Professional Documents
Culture Documents
Download ebook Applications Of Evolutionary Computation 24Th International Conference Evoapplications 2021 Held As Part Of Evostar 2021 Virtual Event April 7 9 Lecture Notes In Computer Science 12694 Pedro A Casti online pdf all chapter docx epub
Download ebook Applications Of Evolutionary Computation 24Th International Conference Evoapplications 2021 Held As Part Of Evostar 2021 Virtual Event April 7 9 Lecture Notes In Computer Science 12694 Pedro A Casti online pdf all chapter docx epub
Download ebook Applications Of Evolutionary Computation 24Th International Conference Evoapplications 2021 Held As Part Of Evostar 2021 Virtual Event April 7 9 Lecture Notes In Computer Science 12694 Pedro A Casti online pdf all chapter docx epub
https://ebookmeta.com/product/graph-transformation-14th-
international-conference-icgt-2021-held-as-part-of-
staf-2021-virtual-event-june-24-25-2021-proceedings-1st-edition-
fabio-gadducci/
Editorial Board
Elisa Bertino
Purdue University, West Lafayette, IN, USA
Wen Gao
Peking University, Beijing, China
Bernhard Steffen
TU Dortmund University, Dortmund, Germany
Gerhard Woeginger
RWTH Aachen, Aachen, Germany
Moti Yung
Columbia University, New York, NY, USA
Founding Editors
Gerhard Goos
Karlsruhe Institute of Technology, Karlsruhe, Germany
Juris Hartmanis
Cornell University, Ithaca, NY, USA
More information about this subseries at http://www.springer.com/
series/7407
Editors
Pedro A. Castillo and Juan Luis Jiménez Laredo
Applications of Evolutionary
Computation
24th International Conference, EvoApplications
2021, Held as Part of EvoStar 2021, Virtual Event,
April 7–9, 2021, Proceedings
1st ed. 2021
Editors
Pedro A. Castillo
ETSIIT-CITIC, University of Granada, Granada, Spain
The publisher, the authors and the editors are safe to assume that the
advice and information in this book are believed to be true and accurate
at the date of publication. Neither the publisher nor the authors or the
editors give a warranty, expressed or implied, with respect to the
material contained herein or for any errors or omissions that may have
been made. The publisher remains neutral with regard to jurisdictional
claims in published maps and institutional affiliations.
Organizing Committee
EvoApplications Conference Chair
Pedro A. Castillo Universidad de Granada, Spain
Local Chair
Federico Divina Universidad Pablo de Olavide, Spain
Publicity Chair
João Correia University of Coimbra, Portugal
Program Committee
Ahmed Kattan Umm Al-Qura University, Saudi Arabia
Aladdin Ayesh De Montfort University, UK
Alberto Tonda INRA, France
Aleš Zamuda University of Maribor, Slovenia
Alessandra Scotto di Freca Università degli studi di Cassino e del Lazio
Meridionale, Italy
Alessandro Niccolai Politecnico di Milano, Italy
Amir Dehsarvi University of Aberdeen, UK
Anabela Simões Coimbra Institute of Engineering, Portugal
Anca Andreica Babeș-Bolyai University, Romania
Anders Christensen University of Southern Denmark, Denmark
Andrea Tettamanzi Université Cô te d’Azur, France
Andres Faina IT University of Copenhagen, Denmark
Andrew Turner Freelance Researcher, UK
Anil Yaman Korea Advanced Institute of Science and Technology,
Korea
Anna Paszyńska Jagiellonian University, Poland
Anthony Clark Pomona College, USA
Antonio Fernández Ares University of Granada, Spain
Antonio Mora García University of Granada, Spain
Antonio Córdoba University of Seville, Spain
Antonio Della Cioppa University of Salerno, Italy
Antonio González Universidad Rey Juan Carlos, Spain
Antonio J. Fernández Leiva Universidad de Má laga, Spain
Arkadiusz Poteralski Silesian University of Technology, Poland
Bernabé Dorronsoro University of Cá diz, Spain
Bing Xue Victoria University of Wellington, New Zealand
Carlotta Orsenigo Polytechnic University of Milan, Italy
Cédric Buche CNRS CERV - Centre Européen de Réalité Virtuelle,
France
Changhe Li China University of Geosciences, China
Chien-Chung Shen University of Delaware, USA
Clara Pizzuti CNR-ICAR, Italy
Daniel Hernandez Data Frontier/Instituto Tecnoló gico de Tijuana,
México
Daniele Gravina University of Malta, Malta
David Megías Universitat Oberta de Catalunya, Spain
David Pelta University of Granada, Spain
Dávid Melhárt University of Malta, Malta
Diego Perez Liebana Queen Mary University of London, UK
Doina Bucur University of Twente, Netherlands
Edoardo Fadda Politecnico di Torino, Italy
Enrico Schumann University of Basel, Switzerland
Ernesto Tarantino ICAR-CNR, Italy
Evelyne Lutton INRAE, France
Fabio Caraffini De Montfort University, UK
Fabio D’Andreagiovanni CNRS, Sorbonne University - UTC, France
