(cluster_assignment, i, j)
-            new_objective = self._calculate_objective(new_cluster_assignment, distance_matrix)
+                # Select random element
+                i = self.rnd.randint(0, len(distance_matrix) - 1)
+                # Get possible swaps
+                possible_exchanges = self._get_exchanges(cluster_assignment, i)
+                if len(possible_exchanges) == 0:
+                    continue
+                # Select random possible swap.
+                j = possible_exchanges[self.rnd.randint(0, len(possible_exchanges) - 1)]
 
-            # Select solution as current if accepted
-            if self._accept(new_objective - objective, temperature):
-                objective = new_objective
-                cluster_assignment = new_cluster_assignment
+                new_cluster_assignment = self._swap(cluster_assignment, i, j)
+                new_objective = self._calculate_objective(new_cluster_assignment, distance_matrix)
 
-            # Cool down temperature
-            temperature = temperature * self.alpha
+                # Select solution as current if accepted
+                if self._accept(new_objective - objective, temperature):
+                    objective = new_objective
+                    cluster_assignment = new_cluster_assignment
 
-        return cluster_assignment
+                # Cool down temperature
+                temperature = temperature * self.alpha
+
+            candidate_solutions.append((objective, cluster_assignment))
+
+            if self.verbose:
+                print(f"Restart {restart + 1} of {self.restarts}")
+
+            # Cold restart, select random cluster assignment
+            cluster_assignment = self._get_random_clusters(num_groups=num_groups, num_elements=len(distance_matrix))
+
+        # Select best solution, maximizing objective
+        _, best_cluster_assignment = max(candidate_solutions, key=lambda x: x[0])
+
+        return best_cluster_assignment
 
     def _calculate_objective(self, cluster_assignment: npt.NDArray[bool], distance_matrix: npt.NDArray[float]) -> float:
         """
@@ -87,7 +103,7 @@ def _calculate_objective(self, cluster_assignment: npt.NDArray[bool], distance_m
 
     def _accept(self, delta: float, temperature: float) -> bool:
         """
-        Simulated annealing acceptance function. Notice d/t is negated because this is a maximisation problem.
+        Simulated annealing acceptance function. Notice d/t is used instead of -d/t because we are maximizing.
         :param delta: Difference in objective
         :param temperature: Current temperature
         :return: Whether the solution is accepted or not.

examples/evaluation.py

@@ -37,7 +37,7 @@
 
 methods: List[AntiClustering] = [
     ExchangeHeuristicAntiClustering(),
-    SimulatedAnnealingHeuristicAntiClustering(alpha=0.95, iterations=10000, starting_temperature=10000),
+    SimulatedAnnealingHeuristicAntiClustering(alpha=0.95, iterations=5000, starting_temperature=1000, restarts=15),
     NaiveRandomHeuristicAntiClustering(),
     ExactClusterEditingAntiClustering(),
 ]