This document develops a model to predict the selection pressure resulting from tournament selection in genetic algorithms. Section 3 develops a model for deterministic environments that shows selection pressure is proportional to the product of population fitness standard deviation and the maximal order statistic for the tournament size. Section 4 extends the model to noisy environments by accounting for noise in fitness evaluations. The model is validated using a test problem over a range of tournament sizes and noise levels. The model provides a way to understand and predict how noise affects the selection pressure and convergence of genetic algorithms using tournament selection.
Breast cancer is one of the most common diseases diagnosed among female cancer patients. Early detection of breast cancer is needed to reduce the risk of fatality of this disease as no cure has been found yet for this illness. This research is conducted to improve the Gradient Vector Flow (GVF) Snake Active Contour segmentation technique in mammography segmentation. Segmentation of the mammogram image is done to segment lesions existence using Chan-Vese Active Contour and Localized Active Contour. Besides that, the effectiveness of these both methods are then compared and chosen to be the best method. Digital Database of Screening Mammograms (DDSM) is used for the purpose of screening. First, the images undergo pre-processing process using the Gaussian Low Pass Filter to remove unwanted noise. After that, contrast enhancement applied to the images. Segmentation of mammograms is then conducted by using Chan-Vese Active Contour and Localized Active Contour method. The result shows that Chan-Vese technique outperforms Localized Active Contour with 90% accuracy.
Breast Mass Segmentation Using a Semi-automatic Procedure Based on Fuzzy C-me...TELKOMNIKA JOURNAL
Mammography is the primary modality that helped in the early detection and diagnosis of women
breast diseases. Further, the process of extracting the masses in mammogram represents a challenging task
facing the radiologists, due to problems such as fuzzy or speculated borders, low contrast and the presence of
intensity inhomogeneities. Aims to help the radiologists in the diagnosis of breast cancer, many approaches
have been conducted to automatically segment the masses in mammograms. Towards this aim, in this paper,
we present a new approach for extraction of tumors from region-of-interest (ROI) using the algorithm of Fuzzy
C-Means (FCM) setting two clusters for semi-automated segmentation. The proposed method meant to select
as input data the set of pixels that enable to get the meaningful information required to segment the masses
with high accuracy. This could be accomplished through eliminating unnecessary pixels, which influence on this
process through separating it outside of the input data using an optimal thresho ld given by monitoring the
change of clusters rate during the process of threshold decrementing. The proposed methodology has
successfully segmented the masses, with an average sensitivity of 82.02% and specificity of 98.23%.
Comparative Study of Different Techniques in Speaker Recognition: ReviewIJAEMSJORNAL
The speech is most basic and essential method of communication used by person.On the basis of individual information included in speech signals the speaker is recognized. Speaker recognition (SR) is useful to identify the person who is speaking. In recent years speaker recognition is used for security system. In this paper we have discussed the feature extraction techniques like Mel frequency cepstral coefficient (MFCC), Linear predictive coding (LPC), Dynamic time wrapping (DTW), and for classification Gaussian Mixture Models (GMM), Artificial neural network (ANN)& Support vector machine (SVM).
IJRET : International Journal of Research in Engineering and Technology is an international peer reviewed, online journal published by eSAT Publishing House for the enhancement of research in various disciplines of Engineering and Technology. The aim and scope of the journal is to provide an academic medium and an important reference for the advancement and dissemination of research results that support high-level learning, teaching and research in the fields of Engineering and Technology. We bring together Scientists, Academician, Field Engineers, Scholars and Students of related fields of Engineering and Technology
Breast cancer is one of the most common diseases diagnosed among female cancer patients. Early detection of breast cancer is needed to reduce the risk of fatality of this disease as no cure has been found yet for this illness. This research is conducted to improve the Gradient Vector Flow (GVF) Snake Active Contour segmentation technique in mammography segmentation. Segmentation of the mammogram image is done to segment lesions existence using Chan-Vese Active Contour and Localized Active Contour. Besides that, the effectiveness of these both methods are then compared and chosen to be the best method. Digital Database of Screening Mammograms (DDSM) is used for the purpose of screening. First, the images undergo pre-processing process using the Gaussian Low Pass Filter to remove unwanted noise. After that, contrast enhancement applied to the images. Segmentation of mammograms is then conducted by using Chan-Vese Active Contour and Localized Active Contour method. The result shows that Chan-Vese technique outperforms Localized Active Contour with 90% accuracy.
Breast Mass Segmentation Using a Semi-automatic Procedure Based on Fuzzy C-me...TELKOMNIKA JOURNAL
Mammography is the primary modality that helped in the early detection and diagnosis of women
breast diseases. Further, the process of extracting the masses in mammogram represents a challenging task
facing the radiologists, due to problems such as fuzzy or speculated borders, low contrast and the presence of
intensity inhomogeneities. Aims to help the radiologists in the diagnosis of breast cancer, many approaches
have been conducted to automatically segment the masses in mammograms. Towards this aim, in this paper,
we present a new approach for extraction of tumors from region-of-interest (ROI) using the algorithm of Fuzzy
C-Means (FCM) setting two clusters for semi-automated segmentation. The proposed method meant to select
as input data the set of pixels that enable to get the meaningful information required to segment the masses
with high accuracy. This could be accomplished through eliminating unnecessary pixels, which influence on this
process through separating it outside of the input data using an optimal thresho ld given by monitoring the
change of clusters rate during the process of threshold decrementing. The proposed methodology has
successfully segmented the masses, with an average sensitivity of 82.02% and specificity of 98.23%.
