Black Box Optimizer

General black-box optimization mainly aiming for optimizing hyperparameters of Machine Learning algorithms.

Current implementations

• Grid Search
• Random Search [1]
• Bayesian Search [2]

Getting started

Installation

    git clone https://github.com/jjakimoto/BBOptimizer
    cd BBOptimizer
    python setup.py install

Dependencies:

• scikit-learn
• numpy
• scipy
• GPy

You can install dependencies by running:

pip install -r requirements.txt

Example

Let's see how to use function while going through optimizing hyperparameters of support vector machine from scikit-learn (SVC).

First, we define the type of parameters from four kinds: continuous, integer, categorical, and fixed. fixed parameters will just be added to additional parameters for score function and not be included in optimization. Here is the example:

space_conf = [
    {'name': 'C', 'domain': (1e-8, 1e5), 'type': 'continuous', 'scale': 'log'},
    {'name': 'gamma', 'domain': (1e-8, 1e5), 'type': 'continuous', 'scale': 'log'},
    {'name': 'kernel', 'domain': 'rbf', 'type': 'fixed'}
]

If you define {'scale': 'log'}, the parameter is optimized and sampled after transformed into 10 ** x. We also have dimensionality option. If you add {'dimensionality': N}, for variable x, you would use x_1, x_2, ..., x_N with the same search space.

Next we need to define a problem to be optimized. Score function has to get argument in the dictionary format and then return scalar values (currently support only single output case). Here is the example of accuracy score function of support vector machine classifier.

from sklearn.svm import SVC
from sklearn.datasets import make_classification
from sklearn.metrics import accuracy_score
from sklearn.model_selection import StratifiedShuffleSplit
from sklearn.preprocessing import StandardScaler


data, target = make_classification(n_samples=2500,
                                   n_features=45,
                                   n_informative=5,
                                   n_redundant=5)

def score_func(params):
    splitter = StratifiedShuffleSplit(n_splits=1, test_size=0.2)
    train_idx, test_idx = list(splitter.split(data, target))[0]
    train_data = data[train_idx]
    train_target = target[train_idx]
    clf = SVC(**params)
    clf.fit(train_data, train_target)
    pred = clf.predict(data[test_idx])
    true_y = target[test_idx]
    score = accuracy_score(true_y, pred)
    return -score

Then, we can execute optimization in the following way:

from bboptimizer import Optimizer

opt = Optimizer(score_func, params_conf, sampler="bayes",
                r_min=10, backend="gpy")
opt.search(num_iter=20)

Here is the list of Optimizer's __init__ arguments

score_func : function
    Takes dictionary as input and returns scalar score.
space: list(dict)
    Define search space. Each element has to the following key
    values: 'name', 'type', and 'domain' (,'num_grid' is optional).
sampler: str
    The name of sample to use: 'grid', 'random', and 'bayes'
init_X: array-like(float), shape=(n_samples, n_dim)
    The list of parameters to initialize sampler
init_y: array-like(float), shape(n_samples,)
    The list of score of init_X
timeout: int, optional
    If specified, it terminates score evaluation after
    timeout seconds has passed.
kwargs:
    These parameters are sent to sampler object

sampler and backend parameters mainly define your optimization method.

sampler parameter defines which optimization in use. It has the following options: grid, random, and bayes.

backend parameter just determines which gaussian process to use for bayesian optimization. It has to be either of gpy or sklearn. Default is gpy.

References

• [1] Random Search for Hyper-Parameter Optimization
• [2] Practical Bayesian Optimization of Machine Learning Algorithms