From 65fb2476b0e5924a0e09dbb13b419d8c6eb01350 Mon Sep 17 00:00:00 2001
From: poveda_g <guillaume.poveda@airbus.com>
Date: Thu, 23 Jan 2025 16:45:48 +0100
Subject: [PATCH] implementing an early stopper for cpsat solver, based on
 optimality gap to best bound. this could be generalized to mip solver in
 further work

---
 .../generic_tools/callbacks/early_stoppers.py | 35 ++++++++++++++++
 .../callbacks/stats_retrievers.py             |  2 +-
 .../test_earlystopobjective_callback.py       | 40 +++++++++++++++++++
 3 files changed, 76 insertions(+), 1 deletion(-)
 create mode 100644 tests/generic_tools/callbacks/test_earlystopobjective_callback.py

diff --git a/discrete_optimization/generic_tools/callbacks/early_stoppers.py b/discrete_optimization/generic_tools/callbacks/early_stoppers.py
index a5d429e0a..ab9550896 100644
--- a/discrete_optimization/generic_tools/callbacks/early_stoppers.py
+++ b/discrete_optimization/generic_tools/callbacks/early_stoppers.py
@@ -8,6 +8,7 @@
 
 from discrete_optimization.generic_tools.callbacks.callback import Callback
 from discrete_optimization.generic_tools.do_solver import SolverDO
+from discrete_optimization.generic_tools.ortools_cpsat_tools import OrtoolsCpSatSolver
 from discrete_optimization.generic_tools.result_storage.result_storage import (
     ResultStorage,
 )
@@ -69,3 +70,37 @@ def on_step_end(
             return True
         else:
             return False
+
+
+class ObjectiveGapCpSatSolver(Callback):
+    """
+    Stop the cpsat solver according to some classical convergence criteria
+    It could be done differently (playing with parameters of cpsat directly)
+    """
+
+    def __init__(
+        self,
+        objective_gap_rel: Optional[float] = None,
+        objective_gap_abs: Optional[float] = None,
+    ):
+        self.objective_gap_rel = objective_gap_rel
+        self.objective_gap_abs = objective_gap_abs
+
+    def on_step_end(
+        self, step: int, res: ResultStorage, solver: OrtoolsCpSatSolver
+    ) -> Optional[bool]:
+        best_sol = solver.clb.ObjectiveValue()
+        bound = solver.clb.BestObjectiveBound()
+        if self.objective_gap_abs is not None:
+            if abs(bound - best_sol) <= self.objective_gap_abs:
+                logger.debug(
+                    f"Stopping search, absolute gap {abs(bound-best_sol)}<{self.objective_gap_abs}"
+                )
+                return True
+        if self.objective_gap_rel is not None:
+            if bound != 0:
+                if abs(bound - best_sol) / abs(bound) <= self.objective_gap_rel:
+                    logger.debug(
+                        f"Stopping search, relative gap {abs(bound-best_sol)/abs(bound)}<{self.objective_gap_rel}"
+                    )
+                    return True
diff --git a/discrete_optimization/generic_tools/callbacks/stats_retrievers.py b/discrete_optimization/generic_tools/callbacks/stats_retrievers.py
index 7a15f911a..b3f9ea007 100644
--- a/discrete_optimization/generic_tools/callbacks/stats_retrievers.py
+++ b/discrete_optimization/generic_tools/callbacks/stats_retrievers.py
@@ -66,7 +66,7 @@ def on_step_end(
             }
         )
         if solver.clb.ObjectiveValue() == solver.clb.BestObjectiveBound():
-            return False
+            return True
 
     def on_solve_start(self, solver: OrtoolsCpSatSolver):
         self.starting_time = perf_counter()
diff --git a/tests/generic_tools/callbacks/test_earlystopobjective_callback.py b/tests/generic_tools/callbacks/test_earlystopobjective_callback.py
new file mode 100644
index 000000000..f466e0d83
--- /dev/null
+++ b/tests/generic_tools/callbacks/test_earlystopobjective_callback.py
@@ -0,0 +1,40 @@
+#  Copyright (c) 2025 AIRBUS and its affiliates.
+#  This source code is licensed under the MIT license found in the
+#  LICENSE file in the root directory of this source tree.
+
+import pytest
+
+from discrete_optimization.generic_tools.callbacks.early_stoppers import (
+    ObjectiveGapCpSatSolver,
+)
+from discrete_optimization.generic_tools.cp_tools import ParametersCp
+from discrete_optimization.knapsack.parser import get_data_available, parse_file
+from discrete_optimization.knapsack.problem import KnapsackProblem, MobjKnapsackModel
+from discrete_optimization.knapsack.solvers.cpsat import CpSatKnapsackSolver
+
+
+def test_knapsack_ortools_objective_callback():
+    model_file = [f for f in get_data_available() if "ks_300_0" in f][0]
+    model: KnapsackProblem = parse_file(model_file, force_recompute_values=True)
+    model: MobjKnapsackModel = MobjKnapsackModel.from_knapsack(model)
+    solver = CpSatKnapsackSolver(model)
+    solver.init_model()
+    objective_gap_rel = 0.1
+    objective_gap_abs = 10
+    mycb = ObjectiveGapCpSatSolver(
+        objective_gap_rel=objective_gap_rel, objective_gap_abs=objective_gap_abs
+    )
+    parameters_cp = ParametersCp.default()
+    result_storage = solver.solve(
+        time_limit=10,
+        parameters_cp=parameters_cp,
+        callbacks=[mycb],
+        ortools_cpsat_solver_kwargs={"log_search_progress": True},
+    )
+    assert (
+        abs(solver.clb.ObjectiveValue() - solver.clb.BestObjectiveBound())
+        <= objective_gap_abs
+        or abs(solver.clb.ObjectiveValue() - solver.clb.BestObjectiveBound())
+        / abs(solver.clb.BestObjectiveBound())
+        <= objective_gap_rel
+    )