desdeo_emo.surrogatemodels

This module provides implementations of EAs which can be used for training surrogate models.

Submodules

Package Contents

Classes

BioGP

EvoNN

Helper class that provides a standard way to create an ABC using

EvoDN2

Helper class that provides a standard way to create an ABC using

surrogateProblem

The base class for the problems.
class desdeo_emo.surrogatemodels.BioGP(training_algorithm: Type[desdeo_emo.EAs.BaseEA.BaseEA] = PPGA, pop_size: int = 500, probability_crossover: float = 0.9, probability_mutation: float = 0.3, max_depth: int = 5, max_subtrees: int = 4, prob_terminal: float = 0.5, complexity_scalar: float = 0.5, error_lim: float = 0.001, init_method: str = 'ramped_half_and_half', model_selection_criterion: str = 'min_error', loss_function: str = 'mse', single_obj_generations: int = 10, function_set=('add', 'sub', 'mul', 'div'), terminal_set=None)

Bases: desdeo_problem.surrogatemodels.SurrogateModels.BaseRegressor

Helper class that provides a standard way to create an ABC using inheritance.

fit(X: pandas.DataFrame, y: pandas.DataFrame)
_create_individuals()
_model_performance(trees: LinearNode, X: numpy.ndarray = None, y: numpy.ndarray = None)
predict(X: numpy.ndarray)
select()
static add(x, y)
static sub(x, y)
static mul(x, y)
static div(x, y)
static sqrt(x)
static log(x)
static sin(x)
static cos(x)
static tan(x)
static neg(x)
class desdeo_emo.surrogatemodels.EvoNN(num_hidden_nodes: int = 20, p_omit: float = 0.2, w_low: float = -5.0, w_high: float = 5.0, activation_function: str = 'sigmoid', loss_function: str = 'mse', training_algorithm: Type[desdeo_emo.EAs.BaseEA.BaseEA] = PPGA, pop_size: int = 500, model_selection_criterion: str = 'akaike_corrected', recombination_type: str = 'evonn_xover_mutation', crossover_type: str = 'standard', mutation_type: str = 'gaussian')[source]

Bases: desdeo_problem.surrogatemodels.SurrogateModels.BaseRegressor

Helper class that provides a standard way to create an ABC using inheritance.

fit(X: numpy.ndarray, y: numpy.ndarray)
_model_performance(first_layer: numpy.ndarray = None, X: numpy.ndarray = None, y_true: numpy.ndarray = None)
predict(X: numpy.ndarray = None, first_layer: numpy.ndarray = None, training: bool = False)
activate(x)
calculate_linear(previous_layer_output)

Calculate the final layer using LLSQ or

Parameters:

non_linear_layer (np.ndarray) – Output of the activation function

Returns:

  • linear_layer (np.ndarray) – The optimized weight matrix of the upper part of the network

  • predicted_values (np.ndarray) – The prediction of the model

  • training_error (float) – The model’s training error

_create_individuals()
select()
class desdeo_emo.surrogatemodels.EvoDN2(num_subnets: int = 4, num_subsets: int = 4, max_layers: int = 4, max_nodes: int = 4, p_omit: float = 0.2, w_low: float = -5.0, w_high: float = 5.0, subsets: list = None, activation_function: str = 'sigmoid', loss_function: str = 'mse', training_algorithm: desdeo_emo.EAs.BaseEA.BaseEA = PPGA, pop_size: int = 500, model_selection_criterion: str = 'min_error', verbose: int = 0)[source]

Bases: desdeo_problem.surrogatemodels.SurrogateModels.BaseRegressor

Helper class that provides a standard way to create an ABC using inheritance.

fit(X: numpy.ndarray, y: numpy.ndarray)
_model_performance(individuals: numpy.ndarray = None, X: numpy.ndarray = None, y_true: numpy.ndarray = None)
_feed_forward(subnets, X)
_calculate_linear(previous_layer_output)

Calculate the final layer using LLSQ or

Parameters:

non_linear_layer (np.ndarray) – Output of the activation function

Returns:

  • linear_layer (np.ndarray) – The optimized weight matrix of the upper part of the network

  • predicted_values (np.ndarray) – The prediction of the model

activate(x)
predict(X)
select()
_create_individuals()
class desdeo_emo.surrogatemodels.surrogateProblem(performance_evaluator)[source]

Bases: desdeo_problem.problem.ProblemBase

The base class for the problems.

All other problem classes should be derived from this.

nadir

Nadir values for the problem, initiated = None

Type:

np.ndarray

ideal

Ideal values for the problem, initiated = None

Type:

np.ndarray

nadir_fitness

Fitness values for nadir, initiated = None

Type:

np.ndarray

ideal_fitness

Fitness values for ideal, initiated = None

Type:

np.ndarray

__n_of_objectives

Number of objectives, initiated = 0

Type:

int

__n_of_variables

Number of variables, initiated = 0

Type:

int

__decision_vectors

Array of decision variable vectors, initiated = None

Type:

np.ndarray

__objective_vectors

Array of objective variable vectors, initiated = None

Type:

np.ndarray

evaluate(model_parameters, use_surrogates=False)

Abstract method to evaluate problem.

Evaluates the problem using an ensemble of input vectors. Uses surrogate models if available. Otherwise, it uses the true evaluator.

Parameters:

  • decision_vectors (np.ndarray) – An array of decision variable

  • vectors. (input) –

  • use_surrogate (bool) – A bool to control whether to use the true, potentially

  • objectives. (expensive function or a surrogate model to evaluate the) –

Returns:

Dict with the following keys:
’objectives’ (np.ndarray): The objective function values for each input

vector.

’constraints’ (Union[np.ndarray, None]): The constraint values of the

problem corresponding each input vector.

’fitness’ (np.ndarray): Equal to objective values if objective is to be

minimized. Multiplied by (-1) if objective to be maximized.

’uncertainity’ (Union[np.ndarray, None]): The uncertainity in the

objective values.

Return type:

(Dict)

evaluate_constraint_values()

Abstract method to evaluate constraint values.

Evaluate just the constraint function values using the attributes decision_vectors and objective_vectors

Note

Currently not supported by ScalarMOProblem

get_variable_bounds()

Abstract method to get variable bounds

get_objective_names()