Add basic support for repeating a trial config. (#642)

Adds very rudimentary support for repeating configs across multiple
trials. We fully expect to expand on more advanced support for this once
we add a proper scheduler (#463).

Additional tests forth coming with #633 and related PRs.

mlos_bench/mlos_bench/config/schemas/cli/cli-schema.json

@@ -67,6 +67,13 @@
             "$ref": "#/$defs/json_config_path"
         },
 
+        "trial_config_repeat_count": {
+            "description": "Number of times to repeat a config.",
+            "type": "integer",
+            "minimum": 1,
+            "examples": [3, 5]
+        },
+
         "storage": {
             "description": "Path to the json config describing the storage backend to use.",
             "$ref": "#/$defs/json_config_path"

mlos_bench/mlos_bench/launcher.py

@@ -75,6 +75,8 @@ class Launcher:
         else:
             config = {}
 
+        self.trial_config_repeat_count = args.trial_config_repeat_count or config.get("trial_config_repeat_count", 1)
+
         log_level = args.log_level or config.get("log_level", _LOG_LEVEL)
         try:
             log_level = int(log_level)
@@ -195,6 +197,10 @@ class Launcher:
             help='Path to the optimizer configuration file. If omitted, run' +
                  ' a single trial with default (or specified in --tunable_values).')
 
+        parser.add_argument(
+            '--trial_config_repeat_count', '--trial-config-repeat-count', required=False, type=int, default=1,
+            help='Number of times to repeat each config. Default is 1 trial per config, though more may be advised.')
+
         parser.add_argument(
             '--storage', required=False,
             help='Path to the storage configuration file.' +

mlos_bench/mlos_bench/run.py

@@ -36,7 +36,8 @@ def _main() -> None:
         storage=launcher.storage,
         root_env_config=launcher.root_env_config,
         global_config=launcher.global_config,
-        do_teardown=launcher.teardown
+        do_teardown=launcher.teardown,
+        trial_config_repeat_count=launcher.trial_config_repeat_count,
     )
 
     _LOG.info("Final result: %s", result)
@@ -48,7 +49,9 @@ def _optimize(*,
               storage: Storage,
               root_env_config: str,
               global_config: Dict[str, Any],
-              do_teardown: bool) -> Tuple[Optional[float], Optional[TunableGroups]]:
+              do_teardown: bool,
+              trial_config_repeat_count: int = 1,
+              ) -> Tuple[Optional[float], Optional[TunableGroups]]:
     """
     Main optimization loop.
 
@@ -66,8 +69,13 @@ def _optimize(*,
         Global configuration parameters.
     do_teardown : bool
         If True, teardown the environment at the end of the experiment
+    trial_config_repeat_count : int
+        How many trials to repeat for the same configuration.
     """
     # pylint: disable=too-many-locals
+    if trial_config_repeat_count <= 0:
+        raise ValueError(f"Invalid trial_config_repeat_count: {trial_config_repeat_count}")
+
     if _LOG.isEnabledFor(logging.INFO):
         _LOG.info("Root Environment:\n%s", env.pprint())
 
@@ -118,16 +126,18 @@ def _optimize(*,
                                config_id, json.dumps(tunable_values, indent=2))
                 config_id = -1
 
-            trial = exp.new_trial(tunables, config={
-                # Add some additional metadata to track for the trial such as the
-                # optimizer config used.
-                # TODO: Improve for supporting multi-objective
-                # (e.g., opt_target_1, opt_target_2, ... and opt_direction_1, opt_direction_2, ...)
-                "optimizer": opt.name,
-                "opt_target": opt.target,
-                "opt_direction": opt.direction,
-            })
-            _run(env_context, opt_context, trial, global_config)
+            for repeat_i in range(1, trial_config_repeat_count + 1):
+                trial = exp.new_trial(tunables, config={
+                    # Add some additional metadata to track for the trial such as the
+                    # optimizer config used.
+                    # TODO: Improve for supporting multi-objective
+                    # (e.g., opt_target_1, opt_target_2, ... and opt_direction_1, opt_direction_2, ...)
+                    "optimizer": opt.name,
+                    "opt_target": opt.target,
+                    "opt_direction": opt.direction,
+                    "repeat_i": repeat_i,
+                })
+                _run(env_context, opt_context, trial, global_config)
 
         if do_teardown:
             env_context.teardown()

mlos_bench/mlos_bench/tests/config/schemas/cli/test-cases/good/full/full-cli.jsonc

@@ -15,6 +15,8 @@
     "optimizer": "optimizers/one_shot_opt.jsonc",
     "storage": "storage/sqlite.jsonc",
 
+    "trial_config_repeat_count": 3,
+
     "random_init": true,
     "random_seed": 42,
 

mlos_bench/mlos_bench/tests/launcher_run_test.py

@@ -93,7 +93,7 @@ def test_launch_main_app_opt(root_path: str, local_exec_service: LocalExecServic
     """
     _launch_main_app(
         root_path, local_exec_service,
-        "--config mlos_bench/mlos_bench/tests/config/cli/mock-opt.jsonc --max_iterations 3",
+        "--config mlos_bench/mlos_bench/tests/config/cli/mock-opt.jsonc --trial_config_repeat_count 3 --max_iterations 3",
         [
             # Iteration 1: Expect first value to be the baseline
             f"^{_RE_DATE} mlos_core_optimizer\\.py:\\d+ " +
@@ -106,6 +106,6 @@ def test_launch_main_app_opt(root_path: str, local_exec_service: LocalExecServic
             r"register DEBUG Score: \d+\.\d+ Dataframe:\s*$",
             # Final result: baseline is the optimum for the mock environment
             f"^{_RE_DATE} run\\.py:\\d+ " +
-            r"_optimize INFO Env: Mock environment best score: 64\.88\d+\s*$",
+            r"_optimize INFO Env: Mock environment best score: 64\.53\d+\s*$",
         ]
     )