customhys

<div> <img align="left" alt="Module Dependency Diagram" src="https://raw.githubusercontent.com/jcrvz/customhys/master/docfiles/chm_logo.png" title="Customhys logo" width="25%"/> </div> <div align="justify"> <b>Customising optimisation metaheuristics via hyper-heuristic search</b> (CUSTOMHyS). This framework provides tools for solving, but not limited to, continuous optimisation problems using a hyper-heuristic approach for customising metaheuristics. Such an approach is powered by a strategy based on Simulated Annealing. Also, several search operators serve as building blocks for tailoring metaheuristics. They were extracted from ten well-known metaheuristics in the literature. </div>

Detailed information about this framework can be found in [1, 2]. Plus, the code for each module is well-documented.

🛠 Requirements:

Check the requirements.txt file.
For Apple Silicon, one may need to install TensorFlow via conda such as:

conda install -c apple tensorflow-deps

Further information can be found at Install TensorFlow on Mac M1/M2 with GPU support by D. Ganzaroli.

🧰 Modules

The modules that comprise this framework depend on some basic Python packages, as well as they liaise each other. The module dependency diagram is presented as follows:

Module Dependency Diagram

NOTE: Each module is briefly described below. If you require further information, please check the corresponding source code.

🤯 Problems (benchmark functions)

This module includes several benchmark functions as classes to be solved by using optimisation techniques. The class structure is based on Keita Tomochika's repository optimization-evaluation.

Source: benchmark_func.py

👯‍♂️ Population

This module contains the class Population. A Population object corresponds to a set of agents or individuals within a problem domain. These agents themselves do not explore the function landscape, but they know when to update the position according to a selection procedure.

Source: population.py

🦾 Search Operators (low-level heuristics)

This module has a collection of search operators (simple heuristics) extracted from several well-known metaheuristics in the literature. Such operators work over a population, i.e., modify the individuals' positions.

Source: operators.py

🤖 Metaheuristic (mid-level heuristic)

This module contains the Metaheuristic class. A metaheuristic object implements a set of search operators to guide a population in a search procedure within an optimisation problem.

Source: metaheuristic.py

👽 Hyper-heuristic (high-level heuristic)

This module contains the Hyperheuristic class. Similar to the Metaheuristic class, but in this case, a collection of search operators is required. A hyper-heuristic object searches within the heuristic space to find the sequence that builds the best metaheuristic for a specific problem.

Source: hyperheuristic.py

🏭 Experiment

This module contains the Experiment class. An experiment object can run several hyper-heuristic procedures for a list of optimisation problems.

Source: experiment.py

🗜️ Tools

This module contains several functions and methods utilised by many modules in this package.

Source: tools.py

🧠 Machine Learning

This module contains the implementation of Machine Learning models which can power a hyper-heuristic model from this framework. In particular, it is implemented a wrapper for a Neural Network model from Tensorflow. Also, contains auxiliar data structures which process sample of sequences to generate training data for Machine Learning models.

Source: machine_learning.py

💾 Data Structure

The experiments are saved in JSON files. The data structure of a saved file follows a particular scheme described below.