Comparative Study of Different Techniques in Speaker Recognition: ReviewIJAEMSJORNAL
The speech is most basic and essential method of communication used by person.On the basis of individual information included in speech signals the speaker is recognized. Speaker recognition (SR) is useful to identify the person who is speaking. In recent years speaker recognition is used for security system. In this paper we have discussed the feature extraction techniques like Mel frequency cepstral coefficient (MFCC), Linear predictive coding (LPC), Dynamic time wrapping (DTW), and for classification Gaussian Mixture Models (GMM), Artificial neural network (ANN)& Support vector machine (SVM).
IJRET : International Journal of Research in Engineering and Technology is an international peer reviewed, online journal published by eSAT Publishing House for the enhancement of research in various disciplines of Engineering and Technology. The aim and scope of the journal is to provide an academic medium and an important reference for the advancement and dissemination of research results that support high-level learning, teaching and research in the fields of Engineering and Technology. We bring together Scientists, Academician, Field Engineers, Scholars and Students of related fields of Engineering and Technology
International Journal of Engineering Research and Applications (IJERA) is an open access online peer reviewed international journal that publishes research and review articles in the fields of Computer Science, Neural Networks, Electrical Engineering, Software Engineering, Information Technology, Mechanical Engineering, Chemical Engineering, Plastic Engineering, Food Technology, Textile Engineering, Nano Technology & science, Power Electronics, Electronics & Communication Engineering, Computational mathematics, Image processing, Civil Engineering, Structural Engineering, Environmental Engineering, VLSI Testing & Low Power VLSI Design etc.
This paper presents a set of methods that uses a genetic algorithm for automatic test-data generation in
software testing. For several years researchers have proposed several methods for generating test data
which had different drawbacks. In this paper, we have presented various Genetic Algorithm (GA) based test
methods which will be having different parameters to automate the structural-oriented test data generation
on the basis of internal program structure. The factors discovered are used in evaluating the fitness
function of Genetic algorithm for selecting the best possible Test method. These methods take the test
populations as an input and then evaluate the test cases for that program. This integration will help in
improving the overall performance of genetic algorithm in search space exploration and exploitation fields
with better convergence rate.
A Novel Hybrid Voter Using Genetic Algorithm and Performance HistoryWaqas Tariq
Triple Modular Redundancy (TMR) is generally used to increase the reliability of real time systems where three similar modules are used in parallel and the final output is arrived at using voting methods. Numerous majority voting techniques have been proposed in literature however their performances are compromised for some typical set of module output value. Here we propose a new voting scheme for analog systems retaining the advantages of previous reported schemes and reduce the disadvantages associated with them. The scheme utilizes a genetic algorithm and previous performances history of the modules to calculate the final output. The scheme has been simulated using MATLAB and the performance of the voter has been compared with that of fuzzy voter proposed by Shabgahi et al [4]. The performance of the voter proposed here is better than the existing voters.
In this work, the hybrid techniques of genetic algorithm (GA) and particle swarm optimization (PSO) are presented. PSO and GA are two population-based heuristic search methods that can be applied to the channel allocation optimization problem. GAPSO is based on a mixture of particle swarm optimization (PSO) and genetic algorithms (GA). Individuals of a new generation are produced in GAPSO by PSO in addition to crossover and mutation operations as in GA. In order to reduce the number of blocked calls and handoff failures in the mobile network, the Hybrid GAPSO algorithm is used to allocate tasks to resources efficiently. The proposed strategy optimizes the channel allocation using the GAPSO.
In this work, the hybrid techniques of genetic algorithm (GA) and particle swarm optimization (PSO) are
presented. PSO and GA are two population-based heuristic search methods that can be applied to the
channel allocation optimization problem. GAPSO is based on a mixture of particle swarm optimization
(PSO) and genetic algorithms (GA). Individuals of a new generation are produced in GAPSO by PSO in
addition to crossover and mutation operations as in GA. In order to reduce the number of blocked calls
and handoff failures in the mobile network, the Hybrid GAPSO algorithm is used to allocate tasks to
resources efficiently. The proposed strategy optimizes the channel allocation using the GAPSO.
Analysis of Common Supervised Learning Algorithms Through Applicationaciijournal
Supervised learning is a branch of machine learning wherein the machine is equipped with labelled data
which it uses to create sophisticated models that can predict the labels of related unlabelled data. the
literature on the field offers a wide spectrum of algorithms and applications. However, there is limited
research available to compare the algorithms making it difficult for beginners to choose the most efficient
algorithm and tune it for their application.
