✨: Add option to pass initial configs

Author: tmechsner

backtesting/backtesting.py

@@ -1262,6 +1262,7 @@ def optimize(self, *,
                  acq_type: str = 'auto',
                  use_constrained_model: bool = False,
                  return_configs: int = 100,
+                 init_configs: list[dict] | None = None,
                  **kwargs) -> Tuple[pd.Series, List]:
         """
         Optimize strategy parameters to an optimal combination.
@@ -1313,6 +1314,9 @@ def optimize(self, *,
 
         `return_configs` is the number of close-to-optimal configurations to return.
 
+        `init_configs` can be a list of dicts with initial configurations. The number of generated
+        initial configurations will be reduced accordingly.
+
         Additional keyword arguments represent strategy arguments with
         list-like collections of possible values. For example, the following
         code finds and returns the "best" of the 7 admissible (of the
@@ -1453,8 +1457,22 @@ def eval_run(config: sp.Configuration):
             initial_runs = min(max_tries, n_initial_points or 20 + 3 * len(kwargs))
 
             if init_strategy == 'latin_hypercube':
-                lhs = LatinHypercubeSampler(space, initial_runs, criterion='maximin')
-                initial_configs = lhs.generate(return_config=True)
+                given_initial_configs = []
+                if init_configs:
+                    for config in init_configs:
+                        try:
+                            given_initial_configs.append(sp.Configuration(configuration_space=space, values=config))
+                        except Exception:
+                            logging.warning("Invalid externally given initial config discovered")
+                            continue
+                initial_runs_to_generate = initial_runs - len(given_initial_configs)
+                if initial_runs_to_generate > 0:
+                    lhs = LatinHypercubeSampler(space, initial_runs_to_generate, criterion='maximin')
+                    generated_initial_configs = lhs.generate(return_config=True)
+                else:
+                    generated_initial_configs = []
+
+                initial_configs = given_initial_configs + generated_initial_configs
 
                 valid_initial_configs = []
                 for config in initial_configs:

backtesting/test/_test.py

@@ -572,15 +572,39 @@ def test_method_skopt(self):
 
     def test_method_openbox(self):
         bt = Backtest(GOOG.iloc[:100], SmaCross)
-        res = bt.optimize(
+        res, best_configs = bt.optimize(
             fast=range(2, 20), slow=np.arange(2, 20, dtype=object), blubb=[-2.0],
             constraint=lambda p: p.fast < p.slow,
             max_tries=30,
             method='openbox',
             return_optimization=False,
             return_heatmap=False,
-            random_state=2)
+            random_state=2,
+        )
+        self.assertIsInstance(res, pd.Series)
+        self.assertIsInstance(best_configs, list)
+        self.assertEqual(len(best_configs), 30)
+
+    def test_method_openbox_with_given_initial_configs(self):
+        bt = Backtest(GOOG.iloc[:100], SmaCross)
+        res, best_configs = bt.optimize(
+            fast=range(2, 20), slow=np.arange(2, 20, dtype=object), blubb=[-2.0],
+            constraint=lambda p: p.fast < p.slow,
+            max_tries=30,
+            method='openbox',
+            init_strategy='latin_hypercube',
+            init_configs=[
+                {'fast': 2, 'slow': 3, 'blubb': -2.0},
+                {'fast': 2, 'slow': 4, 'blubb': -2.0},
+                {'fast': 2, 'slow': 5, 'blubb': -2.0},
+            ],
+            return_optimization=False,
+            return_heatmap=False,
+            random_state=2,
+        )
         self.assertIsInstance(res, pd.Series)
+        self.assertIsInstance(best_configs, list)
+        self.assertEqual(len(best_configs), 30)
 
     # def test_method_openbox_parallel(self):
     #     bt = Backtest(GOOG.iloc[:100], SmaCross)