<details> <summary> Expand structure </summary> <p>

data_frame = {dict: N}
|-- 'problem' = {list: N}
|  |-- 0 = {str}
:  :
|-- 'dimensions' = {list: N}
|  |-- 0 = {int}
:  :
|-- 'results' = {list: N}
|  |-- 0 = {dict: 6}
|  |  |-- 'iteration' = {list: M}
|  |  |  |-- 0 = {int}
:  :  :  :
|  |  |-- 'time' = {list: M}
|  |  |  |-- 0 = {float}
:  :  :  :
|  |  |-- 'performance' = {list: M}
|  |  |  |-- 0 = {float}
:  :  :  :
|  |  |-- 'encoded_solution' = {list: M}
|  |  |  |-- 0 = {int}
:  :  :  :
|  |  |-- 'solution' = {list: M}
|  |  |  |-- 0 = {list: C}
|  |  |  |  |-- 0 = {list: 3}
|  |  |  |  |  |-- search_operator_structure
:  :  :  :  :  :
|  |  |-- 'details' = {list: M}
|  |  |  |-- 0 = {dict: 4}
|  |  |  |  |-- 'fitness' = {list: R}
|  |  |  |  |  |-- 0 = {float}
:  :  :  :  :  :
|  |  |  |  |-- 'positions' = {list: R}
|  |  |  |  |  |-- 0 = {list: D}
|  |  |  |  |  |  |-- 0 = {float}
:  :  :  :  :  :  :
|  |  |  |  |-- 'historical' = {list: R}
|  |  |  |  |  |-- 0 = {dict: 5}
|  |  |  |  |  |  |-- 'fitness' = {list: I}
|  |  |  |  |  |  |  |-- 0 = {float}
:  :  :  :  :  :  :  :
|  |  |  |  |  |  |-- 'positions' = {list: I}
|  |  |  |  |  |  |  |-- 0 = {list: D}
|  |  |  |  |  |  |  |  |-- 0 = {float}
:  :  :  :  :  :  :  :  :
|  |  |  |  |  |  |-- 'centroid' = {list: I}
|  |  |  |  |  |  |  |-- 0 = {list: D}
|  |  |  |  |  |  |  |  |-- 0 = {float}
:  :  :  :  :  :  :  :  :
|  |  |  |  |  |  |-- 'radius' = {list: I}
|  |  |  |  |  |  |  |-- 0 = {float}
:  :  :  :  :  :  :  :
|  |  |  |  |  |  |-- 'stagnation' = {list: I}
|  |  |  |  |  |  |  |-- 0 = {int}
:  :  :  :  :  :  :  :
|  |  |  |  |-- 'statistics' = {dict: 10}
|  |  |  |  |  |-- 'nob' = {int}
|  |  |  |  |  |-- 'Min' = {float}
|  |  |  |  |  |-- 'Max' = {float}
|  |  |  |  |  |-- 'Avg' = {float}
|  |  |  |  |  |-- 'Std' = {float}
|  |  |  |  |  |-- 'Skw' = {float}
|  |  |  |  |  |-- 'Kur' = {float}
|  |  |  |  |  |-- 'IQR' = {float}
|  |  |  |  |  |-- 'Med' = {float}
|  |  |  |  |  |-- 'MAD' = {float}
:  :  :  :  :  :

where:

N is the number of files within data_files folder
M is the number of hyper-heuristic iterations (metaheuristic candidates)
C is the number of search operators in the metaheuristic (cardinality)
P is the number of control parameters for each search operator
R is the number of repetitions performed for each metaheuristic candidate
D is the dimensionality of the problem tackled by the metaheuristic candidate
I is the number of iterations performed by the metaheuristic candidate
search_operator_structure corresponds to [operator_name = {str}, control_parameters = {dict: P}, selector = {str}]

</p> </details>

🏗️ Work-in-Progress

The following modules are available, but they may do not work. They are currently under developing.

🌡️ Characterisation

This module intends to provide metrics for characterising the benchmark functions.

Source: characterisation.py

📊 Visualisation

This module intends to provide several tools for plotting results from the experiments.

Source: visualisation.py

References

Seminal Papers

The seminal papers that describe the framework's theoretical background and software implementation are:

Published Journal Papers

These are the journal articles that have been published using this framework:

G. C. Duque-Gimenez, D. F. Zambrano-Gutierrez, M. Rodriguez-Nieto, J. L. Menchaca, J. M. Cruz-Duarte, D. G. Zárate-Triviño, J. G. Avina-Cervantes, J. C. Ortiz-Bayliss, Viscoelastic Characterization of the Human Osteosarcoma Cancer Cell Line MG-63 Using a Fractional-Order Zener Model Through Automated Algorithm Design and Configuration, Nature Scientific Reports, 15, 31436, 2025.
D. F. Zambrano-Gutierrez, J. M. Cruz-Duarte, H. Castañeda, J. G. Avina-Cervantes, Optimization of Adaptive Sliding Mode Controllers Using Customized Metaheuristics in DC-DC Buck-Boost Converters, Swarm and Evolutionary Computation, 12(23), 370, 2024.
[J. M. Tapia-Avitia, J. M. Cruz‐Duarte, I. Amaya, J. C. Ortiz-Bayliss, H. Terashima-Marín, and N. Pillay, Analysing Hyper-Heuristics based on Neural Networks for the Automatic Design of Population-based Metaheuristics in Continuous Optimisation Problems, Swarm and Evolutionary Computation, 89, 101616, 2024.](https://doi.org/10.1016/j.swev

Customhys

Install / Use

README