This research aims to analyse the performance of common supervised learning algorithms when applied to
sample datasets along with the effect of hyper-parameter tuning. for the research, each algorithm is
applied to the datasets and the validation curves (for the hyper-parameters) and learning curves are
analysed to understand the sensitivity and performance of the algorithms. The research can guide new
researchers aiming to apply supervised learning algorithm to better understand, compare and select the
appropriate algorithm for their application. Additionally, they can also tune the hyper-parameters for
improved efficiency and create ensemble of algorithms for enhancing accuracy.
COMPARISON BETWEEN THE GENETIC ALGORITHMS OPTIMIZATION AND PARTICLE SWARM OPT...IAEME Publication
Close range photogrammetry network design is referred to the process of placing a set of
cameras in order to achieve photogrammetric tasks. The main objective of this paper is tried to find
the best location of two/three camera stations. The genetic algorithm optimization and Particle
Swarm Optimization are developed to determine the optimal camera stations for computing the three
dimensional coordinates. In this research, a mathematical model representing the genetic algorithm
optimization and Particle Swarm Optimization for the close range photogrammetry network is
developed. This paper gives also the sequence of the field operations and computational steps for this
task. A test field is included to reinforce the theoretical aspects.
Analysis of Common Supervised Learning Algorithms Through Applicationaciijournal
Supervised learning is a branch of machine learning wherein the machine is equipped with labelled data which it uses to create sophisticated models that can predict the labels of related unlabelled data. the literature on the field offers a wide spectrum of algorithms and applications. however, there is limited research available to compare the algorithms making it difficult for beginners to choose the most efficient algorithm and tune it for their application.
This research aims to analyse the performance of common supervised learning algorithms when applied to sample datasets along with the effect of hyper-parameter tuning. for the research, each algorithm is applied to the datasets and the validation curves (for the hyper-parameters) and learning curves are analysed to understand the sensitivity and performance of the algorithms. the research can guide new researchers aiming to apply supervised learning algorithm to better understand, compare and select the appropriate algorithm for their application. Additionally, they can also tune the hyper-parameters for improved efficiency and create ensemble of algorithms for enhancing accuracy.
ANALYSIS OF COMMON SUPERVISED LEARNING ALGORITHMS THROUGH APPLICATIONaciijournal
Supervised learning is a branch of machine learning wherein the machine is equipped with labelled data
which it uses to create sophisticated models that can predict the labels of related unlabelled data.the
literature on the field offers a wide spectrum of algorithms and applications.however, there is limited
research available to compare the algorithms making it difficult for beginners to choose the most efficient
algorithm and tune it for their application.
This research aims to analyse the performance of common supervised learning algorithms when applied to
sample datasets along with the effect of hyper-parameter tuning.for the research, each algorithm is applied
to the datasets and the validation curves (for the hyper-parameters) and learning curves are analysed to
understand the sensitivity and performance of the algorithms.the research can guide new researchers
aiming to apply supervised learning algorithm to better understand, compare and select the appropriate
algorithm for their application. Additionally, they can also tune the hyper-parameters for improved
efficiency and create ensemble of algorithms for enhancing accuracy.
بعض (وليس الكل) ملخصات الأبحاث الجيدة المنشورة فى بعض المجلات الجيدة وفيها تنوع من الافكار الابحاث الابتكارية التى يخدم فيها علوم الحاسبات فيها - انها تطبيقات حياتية
AUTOMATIC GENERATION AND OPTIMIZATION OF TEST DATA USING HARMONY SEARCH ALGOR...csandit
Software testing is the primary phase, which is performed during software development and it is
carried by a sequence of instructions of test inputs followed by expected output. The Harmony
Search (HS) algorithm is based on the improvisation process of music. In comparison to other
algorithms, the HSA has gain popularity and superiority in the field of evolutionary
computation. When musicians compose the harmony through different possible combinations of
the music, at that time the pitches are stored in the harmony memory and the optimization can
be done by adjusting the input pitches and generate the perfect harmony. The test case
generation process is used to identify test cases with resources and also identifies critical
domain requirements. In this paper, the role of Harmony search meta-heuristic search
technique is analyzed in generating random test data and optimized those test data. Test data
are generated and optimized by applying in a case study i.e. a withdrawal task in Bank ATM
through Harmony search. It is observed that this algorithm generates suitable test cases as well
as test data and gives brief details about the Harmony search method. It is used for test data
generation and optimization
International Journal of Engineering Research and Applications (IJERA) is an open access online peer reviewed international journal that publishes research and review articles in the fields of Computer Science, Neural Networks, Electrical Engineering, Software Engineering, Information Technology, Mechanical Engineering, Chemical Engineering, Plastic Engineering, Food Technology, Textile Engineering, Nano Technology & science, Power Electronics, Electronics & Communication Engineering, Computational mathematics, Image processing, Civil Engineering, Structural Engineering, Environmental Engineering, VLSI Testing & Low Power VLSI Design etc.
This paper presents a set of methods that uses a genetic algorithm for automatic test-data generation in
software testing. For several years researchers have proposed several methods for generating test data
which had different drawbacks. In this paper, we have presented various Genetic Algorithm (GA) based test
methods which will be having different parameters to automate the structural-oriented test data generation
on the basis of internal program structure. The factors discovered are used in evaluating the fitness
function of Genetic algorithm for selecting the best possible Test method. These methods take the test
populations as an input and then evaluate the test cases for that program. This integration will help in
improving the overall performance of genetic algorithm in search space exploration and exploitation fields
with better convergence rate.