Federico Liberatore Cardiff University, UK
Federico Divina Pablo de Olavide University, Spain
Fernando Lobo University of Algarve, Portugal
Ferrante Neri University of Nottingham, UK
Francesco Fontanella Università di Cassino e del Lazio Meridionale,
Italy
Francisco Chávez Universidad de Extremadura, Spain
Francisco Luna Universidad de Má laga, Spain
Francisco Chicano University of Má laga, Spain
Francisco Fernández de Vega Universidad de Extremadura, Spain
Gabriel Luque University of Má laga, Spain
Geoff Nitschke University of Cape Town, South Africa
Giovanni Fasano University Ca’Foscari of Venice, Italy
Giovanni Squillero Politecnico di Torino, Italy
Giovanni Iacca University of Trento, Italy
Giulio Biondi University of Florence, Italy
Grégoire Danoy University of Luxembourg, Luxembourg
Gregory Gay Chalmers and the University of Gothenburg, Sweden
Günter Rudolph TU Dortmund University, Germany
Guillermo Gómez Trenado Universidad de Granada, Spain
Gürhan Küçük Yeditepe University, Turkey
Gustavo Olague CICESE, México
Heiko Hamann University of Lü beck, Germany
Huthaifa Aljawazneh University of Granada, Spain
Ignacio Hidalgo Universidad Complutense de Madrid, Spain
Ivanoe De Falco ICAR - CNR, Italy
Jacopo Aleotti University of Parma, Italy
James Foster University of Idaho, USA
János Botzheim Budapest University of Technology and Economics,
Hungary
Jarosław Wąs AGH University of Science and Technology, Poland
Jaume Bacardit Newcastle University, UK
Jesús Mayor Universidad Politécnica de Madrid, Spain
Jörg Bremer University of Oldenburg, Germany
Jorge Novo Buján Universidade da Coruñ a, Spain
José Santos University of A Coruñ a, Spain
José Carlos Ribeiro Polytechnic Institute of Leiria, Portugal
José Manuel Colmenar Universidad Rey Juan Carlos, Spain
Juan Luis Jimenez Université du Havre Normandie, France
Julian Miller University of York, UK
Kenji Leibnitz National Institute of Information and Communications
Technology, Japan
Kevin Sim Edinburgh Napier University, UK
Krzysztof Michalak Wrocław University of Economics, Poland
Laura Dipietro Massachusetts Institute of Technology, USA
Leonardo Bocchi University of Florence, Italy
Maciej Smołka AGH University of Science and Technology, Poland
Marco Tomassini University of Lausanne, Switzerland
Marco Villani University of Modena and Reggio Emilia, Italy
Marco Baioletti Universitá degli Studi di Perugia, Italy
Marcos Ortega Hortas University of A Coruñ a, Spain
Mario Köppen Kyushu Institute of Technology, Japan
Mario Giacobini University of Torino, Italy
Mengjie Zhang Victoria University of Wellington, New Zealand
Michael Lones Heriot-Watt University, UK
Michael Guckert Technische Hochschule Mittelhessen, Germany
Mohamad Alissa Edinburgh Napier University, UK
Mohamed Wiem Mkaouer Rochester Institute of Technology, USA
Monica Mordonini University of Parma, Italy
Nadarajen Veerapen University of Lille, France
Neil Urquhart Edinburgh Napier University, UK
Oscar Castillo Tijuana Institute of Technology, México
Oscar Cordón University of Granada, Spain
Pablo Garca-Sánchez University of Granada, Spain
Paolo Burelli IT University of Copenhagen, Denmark
Paolo Mengoni Hong Kong Baptist University, China
Patricia Paderewski University of Granada, Spain
Pedro A. Castillo Valdivieso University of Granada, Spain
Penousal Machado University of Coimbra, Portugal
Petr Pošk Czech Technical University in Prague, Czech Republic
Philip Bontrager New York University, USA
Rafael Villanueva Instituto Universitario de Matemá tica
Multidisciplinar, Spain
Rafael Nogueras Universidad de Má laga, Spain
Rami Abielmona University of Ottawa, Canada
Raneem Qaddoura Philadelphia University, Jordan
Raul Lara Cabrera Universidad Politécnica de Madrid, Spain
Renato Tinós Universidade de Sã o Paulo, Brazil
Rolf Hoffmann TU Darmstadt, Germany
Mohammed Salem University Mustafa Stmboli, Algeria
Sebastian Risi IT University of Copenhagen, Denmark
Sergio Damas University of Granada, Spain
Sevil Şen Hacettepe University, Turkey
Shamik Sural IIT Kharagpur, India
Simon Wells Edinburgh Napier University, UK
Stefano Cagnoni University of Parma, Italy
Stefano Coniglio University of Southampton, UK
Stephen Smith University of York, UK
Thomas Farrenkopf Technische Hochschule Mittelhessen, Germany
Tiago Baptista University of Coimbra, Portugal
Tien-Tsin Wong The Chinese University of Hong Kong, China
Ting Hu Queen’s University, Canada
Valentino Santucci University for Foreigners of Perugia, Italy
Wacław Kuś Silesian University of Technology, Poland
Wolfgang Banzhaf Michigan State University, USA
Yanan Sun Sichuan University, China
Ying-ping Chen National Chiao Tung University, Taiwan
