Getting Started

Download

Clone the repository available on Github

$ git clone https://github.com/yassine-laguel/taco.git
$ cd taco/

Installation

Taco requires several packages one can install through conda and the provided yaml file taco_env.yml:

$ conda env create --file taco_env.yml
$ source activate taco_env

General Usage

The user must provide an instance of the class Problem. Such class has a data attribute which is a dataset stored in the form of a numpy array. The class must provide also 4 implemented methods:

• the objective function and its gradient.

• the constraint function and its gradient.

This problem instance is then used together with a dictionnary of parameters to initialize the Optimizer instance for solving.

Simple Case Demo

Consider the quadratic chance constrained toy problem from our companion paper.

import numpy as np
np.random.seed(42)
mean = np.array([1.0, 1.0])
cov = 20 * np.eye(2)
nb_samples=10000
data = np.random.multivariate_normal(mean, cov, size=nb_samples)
data = np.asarray(data, dtype=np.float64)

We instantiate then a quadratic problem with this dataset. Implementation of this toy problem can be found here.

problem = ToyProblem(data)

A number of parameters can then be provided to instanciate the class Optimizer.

from taco import Optimizer

other_selected_params = {
            'bund_mu_high' : 0.001  # upper bound for the proximal parameter of the bundle
            'bund_mu_low' : 0.001  # lower bound for the proximal parameter of the bundle

            'bund_max_size_bundle_set': 30,  # Maximum size for the smoothing constant
            'superquantile_smoothing_param' : 0.1,  # smoothing constant
        }

safety_proba_threshold =  0.03368421
pb_numba_compliant = True  # Set to true if the input problem is a jitclass.

optimizer = Optimizer(problem, p=0.03368421, numba=pb_numba_compliant, params=other_selected_params)

The user can then run the bundle method and retrieve the solution found.

solution = optimizer.run()