A Novel Hybrid Voter Using Genetic Algorithm and Performance HistoryWaqas Tariq
Triple Modular Redundancy (TMR) is generally used to increase the reliability of real time systems where three similar modules are used in parallel and the final output is arrived at using voting methods. Numerous majority voting techniques have been proposed in literature however their performances are compromised for some typical set of module output value. Here we propose a new voting scheme for analog systems retaining the advantages of previous reported schemes and reduce the disadvantages associated with them. The scheme utilizes a genetic algorithm and previous performances history of the modules to calculate the final output. The scheme has been simulated using MATLAB and the performance of the voter has been compared with that of fuzzy voter proposed by Shabgahi et al [4]. The performance of the voter proposed here is better than the existing voters.
In this work, the hybrid techniques of genetic algorithm (GA) and particle swarm optimization (PSO) are presented. PSO and GA are two population-based heuristic search methods that can be applied to the channel allocation optimization problem. GAPSO is based on a mixture of particle swarm optimization (PSO) and genetic algorithms (GA). Individuals of a new generation are produced in GAPSO by PSO in addition to crossover and mutation operations as in GA. In order to reduce the number of blocked calls and handoff failures in the mobile network, the Hybrid GAPSO algorithm is used to allocate tasks to resources efficiently. The proposed strategy optimizes the channel allocation using the GAPSO.
In this work, the hybrid techniques of genetic algorithm (GA) and particle swarm optimization (PSO) are
presented. PSO and GA are two population-based heuristic search methods that can be applied to the
channel allocation optimization problem. GAPSO is based on a mixture of particle swarm optimization
(PSO) and genetic algorithms (GA). Individuals of a new generation are produced in GAPSO by PSO in
addition to crossover and mutation operations as in GA. In order to reduce the number of blocked calls
and handoff failures in the mobile network, the Hybrid GAPSO algorithm is used to allocate tasks to
resources efficiently. The proposed strategy optimizes the channel allocation using the GAPSO.
Analysis of Common Supervised Learning Algorithms Through Applicationaciijournal
Supervised learning is a branch of machine learning wherein the machine is equipped with labelled data
which it uses to create sophisticated models that can predict the labels of related unlabelled data. the
literature on the field offers a wide spectrum of algorithms and applications. However, there is limited
research available to compare the algorithms making it difficult for beginners to choose the most efficient
algorithm and tune it for their application.
This research aims to analyse the performance of common supervised learning algorithms when applied to
sample datasets along with the effect of hyper-parameter tuning. for the research, each algorithm is
applied to the datasets and the validation curves (for the hyper-parameters) and learning curves are
analysed to understand the sensitivity and performance of the algorithms. The research can guide new
researchers aiming to apply supervised learning algorithm to better understand, compare and select the
appropriate algorithm for their application. Additionally, they can also tune the hyper-parameters for
improved efficiency and create ensemble of algorithms for enhancing accuracy.
COMPARISON BETWEEN THE GENETIC ALGORITHMS OPTIMIZATION AND PARTICLE SWARM OPT...IAEME Publication
Close range photogrammetry network design is referred to the process of placing a set of
cameras in order to achieve photogrammetric tasks. The main objective of this paper is tried to find
the best location of two/three camera stations. The genetic algorithm optimization and Particle
Swarm Optimization are developed to determine the optimal camera stations for computing the three
dimensional coordinates. In this research, a mathematical model representing the genetic algorithm
optimization and Particle Swarm Optimization for the close range photogrammetry network is
developed. This paper gives also the sequence of the field operations and computational steps for this
task. A test field is included to reinforce the theoretical aspects.
Analysis of Common Supervised Learning Algorithms Through Applicationaciijournal
Supervised learning is a branch of machine learning wherein the machine is equipped with labelled data which it uses to create sophisticated models that can predict the labels of related unlabelled data. the literature on the field offers a wide spectrum of algorithms and applications. however, there is limited research available to compare the algorithms making it difficult for beginners to choose the most efficient algorithm and tune it for their application.
This research aims to analyse the performance of common supervised learning algorithms when applied to sample datasets along with the effect of hyper-parameter tuning. for the research, each algorithm is applied to the datasets and the validation curves (for the hyper-parameters) and learning curves are analysed to understand the sensitivity and performance of the algorithms. the research can guide new researchers aiming to apply supervised learning algorithm to better understand, compare and select the appropriate algorithm for their application. Additionally, they can also tune the hyper-parameters for improved efficiency and create ensemble of algorithms for enhancing accuracy.
ANALYSIS OF COMMON SUPERVISED LEARNING ALGORITHMS THROUGH APPLICATIONaciijournal
Supervised learning is a branch of machine learning wherein the machine is equipped with labelled data
which it uses to create sophisticated models that can predict the labels of related unlabelled data.the
literature on the field offers a wide spectrum of algorithms and applications.however, there is limited
research available to compare the algorithms making it difficult for beginners to choose the most efficient
algorithm and tune it for their application.