Yoann Pigné LITIS - Université Le Havre Normandie, France
Contents
Applications of Evolutionary Computation
On Restricting Real-Valued Genotypes in Evolutionary Algorithms
Jørgen Nordmoen, Tønnes F. Nygaard, Eivind Samuelsen and
Kyrre Glette
Towards Explainable Exploratory Landscape Analysis:Extreme
Feature Selection for Classifying BBOB Functions
Quentin Renau, Johann Dreo, Carola Doerr and Benjamin Doerr
Co-optimising Robot Morphology and Controller in a Simulated
Open-Ended Environment
Emma Hjellbrekke Stensby, Kai Olav Ellefsen and Kyrre Glette
Multi-objective Workforce Allocation in Construction Projects
Andrew Iskandar and Richard Allmendinger
Generating Duplex Routes for Robust Bus Transport Network by
Improved Multi-objective Evolutionary Algorithm Based on
Decomposition
Sho Kajihara, Hiroyuki Sato and Keiki Takadama
Combining Multi-objective Evolutionary Algorithms with Deep
Generative Models Towards Focused Molecular Design
Tiago Sousa, Joã o Correia, Vitor Pereira and Miguel Rocha
A Multi-objective Evolutionary Algorithm Approach for Optimizing
Part Quality Aware Assembly Job Shop Scheduling Problems
Michael H. Prince, Kristian DeHaan and Daniel R. Tauritz
Evolutionary Grain-Mixing to Improve Profitability in Farming
Winter Wheat
Md Asaduzzaman Noor and John W. Sheppard
Automatic Modular Design of Behavior Trees for Robot Swarms
with Communication Capabilites
Jonas Kuckling, Vincent van Pelt and Mauro Birattari
Salp Swarm Optimization Search Based Feature Selection for
Enhanced Phishing Websites Detection
Ruba Abu Khurma, Khair Eddin Sabri, Pedro A. Castillo and
Ibrahim Aljarah
Real Time Optimisation of Traffic Signals to Prioritise Public
Transport
Milan Wittpohl, Per-Arno Plö tz and Neil Urquhart
Adaptive Covariance Pattern Search
Ferrante Neri
Evaluating the Success-History Based Adaptive Differential
Evolution in the Protein Structure Prediction Problem
Pedro Henrique Narloch and Má rcio Dorn
Beyond Body Shape and Brain:Evolving the Sensory Apparatus of
Voxel-Based Soft Robots
Andrea Ferigo, Giovanni Iacca and Eric Medvet
Desirable Objective Ranges in Preference-Based Evolutionary
Multiobjective Optimization
Sandra Gonzá lez-Gallardo, Rubén Saborido, Ana B. Ruiz and
Mariano Luque
Improving Search Efficiency and Diversity of Solutions in
Multiobjective Binary Optimization by Using Metaheuristics Plus
Integer Linear Programming
Miguel Á ngel Domínguez-Ríos, Francisco Chicano and Enrique Alba
Automated, Explainable Rule Extraction from MAP-Elites Archives
Neil Urquhart, Silke Hö hl and Emma Hart
Applications of Deep Bioinspired Algorithms
EDM-DRL:Toward Stable Reinforcement Learning Through
Ensembled Directed Mutation
Michael H. Prince, Andrew J. McGehee and Daniel R. Tauritz
Continuous Ant-Based Neural Topology Search
AbdElRahman ElSaid, Joshua Karns, Zimeng Lyu,
Alexander G. Ororbia and Travis Desell
Soft Computing Applied to Games
Playing with Dynamic Systems - Battling Swarms in Virtual Reality
Johannes Bü ttner, Christian Merz and Sebastian von Mammen
EvoCraft:A New Challenge for Open-Endedness
Djordje Grbic, Rasmus Berg Palm, Elias Najarro, Claire Glanois and
Sebastian Risi
A Profile-Based ‘GrEvolutionary’ Hearthstone Agent
Alejandro Romero García and Antonio M. Mora García
Machine Learning and AI in Digital Healthcare and Personalized
Medicine
Modelling Asthma Patients’ Responsiveness to Treatment Using
Feature Selection and Evolutionary Computation
Alejandro Lopez-Rincon, Daphne S. Roozendaal,
Hilde M. Spierenburg, Asta L. Holm, Renee Metcalf, Paula Perez-
Pardo, Aletta D. Kraneveld and Alberto Tonda
Bayesian Networks for Mood Prediction Using Unobtrusive
Ecological Momentary Assessments
Margarita Rebolledo, A. E. Eiben and Thomas Bartz-Beielstein
A Multi-objective Multi-type Facility Location Problem for the
Delivery of Personalised Medicine
Andreea Avramescu, Richard Allmendinger and Manuel Ló pez-
Ibá ñ ez
Evolutionary Computation in Image Analysis, Signal Processing
and Pattern Recognition
RDE-OP:A Region-Based Differential Evolution Algorithm
Incorporation Opposition-Based Learning for Optimising the
Learning Process of Multi-layer Neural Networks
Seyed Jalaleddin Mousavirad, Gerald Schaefer, Iakov Korovin and
Diego Oliva
Estimation of Grain-Level Residual Stresses in a Quenched
Cylindrical Sample of Aluminum Alloy AA5083 Using Genetic
Programming
Laura Millá n, Gabriel Kronberger, J. Ignacio Hidalgo,
Ricardo Ferná ndez, Oscar Garnica and Gaspar Gonzá lez-Doncel
EDA-Based Optimization of Blow-Off Valve Positions for
Centrifugal Compressor Systems
Jacob Spindler, Rico Schulze, Kevin Schleifer and Hendrik Richter
