#!/usr/bin/env python3
# Copyright (c) Meta Platforms, Inc. and affiliates.
#
# This source code is licensed under the MIT license found in the
# LICENSE file in the root directory of this source tree.
# pyre-strict
from logging import Logger
from typing import Any, Dict, List, Optional, Tuple
import numpy as np
from ax.core.base_trial import BaseTrial, TrialStatus
from ax.core.data import Data
from ax.core.experiment import Experiment
from ax.core.optimization_config import (
MultiObjectiveOptimizationConfig,
OptimizationConfig,
)
from ax.core.outcome_constraint import ObjectiveThreshold
from ax.core.trial import Trial
from ax.core.types import ComparisonOp
from ax.global_stopping.strategies.base import BaseGlobalStoppingStrategy
from ax.modelbridge.modelbridge_utils import observed_hypervolume
from ax.plot.pareto_utils import (
get_tensor_converter_model,
infer_reference_point_from_experiment,
)
from ax.utils.common.logger import get_logger
from ax.utils.common.typeutils import checked_cast, not_none
logger: Logger = get_logger(__name__)
[docs]class ImprovementGlobalStoppingStrategy(BaseGlobalStoppingStrategy):
"""
A Global Stopping Strategy which recommends stopping optimization if there
is no significant improvement over recent iterations.
This stopping strategy recommends stopping if there is no significant improvement
over the past `window_size` trials, among those that are feasible
(satisfying constraints). The meaning of a "significant"
improvement differs between single-objective and multi-objective optimizations.
For single-objective optimizations, improvement is as a fraction of the
interquartile range (IQR) of the objective values seen so far. For
multi-objective optimizations (MOO), improvement is as a fraction of the hypervolume
obtained `window_size` iterations ago.
"""
def __init__(
self,
min_trials: int,
window_size: int = 5,
improvement_bar: float = 0.1,
inactive_when_pending_trials: bool = True,
) -> None:
"""
Initialize an improvement-based stopping strategy.
Args:
min_trials: Minimum number of trials before the stopping strategy
kicks in.
window_size: Number of recent trials to check the improvement in.
The first trial that could be used for analysis is
`min_trials - window_size`; the first trial for which stopping
might be recommended is `min_trials`.
improvement_bar: Threshold for considering improvement over the best
point, relative to the interquartile range of values seen so
far. Must be >= 0.
inactive_when_pending_trials: If set, the optimization will not stopped as
long as it has running trials.
"""
if improvement_bar < 0:
raise ValueError("improvement_bar must be >= 0.")
super().__init__(
min_trials=min_trials,
inactive_when_pending_trials=inactive_when_pending_trials,
)
self.window_size = window_size
self.improvement_bar = improvement_bar
self.hv_by_trial: Dict[int, float] = {}
self._inferred_objective_thresholds: Optional[List[ObjectiveThreshold]] = None
def _should_stop_optimization(
self,
experiment: Experiment,
trial_to_check: Optional[int] = None,
objective_thresholds: Optional[List[ObjectiveThreshold]] = None,
**kwargs: Dict[str, Any],
) -> Tuple[bool, str]:
"""
Check if the objective has improved significantly in the past
"window_size" iterations.
For single-objective optimization experiments, it will call
_should_stop_single_objective() and for MOO experiments, it will call
_should_stop_moo(). Before making either of these calls, this function carries
out some sanity checks to handle obvious/invalid cases. For more detail
on what it means to "significantly" improve, see the class docstring.
Args:
experiment: The experiment to apply the strategy on.
trial_to_check: The trial in the experiment at which we want to check
for stopping. If None, we check at the latest trial.
objective_thresholds: Custom objective thresholds to use as reference pooint
when computing hv of the pareto front against. This is used only in the
MOO setting. If not specified, the objective thresholds on the
experiment's optimization config will be used for the purpose.