This research aims to analyse the performance of common supervised learning algorithms when applied to
sample datasets along with the effect of hyper-parameter tuning.for the research, each algorithm is applied
to the datasets and the validation curves (for the hyper-parameters) and learning curves are analysed to
understand the sensitivity and performance of the algorithms.the research can guide new researchers
aiming to apply supervised learning algorithm to better understand, compare and select the appropriate
algorithm for their application. Additionally, they can also tune the hyper-parameters for improved
efficiency and create ensemble of algorithms for enhancing accuracy.
بعض (وليس الكل) ملخصات الأبحاث الجيدة المنشورة فى بعض المجلات الجيدة وفيها تنوع من الافكار الابحاث الابتكارية التى يخدم فيها علوم الحاسبات فيها - انها تطبيقات حياتية
AUTOMATIC GENERATION AND OPTIMIZATION OF TEST DATA USING HARMONY SEARCH ALGOR...csandit
Software testing is the primary phase, which is performed during software development and it is
carried by a sequence of instructions of test inputs followed by expected output. The Harmony
Search (HS) algorithm is based on the improvisation process of music. In comparison to other
algorithms, the HSA has gain popularity and superiority in the field of evolutionary
computation. When musicians compose the harmony through different possible combinations of
the music, at that time the pitches are stored in the harmony memory and the optimization can
be done by adjusting the input pitches and generate the perfect harmony. The test case
generation process is used to identify test cases with resources and also identifies critical
domain requirements. In this paper, the role of Harmony search meta-heuristic search
technique is analyzed in generating random test data and optimized those test data. Test data
are generated and optimized by applying in a case study i.e. a withdrawal task in Bank ATM
through Harmony search. It is observed that this algorithm generates suitable test cases as well
as test data and gives brief details about the Harmony search method. It is used for test data
generation and optimization
AUTOMATIC GENERATION AND OPTIMIZATION OF TEST DATA USING HARMONY SEARCH ALGOR...cscpconf
Software testing is the primary phase, which is performed during software development and it is
carried by a sequence of instructions of test inputs followed by expected output. The Harmony
Search (HS) algorithm is based on the improvisation process of music. In comparison to other
algorithms, the HSA has gain popularity and superiority in the field of evolutionary
computation. When musicians compose the harmony through different possible combinations of
the music, at that time the pitches are stored in the harmony memory and the optimization can
be done by adjusting the input pitches and generate the perfect harmony. The test case
generation process is used to identify test cases with resources and also identifies critical
domain requirements. In this paper, the role of Harmony search meta-heuristic search
technique is analyzed in generating random test data and optimized those test data. Test data
are generated and optimized by applying in a case study i.e. a withdrawal task in Bank ATM
through Harmony search. It is observed that this algorithm generates suitable test cases as well
as test data and gives brief details about the Harmony search method. It is used for test data
generation and optimization.
In this research, a hybrid wrapper model is proposed to identify the featured gene subset from the gene expression data. To balance the gap between exploration
and exploitation, a hybrid model with a popular meta-heuristic algorithm named
spider monkey optimizer (SMO) and simulated annealing (SA) is applied. In the proposed model, ReliefF is used as a filter to obtain the relevant gene subset
from dataset by removing the noise and outliers prior to feeding the data to the
wrapper SMO. To enhance the quality of the solution, simulated annealing is
deployed as local search with the SMO in the second phase, which will guide to the detection of the most optimal feature subset. To evaluate the performance of the proposed model, support vector machine (SVM) as a fitness function to recognize the most informative biomarker gene from the cancer datasets along with University of California, Irvine (UCI) datasets. To further evaluate the model, 4 different classifiers (SVM, na¨ıve Bayes (NB), decision tree (DT), and k-nearest neighbors (KNN)) are used. From the experimental results and analysis, it’s noteworthy to accept that the ReliefF-SMO-SA-SVM performs relatively better than its state-of-the-art counterparts. For cancer datasets, our model performs better in terms of accuracy with a maximum of 99.45%.
Improving Prediction Accuracy Results by Using Q-Statistic Algorithm in High ...rahulmonikasharma
Classification problems in high dimensional information with little sort of observations became furthercommon significantly in microarray information. The increasing amount of text data on internet sites affects the agglomerationanalysis. The text agglomeration could also be a positive analysis technique used for partitioning a huge amount of datainto clusters. Hence, the most necessary draw back that affects the text agglomeration technique is that the presenceuninformative and distributed choices in text documents. A broad class of boosting algorithms is known as actingcoordinate-wise gradient descent to attenuate some potential performs of the margins of a data set. This paperproposes a novel analysis live Q-statistic that comes with the soundness of the chosen feature set to boot to theprediction accuracy. Then we've a bent to propose the Booster of associate degree FS algorithm that enhances theworth of the Q-statistic of the algorithm applied.
