jijzepttools.blackbox_optimization.surrogate_model

Submodules

• jijzepttools.blackbox_optimization.surrogate_model.bfmo_surrogate_model
• jijzepttools.blackbox_optimization.surrogate_model.bocs_surrogate_model
• jijzepttools.blackbox_optimization.surrogate_model.fmo_surrogate_model
• jijzepttools.blackbox_optimization.surrogate_model.interface

Classes

BFMParams	Parameters for the Bayesian Factorization Machine surrogate model.
BFMSurrogateModel	Bayesian Factorization Machine (BFM) surrogate model.
BOCSParams	Parameters for BOCS (Bayesian Optimization of Combinatorial Structures) surrogate model.
BOCSSurrogateModel	BOCS (Bayesian Optimization of Combinatorial Structures) surrogate model.
FMParams	Parameters for Factorization Machine surrogate model.
FMSurrogateModel	Factorization Machine surrogate model implementation.
SurrogateModel	Abstract base class for surrogate models in blackbox optimization.
SurrogateModelParams	Base class for all surrogate model parameters.

Package Contents

class BFMParams

Bases: jijzepttools.blackbox_optimization.surrogate_model.interface.SurrogateModelParams

Parameters for the Bayesian Factorization Machine surrogate model.

init_stdev, alpha0, beta0, gamma_0, mu_0, reg_0の詳細は以下を参照 参考: tohtsky/myFM

latent_dim

Dimension of latent factors.

Type:

int, default=5

tune

Number of burn-in sweeps for Gibbs sampling.

Type:

int, default=100

draws

Number of samples to average after burn-in.

Type:

int, default=10

init_stdev

FMのパラメータ初期値の標準偏差

Type:

float

alpha0, beta0

alpha, lambda_w, lambda_vの事前ガンマ分布の形状を決定するパラメータ

Type:

float

gamma_0

パラメータの正規事前分布の平均mu_w, mu_vが従うハイパー事前分布(正規分布)の精度

Type:

float

mu_0

パラメータの正規事前分布の平均mu_w, mu_vが従うハイパー事前分布(正規分布)の平均

Type:

float

reg_0

バイアス項 w0が従う正規事前分布の精度

Type:

float

latent_dim: int = 5

tune: int = 100

draws: int = 10

init_stdev: float = 0.1

alpha_0: float = 1.0

beta_0: float = 1.0

gamma_0: float = 1.0

mu_0: float = 0.0

reg_0: float = 1.0

class BFMSurrogateModel(params: BFMParams | None = None, random_seed: int = 42)

Bases: jijzepttools.blackbox_optimization.surrogate_model.interface.SurrogateModel

Bayesian Factorization Machine (BFM) surrogate model.

This model wraps the BFM implementation and provides the SurrogateModel interface. It’s effective for problems where interactions between variables are important.

Parameters:

params (BFMParams, optional) – Parameters for the model. If None, default parameters are used.

params = None

bfm_regressor

fit(x: numpy.ndarray, y: numpy.ndarray) → None

Fit the BFM model to the data using Gibbs sampling.

Parameters:

• x (np.ndarray, shape (n_samples, n_features)) – Input features.

• y (np.ndarray, shape (n_samples,)) – Target values.

predict(x: numpy.ndarray) → numpy.ndarray

Predict target values using the fitted BFM model.

surrogate_to_ommx_objective(decision_variables: list[ommx.v1.DecisionVariable])

Create an OMMX expression from the fitted BFM model.

Parameters:

decision_variables (list[ommx.v1.DecisionVariable]) – Decision variables from the OMMX instance.

class BOCSParams

Bases: jijzepttools.blackbox_optimization.surrogate_model.interface.SurrogateModelParams

Parameters for BOCS (Bayesian Optimization of Combinatorial Structures) surrogate model.

random_seed

Random seed for reproducibility.

Type:

int, optional

draws

Number of MCMC samples to draw.

drawed samples are used for the mean of coefficients for the optimization.

Type:

int, default=10

tune

Number of burn-in iterations.

Type:

int, default=100

warm_start

Whether to continue from the previous Gibbs chain.

Type:

bool, default=False

random_seed: int | None = None

draws: int = 10

tune: int = 100

warm_start: bool = False

class BOCSSurrogateModel(params: BOCSParams | None = None)

Bases: jijzepttools.blackbox_optimization.surrogate_model.interface.SurrogateModel

BOCS (Bayesian Optimization of Combinatorial Structures) surrogate model.