3D-2D Registration Using X-Ray Simulation and CMA-ES
Tianci Wen, Radu P. Mihail and Franck P. Vidal
Lateralized Approach for Robustness Against Attacks in Emotion
Categorization from Images
Harisu Abdullahi Shehu, Abubakar Siddique, Will N. Browne and
Hedwig Eisenbarth
Evolutionary Machine Learning
Improved Crowding Distance in Multi-objective Optimization for
Feature Selection in Classification
Peng Wang, Bing Xue, Jing Liang and Mengjie Zhang
Deep Optimisation:Multi-scale Evolution by Inducing and
Searching in Deep Representations
Jamie Caldwell, Joshua Knowles, Christoph Thies, Filip Kubacki and
Richard Watson
Evolutionary Planning in Latent Space
Thor V. A. N. Olesen, Dennis T. T. Nguyen, Rasmus B. Palm and
Sebastian Risi
Utilizing the Untapped Potential of Indirect Encoding for Neural
Networks with Meta Learning
Adam Katona, Nuno Lourenço, Penousal Machado, Daniel W. Franks
and James Alfred Walker
Effective Universal Unrestricted Adversarial Attacks Using a MOE
Approach
Alina Elena Baia, Gabriele Di Bari and Valentina Poggioni
Improving Distributed Neuroevolution Using Island Extinction and
Repopulation
Zimeng Lyu, Joshua Karns, AbdElRahman ElSaid,
Mohamed Mkaouer and Travis Desell
An Experimental Study of Weight Initialization and Lamarckian
Inheritance on Neuroevolution
Zimeng Lyu, AbdElRahman ElSaid, Joshua Karns,
Mohamed Mkaouer and Travis Desell
Towards Feature-Based Performance Regression Using Trajectory
Data
Anja Jankovic, Tome Eftimov and Carola Doerr
Demonstrating the Evolution of GANs Through t-SNE
Victor Costa, Nuno Lourenço, Joã o Correia and Penousal Machado
Optimising Diversity in Classifier Ensembles of Classification
Trees
Carina Ivaşcu, Richard M. Everson and Jonathan E. Fieldsend
WILDA:Wide Learning of Diverse Architectures for Classification
of Large Datasets
Rui P. Cardoso, Emma Hart, David Burth Kurka and Jeremy Pitt
Evolving Character-Level DenseNet Architectures Using Genetic
Programming
Trevor Londt, Xiaoying Gao and Peter Andreae
Transfer Learning for Automated Test Case Prioritization Using
XCSF
Lukas Rosenbauer, David Pä tzel, Anthony Stein and Jö rg Hä hner
On the Effects of Absumption for XCS with Continuous-Valued
Inputs
Alexander R. M. Wagner and Anthony Stein
A NEAT Visualisation of Neuroevolution Trajectories
Stefano Sarti and Gabriela Ochoa
Evaluating Models with Dynamic Sampling Holdout
Celio H. N. Larcher Jr and Helio J. C. Barbosa
Parallel and Distributed Systems
Event-Driven Multi-algorithm Optimization:Mixing Swarm and
Evolutionary Strategies
Mario García-Valdez and Juan J. Merelo
TensorGP – Genetic Programming Engine in TensorFlow
Francisco Baeta, Joã o Correia, Tiago Martins and Penousal Machado
Applications of Nature-Inspired Computing for Sustainability and
Development
A Novel Evolutionary Approach for IoT-Based Water Contaminant
Detection
Claudio De Stefano, Luigi Ferrigno, Francesco Fontanella,
Luca Gerevini and Mario Molinara
Evolutionary Algorithms for Roughness Coefficient Estimation in
River Flow Analyses
Antonio Agresta, Marco Baioletti, Chiara Biscarini, Alfredo Milani
and Valentino Santucci
EA-Based ASV Trajectory Planner for Pollution Detection in Lentic
Waters
Gonzalo Carazo-Barbero, Eva Besada-Portas, José M. Giró n-Sierra
and José A. Ló pez-Orozco
Author Index
Applications of Evolutionary
Computation
© Springer Nature Switzerland AG 2021
P. A. Castillo, J. L. Jiménez Laredo (eds.), Applications of Evolutionary Computation,
Lecture Notes in Computer Science 12694
https://doi.org/10.1007/978-3-030-72699-7_1
Jørgen Nordmoen
Email: jorgehn@ifi.uio.no
Abstract
Real-valued genotypes together with the variation operators, mutation
and crossover, constitute some of the fundamental building blocks of
Evolutionary Algorithms. Real-valued genotypes are utilized in a broad
range of contexts, from weights in Artificial Neural Networks to
parameters in robot control systems. Shared between most uses of real-
valued genomes is the need for limiting the range of individual
parameters to allowable bounds. In this paper we will illustrate the
challenge of limiting the parameters of real-valued genomes and
analyse the most promising method to properly limit these values. We
utilize both empirical as well as benchmark examples to demonstrate
the utility of the proposed method and through a literature review
show how the insight of this paper could impact other research within
the field. The proposed method requires minimal intervention from
Evolutionary Algorithm practitioners and behaves well under repeated
application of variation operators, leading to better theoretical
properties as well as significant differences in well-known benchmarks.