Dynamic Radius Species Conserving Genetic Algorithm for Test Generation for S...ijseajournal
This paper critically examined nine different software models for modeling / developing multi-agent based
systems. The study revealed that the different models examined have their various advantages /
disadvantages and uniqueness in terms of practical development/deployment of the multi-agent based
information system in question. The agentology model is one methodology that can be easily adopted or
adapted for the development of any multi-agent based software prototype because it was inspired by the
best practices and good ideas contained in other agent oriented methodologies like CoMoMas, MASE,
GAIA, Prometheus, HIM, MASim and Tropos.
How to Create Map Views in the Odoo 17 ERPCeline George
The map views are useful for providing a geographical representation of data. They allow users to visualize and analyze the data in a more intuitive manner.
The French Revolution, which began in 1789, was a period of radical social and political upheaval in France. It marked the decline of absolute monarchies, the rise of secular and democratic republics, and the eventual rise of Napoleon Bonaparte. This revolutionary period is crucial in understanding the transition from feudalism to modernity in Europe.
For more information, visit-www.vavaclasses.com
Read| The latest issue of The Challenger is here! We are thrilled to announce that our school paper has qualified for the NATIONAL SCHOOLS PRESS CONFERENCE (NSPC) 2024. Thank you for your unwavering support and trust. Dive into the stories that made us stand out!
This is a presentation by Dada Robert in a Your Skill Boost masterclass organised by the Excellence Foundation for South Sudan (EFSS) on Saturday, the 25th and Sunday, the 26th of May 2024.
He discussed the concept of quality improvement, emphasizing its applicability to various aspects of life, including personal, project, and program improvements. He defined quality as doing the right thing at the right time in the right way to achieve the best possible results and discussed the concept of the "gap" between what we know and what we do, and how this gap represents the areas we need to improve. He explained the scientific approach to quality improvement, which involves systematic performance analysis, testing and learning, and implementing change ideas. He also highlighted the importance of client focus and a team approach to quality improvement.
Ethnobotany and Ethnopharmacology:
Ethnobotany in herbal drug evaluation,
Impact of Ethnobotany in traditional medicine,
New development in herbals,
Bio-prospecting tools for drug discovery,
Role of Ethnopharmacology in drug evaluation,
Reverse Pharmacology.
We all have good and bad thoughts from time to time and situation to situation. We are bombarded daily with spiraling thoughts(both negative and positive) creating all-consuming feel , making us difficult to manage with associated suffering. Good thoughts are like our Mob Signal (Positive thought) amidst noise(negative thought) in the atmosphere. Negative thoughts like noise outweigh positive thoughts. These thoughts often create unwanted confusion, trouble, stress and frustration in our mind as well as chaos in our physical world. Negative thoughts are also known as “distorted thinking”.
The Art Pastor's Guide to Sabbath | Steve ThomasonSteve Thomason
What is the purpose of the Sabbath Law in the Torah. It is interesting to compare how the context of the law shifts from Exodus to Deuteronomy. Who gets to rest, and why?
The Art Pastor's Guide to Sabbath | Steve Thomason
10.1.1.30.6625 (1)
1. Genetic Algorithms,
Tournament Selection,
and the E ects of Noise
Brad L. Miller
Dept. of Computer Science
University of Illinois
at Urbana-Champaign
David E. Goldberg
Dept. of General Engineering
University of Illinois
at Urbana-Champaign
IlliGAL Report No. 95006
July 1995
Department of General Engineering
University of Illinois at Urbana-Champaign
117 Transportation Building
104 South Mathews Avenue
Urbana, IL 61801
2. Genetic Algorithms, Tournament Selection, and the E ects of Noise
Brad L. Miller
Dept. of Computer Science
University of Illinois
at Urbana-Champaign
bmiller@uiuc.edu
David E. Goldberg
Dept. of General Engineering
University of Illinois
at Urbana-Champaign
deg@uiuc.edu
July 12, 1995
Abstract
Tournament selection is a useful and robust selection mechanism commonly used by genetic algorithms.
The selection pressure of tournament selection directly varies with the tournament size | the more com-
petitors, the higher the resulting selection pressure. This article develops a model, based on order statistics,
that can be used to quantitatively predict the resulting selection pressure of a tournament of a given size.
This model is used to predict the convergence rates of genetic algorithms utilizing tournament selection.
While tournament selection is often used in conjunction with noisy (imperfect) tness functions, little
is understood about how the noise a ects the resulting selection pressure. The model is extended to
quantitatively predict the selection pressure for tournament selection utilizing noisy tness functions. Given
the tournament size and noise level of a noisy tness function, the extended model is used to predict the
resulting selection pressure of tournament selection. The accuracy of the model is veri ed using a simple
test domain, the onemax (bit-counting) domain. The model is shown to accurately predict the convergence
rate of a genetic algorithm using tournament selection in the onemax domain for a wide range of tournament
sizes and noise levels.
The model developed in this paper has a number of immediate practical uses as well as a number of
longer term rami cations. Immediately, the model may be used for determining appropriate ranges of
control parameters, for estimating stopping times to achieve a speci ed level of solution quality, and for
approximating convergence times in important classes of function evaluations that utilize sampling. Longer
term, the approach of this study may be applied to better understand the delaying e ects of function noise in
other selection schemes or to approximate the convergence delays that result from inherently noisy operators
such as selection, crossover, and mutation.
