io-shaker

io-shaker provides Python bindings for the C++ Intelligent Optimization heuristics. The package exposes two local search heuristics for continuous function optimization:

• the inertial shaker [1, 3]: a "search box" is maintained around the current best point and its sizes along the coordinates are expanded and contracted according to the outcomes of explorative evaluations along the coordinated axes;

• the affine shaker [1, 2, 3], which maintains a more flexible search region whose size and shape are again determined by explorative evaluations.

References

[1] R. Battiti and G. Tecchiolli. Learning with first, second, and no derivatives: a case study in high energy physics. Neurocomputing, 6:181–206, 1994.

[2] M. Brunato and R. Battiti. RASH: A self-adaptive random search method. In Carlos Cotta, Marc Sevaux, and Kenneth Sörensen, editors, Adaptive and Multilevel Meta- heuristics, volume 136 of Studies in Computational Intelligence. Springer, 2008. ISBN 978-3-540-79437-0.

[3] Roberto Battiti, Kevin Tierney, Mauro Brunato. Intelligent Optimization - Optimization meets Machine Learning. LION association, Italy, 2026. Downloadable at https://intelligent-optimization.org/iobook/

Installing

From PyPi

Install the io-shaker package:

python -m pip install io-shaker

The package contains binding to a C++ library; therefore, a C++ compiler and Python development headers may be required if a suitable pre-compiled platform wheel is not found.

In such case, please follow your platform's instructions to install the compiler and Python headers, then repeat the above command.

Note: while the official package name is io-shaker(with a dash), the import command requires an underscore:

from io_shaker import reactive_affine_shaker, inertial_shaker

From the source tree

A C++ compiler and Python development headers are required when installing from the source tree. SWIG is not required because its generated files, io_shaker.py and io_shaker_wrap.cxx, are included in the package.

python -m pip install .

To build a wheel for the current Python and platform:

python -m pip wheel .

Setuptools selects the platform compiler and gives the native _io_shaker extension the correct filename for Linux, macOS, or Windows.

If needed, the SWIG-generated files io_shaker.py and io_shaker_wrap.cxx can be recreated by

swig -python -c++ io_shaker.i

Native library

The following cross-platform commands compile a standalone C++ library (cmake build system needed):

cmake -S . -B cbuild
cmake --build cbuild --config Release

Usage and example

Python library

See test/test.py:

from io_shaker import reactive_affine_shaker, inertial_shaker

# Function to be optimized
def f(x: list[float]) -> float:
    return (x[0]-.5)**2 + (x[1]-.45)**2

# OPTIONAL: called whenever current best is updated
def print_update(evals: int, value: float, x: list[float]) -> None:
    print('\tNew best @', evals, x, value)

# Test the solvers 10 times with different seeds
for seed in range(10):
    best_x, best_f = reactive_affine_shaker(f, 2, seed=seed, update_callback=print_update)
    print('BEST AFFINE', best_x, best_f)
    best_x, best_f = inertial_shaker(f, 2, seed=seed, update_callback=print_update)
    print('BEST INERTIAL', best_x, best_f)

The two functions reactive_affine_shaker and inertial_shaker accept the following parameters:

• function (mandatory, type Callable[[list[float]],float]): a Python function accepting a float array and returning a float value.
• dimension (mandatory, type int): the function's domain dimension.
• seed (type int, default 0): randome number generator seed, for reproducibility.
• lb (type list[float], default [0.0]*dimension): domain lower bounds.
• ub (type list[float], default [1.0]*dimension): domain upper bounds.
• fraction (type float, default .1): initial size of the search box as a fraction of the [lb,ub] interval along each dimension.
• reducefactor (type float, default .9): search box reduction factor upon failed improvement.
• expandfactor (type float, default 1.0/reducefactor): search box expansion factor upon succeeded improvement.
• update_callback (type Callable[[int,float,list[float]],None], default None): if passed, function to be invoked at every best value update; receives three arguments: number of function evaluations, best value, and optimizer coordinates.

Both functions return a two-element tuple containing the minimizer coordinates and the corresponding value (Tuple[List[float],float]).

Native library

See test/test.cpp:

#include <iostream>

#include "function.h"
#include "c_io_shaker.h"

using namespace std;

class MyFunction: public Function {
protected:
    // variable holding the function's latest evaluation
    double result_value;
    // evaluation function
    virtual const double *evaluate (const double *x);
public:
    // Constructor
    MyFunction();
};

MyFunction::MyFunction():
    // base constructor: domain and codomain dimensions, random rotations (useful for tests)
    Function(2, 1, false)
{
    // Basic initialization: lower and upper bound for each coordinate
    xmin[0] = xmin[1] = 0.0;
    xmax[0] = xmax[1] = 1.0;
}

const double *MyFunction::evaluate(const double *x) {
    double
        x0 = x[0]-.5,
        x1 = x[1]-.45;
    result_value = x0*x0 + x1*x1;
    // The function returns a pointer to the value
    return &result_value;
}

int main() {
    MyFunction f;
    double best_x[2];
    // Test the solvers ten times with different seeds
    for (int seed = 0; seed < 10; seed++ ) {
        double best_f = reactive_affine_shaker(f, best_x, seed);
        cout << "BEST AFFINE " << best_f << " @ " << best_x[0] << ',' << best_x[1] << endl;
        best_f = inertial_shaker(f, best_x, seed);
        cout << "BEST INERTIAL " << best_f << " @ " << best_x[0] << ',' << best_x[1] << endl;
    }
    return 0;
}

To compile from inside test, add .. and ../libRSO as include directories and link the object file with ../cbuild/libIOShaker.a:

cd test
g++ -c -I.. -I../libRSO test.cpp
g++ -o test -L../cbuild test.o -lIOShaker
./test

The two functions reactive_affine_shaker and inertial_shaker accept a parameter list similar to their Python counterparts:

• function (class Function): a reference to an object from a subclass of Function.
• best_x (type double[function.dimension]): a double array apt at containing the minimizer's coordinates.
• seed (type int, default 0): randome number generator seed, for reproducibility.
• fraction (type double, default .1): initial size of the search box as a fraction of the [lb,ub] interval along each dimension.
• reducefactor (type double, default .9): search box reduction factor upon failed improvement.
• expandfactor (type double, default 1.0/reducefactor): search box expansion factor upon succeeded improvement.

The differences from the Python functions are the following:

• function is a reference to an object from a subclass of Function instead of a generic callable object.
• best_x must be provided as argument, while in the Python version it is created and returned by the solver.
• parameters dimension, lb and ub are omitted because the domain is specified in the dimension, xmin and xmax members of Function.

As shown in the example above, the Function subclass must define at least two methods:

• a constructor invoking the one from the base class, setting the domain's and codomain's dimension (the latter must be 1), and filling the lower and upper bound arrays;
• a protected evaluate function that accepts a constant double array and provides a pointer to a double variable containing the result.