Keywords Evolutionary algorithms – Bounce-back – Real-value –
Restricting – Genome
1 Introduction
Evolutionary Algorithms (EAs) are a class of optimization algorithms
that take inspiration from nature [5]. By taking inspiration from
biological concepts such as hereditary traits, genotype - phenotype
distinction, mutation, and survival of the fittest, EAs have been used to
solve many challenging problems [3]. Many different encodings can be
used to implement a genotype [18], among them is the real-valued
genotype—a vector in [8]. One often overlooked aspect of real-
valued genotypes is the necessity to restrict the values to task specific
bounds [21]. Restricting the values is fundamental since very few
problems have infinite domains and because restrictions makes the
search space feasible to explore [20]. Other considerations, such as the
proclivity of generating invalid solutions [12], also make genotype
restriction an important topic within the field.
In this paper, we demonstrate the challenges of restricting real-
valued genotypes and the effect these limitations have on the
distribution of values in the genome. We start by introducing the
theoretical problem and how restriction can affect the distribution. We
then show that this problem is not simply a theoretical possibility; by
optimizing benchmark functions we show that results diverge by solely
varying the restriction function. We also look at the literature to gain an
impression of how this problem could affect other EA practitioners.
The contribution of our paper is a better understanding of how
value restriction affects real-valued genotypes in EAs. In addition we
show that this is a topic deserving of more scrutiny by the wider EA
community through a literature overview of both well known EA
frameworks and a review of conference proceedings. By showing both
the theoretical as well as the practical side of this challenge, we hope
that other researchers become aware of the need for better restriction
functions for real-valued genotypes and can utilize the results
presented here in future research.
Fig. 1. By clamping the values in a real-valued genome the values will tend towards
the limits of the bounds. The proposed Bounce-back function reflects values that
exceed the limits back inside the bounds, resulting in a uniform distribution inside
the bounds.
2 Background
When designing a restriction function1 for real values, it is intuitive to
create some form of a clamp method which ensures that a value does
not exceed desired bounds2. Although restricting the values of genomes
is nothing new [19], few contemporary articles take a conscious choice
when it comes to restricting genomes. In [1] the authors compare
different restriction methods on the result of evolving with Differential
Evolution (DE) and showed that the results are impacted by the choice
of restriction function. The paper [21] extended on this work by
understanding which type of inheritance model functioned with
different restriction functions and as a by-product illustrated how these
different restriction functions can alter the search landscape. A larger
review of different components that contribute to the evolutionary
process in DE were undertaken in [20], in this review article the
authors showed that the choice of restriction function can have a
significant impact on the results, a result backed by a very
comprehensive comparison on different benchmark functions.
However, even with these articles, that show significant impact on
results solely on the basis of the restriction function, it is our
impression that few EA practitioners heed these warnings and make
appropriate accommodations in regards to restricting real-valued
genotypes.
To gain insight into the practises of restricting real-valued
genotypes within the wider EA community, we conducted a limited
literature review. We first began with identifying some of the larger
open-source frameworks for implementing EAs to see how value
restriction is implemented.
DEAP [6] is one of the more popular3 Python based frameworks for
implementing EAs. Looking through the source code of this framework,
two functions implement genome value restriction,
‘cxSimulatedBinaryBounded’ and ‘mutPolynomialBounded’. Both of these
functions utilize the Clamped function to limit real-valued genomes.
SFERESv2 [14] is another well-known framework implemented in C++.
It too implements value restriction for real-valued genotypes as can be
seen in the genotype definition in ‘evo_float.hpp’. The restriction
function is implemented in ‘put_in_range’ and is equivalent to the
Clamped function. Both of these frameworks cite the same source when
it comes to their implementation of real-valued genomes and
restriction, NSGA-II, a widely used Multi-Objective Evolutionary
Algorithm [4]. Consulting the source-code4 of this algorithm, it can be
seen that this too implements restriction through the Clamped function.
To evaluate whether or not the observations about evolutionary
frameworks are representative for the community, we conducted a
small review of the main proceedings from GECCO 2019 and the 2019
and 2020 EvoAPPS proceedings. To evaluate if the results of a paper
could be susceptible to the challenges, identified in this paper, we first
identified experiments utilizing real-valued genotypes5, we then tried
to identify if the authors discuss their strategy for value restriction or if
anything could be discerned from the source of the experiments. The
summary of our results can be seen in Table 1.
Table 1. Overview of real-valued genomes and limitation function in previous
conference main proceedings.
GECCO’19 EvoAPPS 2019 EvoAPPS 2020
Total number of papers 173 42 44
Uses real-valued genotype 37 ( ) 11 ( ) 10 ( )
3 Methods
Several possible functions can be devised for limiting individual genes
in real-valued genotypes [1]. The easiest and most straight forward
limitation function is to clamp the value to the given bounds. The
function is defined as follows:
(1)
where v is the value to limit and [min, max] are the bounds to apply for
the value. This function will be used as the baseline to compare against
the proposed Bounce-back method as it can be seen as the default when
limiting real numbers6 and is the only one implemented in the EA
frameworks reviewed. We will refer to this function as Clamped for the
remainder of this text.