If no thresholds are provided, they are automatically inferred. They are
only inferred once for each instance of the strategy (i.e. inferred
thresholds don't update with additional data).
kwargs: Unused.
Returns:
A Tuple with a boolean determining whether the optimization should stop,
and a str declaring the reason for stopping.
"""
max_completed_trial = max(
experiment.trial_indices_by_status[TrialStatus.COMPLETED]
)
if trial_to_check is None:
trial_to_check = max_completed_trial
elif trial_to_check > max_completed_trial:
raise ValueError(
"trial_to_check is larger than the total number of "
f"trials (={max_completed_trial})."
)
# Only counting the trials up to trial_to_check.
num_completed_trials = sum(
index <= trial_to_check
for index in experiment.trial_indices_by_status[TrialStatus.COMPLETED]
)
min_required_trials = max(self.min_trials, self.window_size)
if num_completed_trials < min_required_trials:
stop = False
message = (
"There are not enough completed trials to make a stopping decision "
f"(completed: {num_completed_trials}, required: {min_required_trials})."
)
return stop, message
if isinstance(experiment.optimization_config, MultiObjectiveOptimizationConfig):
if objective_thresholds is None:
# self._inferred_objective_thresholds is cached and only computed once.
if self._inferred_objective_thresholds is None:
# only infer reference point if there is data on the experiment.
data = experiment.fetch_data()
if not data.df.empty:
# We infer the nadir reference point to be used by the GSS.
self._inferred_objective_thresholds = (
infer_reference_point_from_experiment(
experiment=experiment, data=data
)
)
# TODO: move this out into a separate infer_objective_thresholds
# instance method or property that handles the caching.
objective_thresholds = self._inferred_objective_thresholds
return self._should_stop_moo(
experiment=experiment,
trial_to_check=trial_to_check,
objective_thresholds=not_none(objective_thresholds),
)
else:
return self._should_stop_single_objective(
experiment=experiment, trial_to_check=trial_to_check
)
def _should_stop_moo(
self,
experiment: Experiment,
trial_to_check: int,
objective_thresholds: List[ObjectiveThreshold],
) -> Tuple[bool, str]:
"""
This is the "should_stop_optimization" method of this class, specialized
to MOO experiments.
It computes the (feasible) hypervolume of the Pareto front at
`trial_to_check` trial and `window_size` trials before, and suggest to stop the
optimization if the improvment in hypervolume over the past
`window_size` trials, as a fraction of the hypervolume at the start of
the window, is less than `self.improvement_bar`. When the hypervolume is
zero at the beginning of the window, stopping is never recommended.
Becaues hypervolume computations are expensive, these are stored to
increase the speed of future checks.
Args:
experiment: The experiment to apply the strategy on.
trial_to_check: The trial in the experiment at which we want to check
for stopping. If None, we check at the latest trial.
objective_thresholds: Custom objective thresholds to use as reference pooint
when computing hv of the pareto front against. This is used only in the
MOO setting. If not specified, the objective thresholds on the
experiment's optimization config will be used for the purpose.
Returns:
A Tuple with a boolean determining whether the optimization should stop,
and a str declaring the reason for stopping.
"""
reference_trial_index = trial_to_check - self.window_size + 1
data_df = experiment.fetch_data().df
data_df_reference = data_df[data_df["trial_index"] <= reference_trial_index]
data_df = data_df[data_df["trial_index"] <= trial_to_check]
# Computing or retrieving HV at "window_size" iteration before
if reference_trial_index in self.hv_by_trial:
hv_reference = self.hv_by_trial[reference_trial_index]
else:
mb_reference = get_tensor_converter_model(
experiment=experiment, data=Data(data_df_reference)
)
hv_reference = observed_hypervolume(
modelbridge=mb_reference, objective_thresholds=objective_thresholds
)
self.hv_by_trial[reference_trial_index] = hv_reference
if hv_reference == 0:
message = "The reference hypervolume is 0. Continue the optimization."
