utility_function_implementation=ConfidenceBoundUtilityFunction.
__name__,
numeric_optimizer_implementation=RandomSearchOptimizer.__name__,
confidence_bound_utility_function_config=
confidence_bound_utility_function_config_store.default,
random_search_optimizer_config=random_search_optimizer_config_store.
default,
fraction_random_suggestions=0.5))
experiment_designer_config_store.add_config_by_name(
config_name="default_glow_worm_config",
config_point=Point(
utility_function_implementation=ConfidenceBoundUtilityFunction.
__name__,
numeric_optimizer_implementation=GlowWormSwarmOptimizer.__name__,
confidence_bound_utility_function_config=
confidence_bound_utility_function_config_store.default,
glow_worm_swarm_optimizer_config=glow_worm_swarm_optimizer_config_store
.default,
fraction_random_suggestions=0.5),
description=
"Experiment designer config with glow worm swarm optimizer as a utility function optimizer."
)
experiment_designer_config_store.add_config_by_name(
config_name="default_multi_objective_config",
config_point=Point(
utility_function_implementation=
MultiObjectiveProbabilityOfImprovementUtilityFunction.__name__,
numeric_optimizer_implementation=RandomSearchOptimizer.__name__,
multi_objective_probability_of_improvement_config=
multi_objective_probability_of_improvement_utility_function_config_store
min=2,
max=10000)
]).join(
subgrid=homogeneous_random_forest_config_store.parameter_space,
on_external_dimension=CategoricalDimension(
name="surrogate_model_implementation",
values=[
HomogeneousRandomForestRegressionModel.__name__
])).join(
subgrid=experiment_designer_config_store.parameter_space,
on_external_dimension=CategoricalDimension(
name="experiment_designer_implementation",
values=[ExperimentDesigner.__name__])),
default=Point(
surrogate_model_implementation=HomogeneousRandomForestRegressionModel.
__name__,
experiment_designer_implementation=ExperimentDesigner.__name__,
min_samples_required_for_guided_design_of_experiments=10,
homogeneous_random_forest_regression_model_config=
homogeneous_random_forest_config_store.default,
experiment_designer_config=experiment_designer_config_store.default),
description="TODO")
# Add a config with homogeneous random forest where the decision trees refit for every new observation.
#
optimizer_config = bayesian_optimizer_config_store.default
optimizer_config.homogeneous_random_forest_regression_model_config.decision_tree_regression_model_config.n_new_samples_before_refit = 1
optimizer_config.homogeneous_random_forest_regression_model_config.n_estimators = 50
bayesian_optimizer_config_store.add_config_by_name(
config_name='default_refit_tree_every_time', config_point=optimizer_config)
default=Point(
surrogate_model_implementation=HomogeneousRandomForestRegressionModel.
__name__,
experiment_designer_implementation=ExperimentDesigner.__name__,
min_samples_required_for_guided_design_of_experiments=10,
homogeneous_random_forest_regression_model_config=
homogeneous_random_forest_config_store.default,
experiment_designer_config=experiment_designer_config_store.default),
description="TODO")
# Add a config with homogeneous random forest where the decision trees refit for every new observation.
#
optimizer_config = bayesian_optimizer_config_store.default
optimizer_config.homogeneous_random_forest_regression_model_config.decision_tree_regression_model_config.n_new_samples_before_refit = 1
optimizer_config.homogeneous_random_forest_regression_model_config.n_estimators = 50
bayesian_optimizer_config_store.add_config_by_name(
config_name='default_refit_tree_every_time', config_point=optimizer_config)
# Add a default config with glowworm swarm optimizer
#
bayesian_optimizer_config_store.add_config_by_name(
config_name="default_with_glow_worm",
config_point=Point(
surrogate_model_implementation=HomogeneousRandomForestRegressionModel.
__name__,
experiment_designer_implementation=ExperimentDesigner.__name__,
min_samples_required_for_guided_design_of_experiments=10,
homogeneous_random_forest_regression_model_config=
homogeneous_random_forest_config_store.default,
experiment_designer_config=experiment_designer_config_store.
get_config_by_name("default_glow_worm_config")))
utility_function_implementation=ConfidenceBoundUtilityFunction.