The proposed limit function, which we will call Bounce-back
restriction (also known as reflection [1] or the internal reflection
rule [19]), is defined as follows:
(2)
where, again, v is the value to limit and [min, max] are the bounds to
apply for the value. The effect of the Bounce-back function is to redirect
out-of-bounds values by the amount that the value is outside of the
limits. The effect is illustrated in Fig. 1. The function is independent of
both mutation and crossover operators and can be applied in-between
or after variation to ensure the genome is within given bounds.
Compared to other restriction functions, such as wrapping, re-
initialization and re-sampling [1], the Bounce-back function results in
values in the current vicinity of the solution while also having minimum
computational impact. The computational impact can be seen in
contrast to re-sampling which depending on the variational operators
can be computationally more costly to apply. The wrapping restriction
function would share the same properties in regard to distribution as
the Bounce-back function, however, the act of wrapping around can lead
to discontinuous jumps in genotype space which could slow down the
optimization process.
One thing to note about the Bounce-back function, as defined in
Eq. 2, is that it is not guaranteed to result in values within the given
bounds. This can happen when the value or
. This can either be solved by continued
application of the restriction function until the value is within bounds
or limiting the difference in Eq. 2 with the Clamped function. The later
solution was utilized when limiting the distribution in Fig. 2.
To demonstrate the unwanted properties of the Clamped function
and how these properties are not present in the Bounce-back function,
we will first take an empirical approach to restricting real-valued
genotypes with Gaussian perturbation.
4 Empirical Analysis
To understand why using the Clamped function with real-valued
genomes can be problematic we will begin by looking at how the
function affects the Gaussian distribution. In Fig. 2 the result of
applying the two limiting functions to the Gaussian distribution is
shown. The grey area shows the original Gaussian distribution while
the colored areas represent the respective limitation functions as
applied. This example is akin to mutating a value that is on or near the
bound with Gaussian perturbation before restricting the value to be
less than the bound. As can be seen from the figure, the Bounce-back
function result in a distribution that is equivalent with the Gaussian
distribution. Applying the clamping function results in a distribution
that is heavily biased towards the ‘limit’. This shows the problem with
simply clamping a value to the desired bounds, the value will be skewed
towards the bounds because all values above the limit is restricted to
become exactly the limit, as illustrated in Fig. 1.
Fig. 2. The resulting distributions of applying the limit functions to a Gaussian
distribution, in grey, with mean equal to the ‘limit’ value marked with a dashed line.
This is analogous to mutating a value at ‘limit’ with Gaussian mutation and then
restricting the value to be less than the ‘limit’. The Bounce-back distribution follows
the Gaussian distribution while the Clamped distribution is skewed towards the
‘limit’.
value to the range [0, 1]. The perturbation-limitation cycle was run 100 times.
To understand how the input parameters of Algorithm 1 impact the
output, we varied the two parameters, cycles and . We will postulate
that the number of reals, n, will not impact the underlying distribution
and the effect of changing this parameter is to give a better or worse
impression of the underlying distribution. The results of changing the
input parameters are shown in Fig. 4 and Fig. 5, for number of cycles
and varying the standard deviation respectively. Changing the number
of cycles, shown in Fig. 4, does not have an effect on the resulting
distribution which is as expected. Since the probability of the Gaussian
mutation going out- or staying inside the bounds is symmetrical, the
number of times mutation is applied should not affect the resulting
distribution. For the standard deviation, shown in Fig. 5, the results are
slightly different. Here we can see an effect of increasing , which can
be explained as a larger part of the initial uniform distribution having a
probability of going out of bounds. For the Clamped function this results
in more values restricted at the bounds as the standard deviation of the
mutation increases. On the other hand, we can see that the Bounce-back
function is not affected by changes in mutation and continues to be
uniformly distributed.
resulting distribution. The distribution in the figure were generated from 1000
uniform random values and perturbed 100 times.
5 Benchmark Functions
To illustrate the potential impact of the limitation function, we applied
the two limitation functions to a selection of benchmark problems [22].
All functions were optimized with a single objective EA using
tournament selection, Gaussian mutation, and no crossover operator.
The benchmark problems are included to illustrate the challenge of
limiting the genome to a specific range outside of theoretical
considerations, as demonstrated in the previous section.
The following four functions are used as benchmark problems:
(3)
(4)
(5)
(6)
Size - n† 30
Repetitions 50
Tournament 10
P(M|i) 0.05
Mutation - ‘ ’ 0.005 0.05 0.05 0.2
Range
where N is the size of the genome and is the value of gene i in the
genome. We will refer to Eq. (3) as Griewank, Eq. (4) as Rastrigin,
Eq. (5) as Schaffer and Eq. (6) as Schwefel. The genotype was encoded
as a vector of reals with a range of [0, 1], before being transformed into
the range required by each benchmark task. Genotypes are randomly
initiated based on a uniform distribution. To generate data, we first
optimized the parameters for each function, selecting the best mutation
rate , tournament size and probability of applying Gaussian mutation
to gene i - P(M|i). The final parameters used to generate data are shown
in Table 2.
Fig. 6. The mean and confidence interval for the population minimum
aggregated over 50 runs of each benchmark function. Differences are marked at the
bottom of each graph in black, the difference is statistically significant using a
Wilcoxon Rank Sum test over an interval of 10 generations adjusted for the number
of tests using Holm correction.