return False, message
# Computing HV at current trial
mb = get_tensor_converter_model(experiment=experiment, data=Data(data_df))
hv = observed_hypervolume(mb, objective_thresholds=objective_thresholds)
self.hv_by_trial[trial_to_check] = hv
hv_improvement = (hv - hv_reference) / hv_reference
stop = hv_improvement < self.improvement_bar
if stop:
message = (
f"The improvement in hypervolume in the past {self.window_size} "
f"trials (={hv_improvement:.3f}) is less than improvement_bar "
f"(={self.improvement_bar}) times the hypervolume at the start "
f"of the window (={hv_reference:.3f})."
)
else:
message = ""
return stop, message
def _should_stop_single_objective(
self, experiment: Experiment, trial_to_check: int
) -> Tuple[bool, str]:
"""
This is the `_should_stop_optimization` method of this class,
specialized to single-objective experiments.
It computes the interquartile range (IQR) of feasible objective values
found so far, then computes the improvement in the best seen over the
last `window_size` trials. If the recent improvement as a fraction of
the IQR is less than `self.improvement_bar`, it recommends stopping.
Args:
experiment: The experiment to apply the strategy on.
trial_to_check: The trial in the experiment at which we want to check
for stopping. If None, we check at the latest trial.
Returns:
A Tuple with a boolean determining whether the optimization should stop,
and a str declaring the reason for stopping.
"""
objectives = []
is_feasible = []
for trial in experiment.trials_by_status[TrialStatus.COMPLETED]:
if trial.index <= trial_to_check:
tr = checked_cast(Trial, trial)
objectives.append(tr.objective_mean)
is_feasible.append(constraint_satisfaction(tr))
if checked_cast(
OptimizationConfig, experiment.optimization_config
).objective.minimize:
selector, mask_val = np.minimum, np.inf
else:
selector, mask_val = np.maximum, -np.inf
# Replace objective value at infeasible iterations with mask_val
masked_obj = np.where(is_feasible, objectives, mask_val)
running_optimum = selector.accumulate(masked_obj)
# Computing the interquartile for scaling the difference
feasible_objectives = np.array(objectives)[is_feasible]
if len(feasible_objectives) <= 1:
message = "There are not enough feasible arms tried yet."
return False, message
q3, q1 = np.percentile(feasible_objectives, [75, 25])
iqr = q3 - q1
relative_improvement = np.abs(
(running_optimum[-1] - running_optimum[-self.window_size]) / iqr
)
stop = relative_improvement < self.improvement_bar
if stop:
message = (
f"The improvement in best objective in the past {self.window_size} "
f"trials (={relative_improvement:.3f}) is less than "
f"{self.improvement_bar} times the interquartile range (IQR) of "
f"objectives attained so far (IQR={iqr:.3f})."
)
else:
message = ""
return stop, message
[docs]def constraint_satisfaction(trial: BaseTrial) -> bool:
"""
This function checks whether the outcome constraints of the
optimization config of an experiment are satisfied in the
given trial.
Args:
trial: A single-arm Trial at which we want to check the constraint.
Returns:
A boolean which is True iff all outcome constraints are satisfied.
"""
outcome_constraints = not_none(
trial.experiment.optimization_config
).outcome_constraints
if len(outcome_constraints) == 0:
return True
df = trial.lookup_data().df
for constraint in outcome_constraints:
bound = constraint.bound
metric_name = constraint.metric.name
metric_data = df.loc[df["metric_name"] == metric_name]
mean, sem = metric_data.iloc[0][["mean", "sem"]]
if sem > 0.0:
logger.warning(
f"There is observation noise for metric {metric_name}. This may "
"negatively affect the way we check constraint satisfaction."
)
if constraint.op is ComparisonOp.LEQ:
satisfied = mean <= bound
else:
satisfied = mean >= bound
if not satisfied:
return False
return True