This model uses sparse Bayesian linear regression on quadratic features to approximate the objective function. It automatically generates quadratic interaction terms and applies Bayesian regularization.

Parameters:

params (BOCSParams, optional) – Parameters for the model. If None, default parameters are used.

sblr

draws = 10

tune = 100

n_features: int | None = None

triu_indices: numpy.ndarray | None = None

fit(x: numpy.ndarray, y: numpy.ndarray) → None

Fit the BOCS model to the data.

This involves: 1. Generating quadratic interaction terms 2. Fitting sparse Bayesian linear regression

Parameters:

• x (np.ndarray, shape (n_samples, n_features)) – Input features.

• y (np.ndarray, shape (n_samples,)) – Target values.

predict(x: numpy.ndarray) → numpy.ndarray

Predict target values using the fitted BOCS model.

This generates quadratic interaction terms for the input features and uses the fitted model to make predictions.

Parameters:

x (np.ndarray, shape (n_samples, n_features)) – Input features.

Returns:

Predicted target values.

Return type:

np.ndarray, shape (n_samples,)

surrogate_to_ommx_objective(decision_variables: list[ommx.v1.DecisionVariable])

Create OMMX expression for the fitted BOCS model.

Notes: - The order of the decision variables is the same as the order of the features.

Parameters:

decision_variables (list[ommx.v1.DecisionVariable]) – Decision variables from OMMX instance.

class FMParams

Bases: jijzepttools.blackbox_optimization.surrogate_model.interface.SurrogateModelParams

Parameters for Factorization Machine surrogate model.

latent_dim

Dimension of latent factors in FM.

Type:

int, default=5

n_epochs

Number of training epochs.

Type:

int, default=100

optimizer_params

Parameters for PyTorch optimizer (e.g., {‘lr’: 0.001, ‘weight_decay’: 1e-4}).

Type:

dict, default={}

batch_size

Batch size for training.

Type:

int, default=1

latent_dim: int = 5

n_epochs: int = 100

optimizer_params: dict

batch_size: int = 1

class FMSurrogateModel(n_features: int, params: FMParams | None = None)

Bases: jijzepttools.blackbox_optimization.surrogate_model.interface.SurrogateModel

Factorization Machine surrogate model implementation.

This model uses second-order feature interactions to approximate the objective function. The mathematical form is: f(x) = w0 + Σ wi*xi + ΣΣ <vi,vj>*xi*xj

Parameters:

• n_features (int) – Number of input features.

• params (FMParams, optional) – Parameters for the model. If None, default parameters are used.

n_features

fm_trainer

fit(x: numpy.ndarray, y: numpy.ndarray) → None

Fit the Factorization Machine to the data.

Parameters:

• x (np.ndarray, shape (n_samples, n_features)) – Input features.

• y (np.ndarray, shape (n_samples,)) – Target values.

predict(x: numpy.ndarray) → numpy.ndarray

Predict target values using the fitted FM model.

surrogate_to_ommx_objective(decision_variables: list[ommx.v1.DecisionVariable])

Create OMMX expression for the fitted FM model.

Parameters:

decision_variables (list[ommx.v1.DecisionVariable]) – Decision variables from OMMX instance.

class SurrogateModel

Bases: abc.ABC

Abstract base class for surrogate models in blackbox optimization.

This interface provides a unified way to interact with different surrogate models such as Factorization Machine and BOCS. The typical workflow is: 1. Instantiate a concrete surrogate model with parameters 2. Fit the model to observed data using fit() 3. Create optimization objective using create_optimization_model()

abstract fit(x: numpy.ndarray, y: numpy.ndarray) → None

Fit the surrogate model to observed data.

Parameters:

• x (np.ndarray, shape (n_samples, n_features)) – Input features from blackbox function evaluations.

• y (np.ndarray, shape (n_samples,)) – Target values for a single objective.

abstract surrogate_to_ommx_objective(decision_variables: list[ommx.v1.DecisionVariable])

Create an OMMX expression representing the surrogate model objective.

Parameters:

decision_variables (list[ommx.v1.DecisionVariable]) – List of decision variables from the OMMX instance.

class SurrogateModelParams

Base class for all surrogate model parameters.

This class serves as a base for parameter classes used by different surrogate models (e.g., FMParams, BOCSParams). Each concrete implementation should define the specific parameters needed for that model type.

This provides a unified interface for parameter classes of different surrogate models, allowing for type-safe and extensible parameter handling.