Figure 6 shows the mean and 95% confidence interval of the
population minimum for each benchmark function. Statistical
significant differences between the two limitation functions are marked
in black at the bottom of each plot and is the result of applying the
Wilcoxon Rank Sum test on intervals of 10 generations. To correct the
p-value for the number of successive tests performed per row, Holm
correction [9] was applied. The result show that only one benchmark
function is sufficiently affected to lead to different results, however,
both Rastrigin and Schaffer show diverging results at some points
before converging. Based on the individual search parameters, shown
in Table 2, we can also observe that larger seems to induce larger
differences.
6 Discussion
From the empirical analysis performed, it can be seen that the Clamped
function skews the distribution of values towards the bounds of the
range, as shown in Fig. 2 and Fig. 3. This property can be challenging
when the genotype is based on real values that are translated into task
applicable ranges during the phenotype conversion. The proposed
Bounce-back function did not exhibit these properties resulting in no
alteration of the underlying value distribution. When experimenting
with different facets of mutation, Fig. 4 and Fig. 5, it was shown that the
magnitude of the standard deviation had the most effect on the
resulting value distribution. This can be understood as expanding the
range of the Gaussian distribution applied at each value. As more values
have the potential of being mutated near the bounds, more values will
be restricted and end up at the extremities of the limit with the
Clamped function. This leads to the observation that using a larger
standard deviation in the mutation operator require more careful
thought to which restriction function to apply.
The results in Fig. 6 shows that the choice of restriction function can
have an effect on benchmark problems, and is not just a theoretical
problem. The effect of the restriction function followed closely the
magnitude of the standard deviation used for mutation, detailed in
Table 2, and illustrates the challenge that it can be difficult a priori to
know the effect of the restriction function on any given task. One thing
to point out about the benchmark tasks is that, except for Schwefel, the
target value for all genes is , which is in the middle of the range,
away from the bounds. As the target value moves closer to the bounds,
we would expect the effect of the restriction function to grow, which
could also explain the larger difference observed with the Schwefel
benchmark function. With real-world tasks where the optimal value is
difficult or impossible to confirm, it is even more challenging to a priori
predict the effect and interactions of the restriction function, making it
paramount that the restriction function has minimal impact on the
optimization process.
The literature review conducted showed that two popular EA
frameworks utilizes the Clamped function for restriction, and that while
many papers published at GECCO’19 and EvoAPPS 2019 and 2020
utilized real-valued genotypes, few papers discussed their use of
restriction functions. A deeper dive into the source of NSGA-II, a
foundational algorithm within the Multi-Objective Evolutionary
Algorithm literature, also revealed the use of the Clamped function
which could affect re-implementations. Although limited, the literature
review does underscore how the findings in this paper could impact the
broader research field. As shown previously, the restriction function is
not guaranteed to lead to significantly different results, however, it is
difficult to predict the effect beforehand.
The results shown in this paper has focused on restricting real-
valued genotypes in EAs. One interesting direction to take this work in
the future is to apply the same Bounce-back technique on particles in
Particle Swarm Optimization (PSO) [11]. Because of the additional
velocity attribute present in PSO the restriction function should take
this into account when limiting so that the particles do not keep moving
towards the bounds. In the same vein, it would be interesting to know if
PSO is sensitive to the challenges presented in this paper, or if the
collective behavior can mitigate the boundary effect shown in this
paper.
7 Conclusion
In this paper we have shown that one of the most used functions for
restricting real-valued genotypes—clamping values to strictly lie
between a minimum and maximum—can bias the genome to the
extremities of the bounds under repeated application of the variation
operators. We therefore suggest a different function which result in a
uniform distribution of values within the genome. The Bounce-back
function is shown both empirically and in practice to lead to a uniform
distribution without bias towards the bounds and requires minimal
intervention from existing Evolutionary Algorithm implementations,
while having a minimal computational impact. We also conducted a
limited literature review which illustrated that other practitioners in
the field of Evolutionary Algorithms could be susceptible to the
complications detailed in this paper. We hope that by illuminating this
problem other researchers will become aware of the requirements for
restricting real-valued genotypes, and can hopefully mitigate it with
minimal effort in future work.
Acknowledgement
This work is partially supported by the Research Council of Norway
through its Centres of Excellence scheme under grant agreements
240862 and 262762.
References
1. Arabas, J., Szczepankiewicz, A., Wroniak, T.: Experimental comparison of methods
to handle boundary constraints in differential evolution. In: Schaefer, R., Cotta, C.,
Kołodziej, J., Rudolph, G. (eds.) PPSN 2010. LNCS, vol. 6239, pp. 411–420. Springer,
Heidelberg (2010). https://doi.org/10.1007/978-3-642-15871-1_42
[Crossref]
2.
Brest, J., Greiner, S., Boskovic, B., Mernik, M., Zumer, V.: Self-adapting control
parameters in differential evolution: a comparative study on numerical
benchmark problems. IEEE Trans. Evol. Comput. 10(6), 646–657 (2006)
[Crossref]
3.
Chiong, R., Weise, T., Michalewicz, Z.: Variants of Evolutionary Algorithms for
Real-World Applications. Springer, Berlin Heidelberg (2012). https://doi.org/10.
1007/978-3-642-23424-8
[Crossref]
4.