1 Introduction
There are many selection schemes for genetic algorithms (GAs), each with di erent characteristics. An ideal
selection scheme would be simple to code, and e cient for both non-parallel and parallel architectures. Fur-
thermore, a selection scheme should be able to adjust its selection pressure so as to tune its performance for
di erent domains. Tournamentselection is increasingly being used as a GA selection scheme because it satis es
all of the above criteria. It is simple to code and is e cient for both non-parallel and parallel architectures.
Tournamentselection can also adjust the selection pressure to adapt to di erent domain. Tournamentselection
pressure is increased (decreased) by simply increasing (decreasing) the tournament size. All of these factors
have contributed to the increased usage of tournament selection as a selection mechanism for GAs.
Good progress was made some time ago (Goldberg and Deb, 1991) in understanding the convergence rates
of various selection schemes, including tournament selection. Recently, building on work by Muhlenbein and
Schlierkamp-Voosen (1993), this understanding has been re ned to better understand the timing and degree of
convergence more accurately (Thierens and Goldberg, 1994). Despite this progress, this detailed timing and
degree of convergence analysis has not yet been extended to tournaments other than binary (s = 2); nor has
the analysis been applied to domains other than deterministic ones. In this paper, we do these two things.
The purpose of this paper is to develop a model for the selection pressure of tournament selection. This
model, based on order statistics, quantitatively predicts the selection pressure resulting from both di erent
1
3. 2
tournament sizes and noise levels. Given the current population tness mean and variance, the model can
predict the average population tness of the next generation. The model can also be used iteratively to predict
the convergence rate of the GA over time. The predictive model is veri ed, using the onemax domain, under
a range of tournament sizes and noise levels.
Section 2 provides the reader with background information needed to understand this paper, including
tournament selection, noise, and order statistics. Sections 3 and 4 develop the predictive model for tournament
selection. Section 3 develops a predictive model that handles varying tournament sizes for noiseless envir-
onments, and section 4 extends this model for noisy environments. Section 5 assesses the accuracy of the
predictive model, using the onemax domain, for a variety of tournament sizes and noise levels. Application of
the model for other research issues is described in Section 6. Some general conclusions from this research are
presented in section 7.
2 Background
This section gives somebackground informationneeded to understand this paper. The rst subsection describes
selection schemes, selection pressure, and tournamentsection. The second subsection details noise, noisy tness
functions, and approximate tness functions. Lastly, a brief overview of order statistics is given, focusing on
the maximal order statistic for normal distributions.
2.1 Tournament Selection
Genetic algorithms use a selection mechanism to select individuals from the population to insert into a mating
pool. Individuals from the mating pool are used to generate new o spring, with the resulting o spring forming
the basis of the next generation. As the individuals in the mating pool are the ones whose genes are inherited
by the next generation, it is desirable that the mating pool be comprised of good" individuals. A selection
mechanism in GAs is simply a process that favors the selection of better individuals in the population for the
mating pool. The selection pressure is the degree to which the better individuals are favored: the higher the
selection pressure, the morethe better individualsare favored. This selection pressure drives the GAto improve
the population tness over succeeding generations. The convergence rate of a GA is largely determined by the
selection pressure, with higher selection pressures resulting in higher convergence rates. Genetic algorithms
are able to to identify optimal or near-optimal solutions under a wide range of selection pressure (Goldberg,
Deb, and Thierens, 1993). However, if the selection pressure is too low, the convergence rate will be slow, and
the GA will unnecessarily take longer to nd the optimal solution. If the selection pressure is too high, there
is an increased chance of the GA prematurely converging to an incorrect (sub-optimal) solution.
Tournament selection provides selection pressure by holding a tournament among s competitors, with s
being the tournament size. The winner of the tournament is the individual with the highest tness of the
s tournament competitors, and the winner is then inserted into the mating pool. The mating pool, being
comprised of tournament winners, has a higher average tness than the average population tness. This
tness di erence provides the selection pressure, which drives the GA to improve the tness of each succeeding
generation. Increased selection pressure can be provided by simply increasing the tournament size s, as the
winner froma larger tournamentwill, on average, have a higher tness than the winner of a smallertournament.
2.2 Noise and Noisy Fitness Functions
The noise inherent in noisy tness functions causes the tournament selection process to also be noisy. We
assume that a noisy tness function returns a tness score for an individual equal to the sum of the real tness
of the individual plus some noise. In this paper, we assume that the noise is normallydistributed and unbiased
(mean of zero). This assumption is true or approximately true in many noisy domains, and allows the e ects
of noise to be more easily modeled.
There are many factors that may necessitate the use of noisy tness functions. In some domains, there may
be no known tness function that can accurately assess an individual's tness, so an approximate(noisy) tness
function must be used. Noisy information can also negatively a ect the tness evaluation. Noisy information
can comefroma varietyof sources, including noisysensor input, noisy data, knowledge uncertainty, and human
error. To improve run-time performance, some GAs will utilize fast, but noisier, tness functions instead of
4. 3
more accurate, but slower, tness functions that may also be available. Sampling tness functions are a good
example of this phenomena, as a tness function that uses sampling to assess an individual's tness can use
smaller sample sizes to increase run-time speed, at the expense of decreased accuracy of the tness evaluation.