__name__,
numeric_optimizer_implementation=RandomSearchOptimizer.__name__,
confidence_bound_utility_function_config=
confidence_bound_utility_function_config_store.default,
random_search_optimizer_config=random_search_optimizer_config_store.
default,
fraction_random_suggestions=0.5))
experiment_designer_config_store.add_config_by_name(
config_name="default_glow_worm_config",
config_point=Point(
utility_function_implementation=ConfidenceBoundUtilityFunction.
__name__,
numeric_optimizer_implementation=GlowWormSwarmOptimizer.__name__,
confidence_bound_utility_function_config=
confidence_bound_utility_function_config_store.default,
glow_worm_swarm_optimizer_config=glow_worm_swarm_optimizer_config_store
.default,
fraction_random_suggestions=0.5),
description=
"Experiment designer config with glow worm swarm optimizer as a utility function optimizer."
)
class ExperimentDesigner:
""" Portion of a BayesianOptimizer concerned with Design of Experiments.
The two main components of a Bayesian Optimizer are:
* the surrogate model - responsible for fitting a regression function to try to predict some performance metric(s)
based on suggested config, and context information
* the experiment designer - responsible for suggesting the next configuration to try against the real system.
),
default=Point(
uncertainty_type="constant",
use_objective_function=True,
predicted_value_degrees_of_freedom=10,
constant_uncertainty_config=Point(value=1),
objective_function_config=objective_function_config_store.get_config_by_name("three_level_quadratic")
),
description=""
)
multi_objective_pass_through_model_config_store.add_config_by_name(
config_name="three_level_quadratic",
config_point=Point(
uncertainty_type="constant",
use_objective_function=True,
predicted_value_degrees_of_freedom=10,
constant_uncertainty_config=Point(value=1),
objective_function_config=objective_function_config_store.get_config_by_name("three_level_quadratic")
)
)
multi_objective_pass_through_model_config_store.add_config_by_name(
config_name="multi_objective_waves_3_params_2_objectives_half_pi_phase_difference",
config_point=Point(
uncertainty_type="coefficient_of_variation",
use_objective_function=True,
predicted_value_degrees_of_freedom=10,
coefficient_of_variation_config=Point(value=0.1),
objective_function_config=objective_function_config_store.get_config_by_name("multi_objective_waves_3_params_2_objectives_half_pi_phase_difference")
)
)
* velocity_convergence_threshold - when an incumbent's velocity drops below this threshold, it is assumed to have converged.
* max_num_iterations - cap on the number of iterations. A failsafe - should be higher than what the algorithm needs to converge on average.
* num_neighbors - how many random neighbors to generate for each incumbent?
* num_cached_good_params - how many good configurations should this optimizer cache for future use?
* initial_points_pareto_weight - what proportion of initial points should come from the pareto frontier?
* initial_points_cached_good_params_weight - what proportion of initial points should come from the good params cache?
* initial_points_random_params_weight - what proportion of initial points should be randomly generated?
""")
random_near_incumbent_optimizer_config_store.add_config_by_name(
config_name="20_incumbents_50_neighbors",
config_point=Point(num_starting_configs=20,
initial_velocity=0.2,
velocity_update_constant=0.3,
velocity_convergence_threshold=0.01,
max_num_iterations=15,
num_neighbors=50,
num_cached_good_params=2**10,
initial_points_pareto_weight=0.5,
initial_points_cached_good_params_weight=0.3,
initial_points_random_params_weight=0.2),
description="More thorough and more expensive than the default.")
class RandomNearIncumbentOptimizer(UtilityFunctionOptimizer):
""" Searches the utility function for maxima using the random near incumbent strategy.
Starting from an incumbent configuration, this optimizer creates a 'cloud' of random points in the incumbent's vicinity
and evaluates the utility function for each of these points. If any of the new points has a higher utility value than the
incumbent then it gets promoted to the incumbent and we repeat the process.