Deb, K., Agrawal, S., Pratap, A., Meyarivan, T.: A fast elitist non-dominated sorting
genetic algorithm for multi-objective optimization: NSGA-II. In: Schoenauer, M.,
et al. (eds.) PPSN 2000. LNCS, vol. 1917, pp. 849–858. Springer, Heidelberg
(2000). https://doi.org/10.1007/3-540-45356-3_83
[Crossref]
5.
Floreano, D., Mattiussi, C.: Bio-Inspired Artificial Intelligence: Theories, Methods,
and Technologies. Intelligent Robotics and Autonomous Agents. MIT Press,
Cambridge (2008). http://infoscience.epfl.c h/record/118584
6.
Fortin, F.A., De Rainville, F.M., Gardner, M.A., Parizeau, M., Gagné, C.: DEAP:
evolutionary algorithms made easy. J. Mach. Learn. Res. 13, 2171–2175 (2012)
[MathSciNet]
7.
Glasmachers, T.: Challenges of convex quadratic bi-objective benchmark
problems. In: Proceedings of the Genetic and Evolutionary Computation
Conference. GECCO 2019. pp. 559–567. Association for Computing Machinery,
New York (2019)
8.
Herrera, F., Lozano, M., Verdegay, J.L.: Tackling real-coded genetic algorithms:
operators and tools for behavioural analysis. Artif. Intell. Rev. 12(4), 265–319
(1998)
[Crossref]
9.
Holm, S.: A simple sequentially rejective multiple test procedure. Scand. J. Stat.
6(2), 65–70 (1979)
[MathSciNet][zbMATH]
10.
Iacca, G., Caraffini, F.: Compact optimization algorithms with re-sampled
inheritance. In: Kaufmann, P., Castillo, P.A. (eds.) EvoApplications 2019. LNCS, vol.
11454, pp. 523–534. Springer, Cham (2019). https://doi.org/10.1007/978-3-030-
16692-2_35
[Crossref]
11.
Kennedy, J., Eberhart, R.: Particle swarm optimization. In: Proceedings of ICNN
1995-International Conference on Neural Networks, Perth, WA, Australia, vol. 4,
pp. 1942–1948. IEEE (1995)
12.
Kononova, A.V., Caraffini, F., Bäck, T.: Differential evolution outside the box. arXiv
preprint arXiv:2004.10489 (2020)
13.
Massey Jr., F.J.: The Kolmogorov-Smirnov test for goodness of fit. J. Am. Stat.
Assoc. 46(253), 68–78 (1951)
[Crossref]
14.
Mouret, J.B., Doncieux, S.: SFERESv2: evolvin’ in the multi-core world. In:
Proceedings of Congress on Evolutionary Computation (CEC), Barcelona, Spain,
pp. 4079–4086. IEEE (2010)
15.
Nordmoen, J., Nygaard, T.F., Ellefsen, K.O., Glette, K.: Evolved embodied phase
coordination enables robust quadruped robot locomotion. In: Proceedings of the
Genetic and Evolutionary Computation Conference. GECCO 2019. pp. 133–141.
Association for Computing Machinery, New York (2019)
16.
Nygaard, T.F., Martin, C.P., Torresen, J., Glette, K.: Evolving robots on easy mode:
towards a variable complexity controller for quadrupeds. In: Kaufmann, P.,
Castillo, P.A. (eds.) EvoApplications 2019. LNCS, vol. 11454, pp. 616–632.
Springer, Cham (2019). https://doi.org/10.1007/978-3-030-16692-2_41
[Crossref]
17.
Pontes-Filho, S., et al.: EvoDynamic: a framework for the evolution of generally
represented dynamical systems and its application to criticality. In: Castillo, P.A.,
Jiménez Laredo, J.L., Fernández de Vega, F. (eds.) EvoApplications 2020. LNCS, vol.
12104, pp. 133–148. Springer, Cham (2020). https://doi.org/10.1007/978-3-030-
43722-0_9
[Crossref]
18.
Rothlauf, F.: Representations for Genetic and Evolutionary Algorithms, 2nd edn.,
pp. 9–32. Springer, Heidelberg (2006). https://doi.org/10.1007/3-540-32444-5
[Crossref]
19.
Schaefer, R.: Foundations of Global Genetic Optimization, vol. 74. Springer,
Heidelberg (2007)
[Crossref]
20.
Tanabe, R., Ishibuchi, H.: Review and analysis of three components of the
differential evolution mutation operator in MOEA/D-DE. Soft. Comput. 23(23),
12843–12857 (2019). https://doi.org/10.1007/s00500-019-03842-6
[Crossref]
21.
Wessing, S.: Repair methods for box constraints revisited. In: Esparcia-Alcázar,
A.I. (ed.) EvoApplications 2013. LNCS, vol. 7835, pp. 469–478. Springer,
Heidelberg (2013). https://doi.org/10.1007/978-3-642-37192-9_47
[Crossref]
22.
Yao, X., Liu, Y., Lin, G.: Evolutionary programming made faster. IEEE Trans. Evol.
Comput. 3(2), 82–102 (1999)
[Crossref]
Footnotes
1 Other names include repair operator or simply repairing the genome.
3 Based on citations.
Ragyog a hold…
Updated editions will replace the previous one—the old editions will
be renamed.