PyQEA 0.1.5

General use optimizer for non-convex cost functions with non-linear constraints

PyQEA

Library for Quantum Inspired optimization in python

PyQEA is a extensive research library for Quantum inspired hyper-normal based optimization in python.

It is intended for the solution of global optimization problems where conventional genetic algorithms or PSO yield sub-optimal results. The current implementation of the algorithm allows for a fast deployment of any optimization problem, regardless of the non-linearity of its constraints or the complexity of the cost function.

The library has the following features:

Features

• High level module for Quantum Inspired optimization
• Built-in set of objective cost functions to test the optimization algorithm
• Capacity to implement non-linear restrictions
• Capcity to implement integral-only variables

To install PyQEA, run this command in your terminal (Currently in test PyPI):

$ pip install -i https://test.pypi.org/simple/ PyQEA==0.1.5

To install the development version as it is, clone the development branch of the repo and then run:

$ cd PyQEA
$ python setup.py install

or:

$ cd PyQEA
$ pip install .

Basic Usage:

PyQEA provides a high level implementation of the proposed Quantum Inspired algorithm that allows a fast implementation and usage. It aims to be user-friendly despite the non-trivial nature of its hyper-parameters. We now show the optimization process of a paraboloid (Sphere function) of input dimension n centered in the vector: [3.8, 3.8, 3.8, 3.8, ...].

Use case example:

The optimizer setup is as follows:

import numpy as np

from PyQEA import QuantumEvAlgorithm
from PyQEA.utils.cost_functions import f

n_dims = 10 # Input dimensions of f(x)
up = 5.12 *np.ones(n_dims) # Upper bound defined for the input variables
low = -5*np.ones(n_dims)  # Lower bound defined for the input variables

integrals = np.full(n_dims, False) #Boolean vector defining which variables are integral

cost_function = f

optimizer = QuantumEvAlgorithm(cost_function, n_dims=n_dims, upper_bound=up,
                                     lower_bound=low, integral_id=integrals,
                                     sigma_scaler=1.003,
                                     mu_scaler=20, elitist_level=6,
                                     restrictions=[])

results = optimizer.training(N_iterations=4000, sample_size=20)

The following example showcases the usage of the optimizer for a constrained-optimization problem

from PyQEA import QuantumEvAlgorithm
from PyQEA.utils.cost_functions import f

def h(x: np.ndarray):
    """Definition of the restriction to be applied to the opt problem"""

    if x.ndim == 1:
        x = x[None]

    n_dims = x.shape[1]

    return  -x[:,1 ] - x[:,0] + 6.1

n_dims = 10
up = 5*np.ones(n_dims)
low = -5*np.ones(n_dims)
integrals = np.full(n_dims, False)
integrals[0:2] = True

optimizer = QuantumEvAlgorithm(f, n_dims=n_dims, upper_bound=up,
                                     lower_bound=low, integral_id=integrals,
                                     sigma_scaler=1.003,
                                     mu_scaler=20, elitist_level=6,
                                     restrictions=[h])

results = optimizer.training(N_iterations=4000, sample_size=20, save=False,
                             filename='q11.npz')

Parameter tuning

The main limitation that the user may encounter in the use of this optimizer is the non-trivial character of it's hyper-parameters. The critical hyper-parameters are the ones that regulate the update of hyper-normal distribution after the evaluation of the sampled population.

The recommended rule of thumb is the following:

• mu_scaler ~ 20 (It is not as critical for performance)
• sigma_scaler ~ (1 + 1/(10*n)) being n the number of input dimensions of the problem

The key concept to bear in mind is that, as the dimensionality of the problem increases, it is necessary to make the algorithm more "cautious", therefore minimizing the difference between before and after distributions. In practical terms, as the complexity of a given problem increases, sigma_scaler must tend to ~1.

Pending to be uploaded to PyPI