2.3 Order Statistics
This paper uses order statistics to further our understanding of tournament selection, and this section brie y
reviews them. For a detailed description of order statistics, the reader should see David (1981).
If a random sample of size n is arranged in ascending order of magnitude and then written as
x1:n x2:n ::: xn:n;
we can let the random variable Xi:n represent the distribution of the corresponding xi:n over the space of all
possible samples of size n. The variable Xi:n is called the ith order statistic. The eld of order statistics deals
with the properties and applications of these random variables.
Of particular interest is the maximalorder statistic Xn:n, which represents the distribution of the maximum
member of a sample of size n. This is directly analogous to tournament selection, where the competitor with
the maximum tness is selected as the tournament winner.
The probability density function pi:n(x) of the ith order statistic, Xi:n, gives the probability that the ith
highest individual from a sample of size n will have a value of x. The value of pi:n(x) is calculated by
pi:n(x) = n n? 1
i ?1 P(x)i?1(1?P(x))n?i;
where P(x)represents the cumulativedistributionfunction ofx (the probabilitythat fX xg). The probability
that a single combination will have i? 1 individuals less than or equal to x and n? i individuals greater than
x is given by the product P(x)i?1(1 ?P(x))n?i. However, there are many possible sample combinations that
will yield the desired distribution of having i?1 individuals less than x and n?i individuals greater or equal
to x. For n individuals, there are n slots that the ith greatest individual could occupy. For each of these slots,
there are
?n?1
i?1 di erent ways of arranging the i ? 1 individuals that are less than or equal to x among the
n? 1 remaining slots.
The expected value (mean) ui:n of an order statistic Xi:n can thus be determined by
ui:n =
+1Z
?1
x pi:n(x)dx;
= n n? 1
i ?1
+1Z
?1
x P(x)i?1(1? P(x))n?idP(x):
For the maximal order statistic (i = n), the mean un:n simpli es to
un:n = n
+1Z
?1
x P(x)n?1dP(x):
In this paper we are particularly interested in the normal distribution N( ; 2), where and 2 are the
mean and variance, respectively, of the normaldistribution. For the standard normaldistribution N(0;1), P(x)
is the cumulative distribution function for the unit normal (x), and thus dP(x) is (x)dx = 1p2 e?x2
2 dx. The
expected value (mean) of the maximal order statistic for the standard normal distribution is thus
un:n = n
+1Z
?1
x (x)n?1 (x)dx: (1)
5. 4
For samples of size n = f2;3;4;5g, Equation 1 for un:n can be solved exactly in terms of elementary func-
tions. Table 1 gives the values for the mean of the maximalorder statistic for n = f2;3;4;5g(see Balakrishnan
and Cohen (1991) for derivations). For larger values of n, the means of the order statistics for the standard
normal distribution have been tabulated extensively (Harter, 1961). The variances and covariances of the
standard normal distribution order statistics can also be calculated, and are tabulated for n 20 in Teichroew
(1956), and for n 50 in Tietjen, Kahaner, and Beckman (1977).
n n:n Values of n:n
2 1p 0.5642
3 3
2p 0.8463
4 6p tan?1(
p
2) 1.0294
5 5
4p + 15
2 p sin?1(1
3) 1.1630
Table 1: Expected Value of Maximal Order Statistic for Standard Normal Distribution.
3 Tournament Selection in Deterministic Environments
This section develops a predictive model for the selection pressure resulting from a tournament of size s in a
deterministic (noiseless) environment. In a noiseless environment,the tness function can accurately assess the
true tness of an individual. We show that for a population whose tness is normally distributed, the resulting
tournament selection pressure is proportional to the product of the standard deviation of the population tness
and the maximal order statistic s:s.
In a deterministic environment, the tness function returns the true tness value of an individual. The
population's tness values, after crossover and mutation, are assumed to be normally distributed over the
population. Although tournament selection by itself will generate a skewed (non-normal) distribution, the
crossover and mutationoperations `remix'the population,which forces the distributionto become morenormal.
This normalizing e ect of crossover and mutation allows the assumption of normally distributed population
tness to be reasonable for a wide variety of domains.
Let the population tness in generation t be normally distributed N( F;t; 2
F;t). The probability that an
individual with tness f will win a tournament of s individuals randomly picked from the population is given
by
p(f = max(f1 :::fs)) = s P(F < f)s?1p(f);
which represents the probability of an individual with tness f occurring along with s ? 1 individuals having
lower tness scores. There are s di erent ways of arranging the s?1 losers" and the winner." The expected
value of the tournament winner F;t+1 from a tournament of size s can thus be calculated by
F;t+1 = E f = max(f1 :::fs)];
=
+1Z
?1
f p(f = max(f1 :::fs))df;
= s
+1Z
?1
f P(f)s?1p(f)df:
However, for a normally distributed population N( F;t; 2F;t), P(f) = (f? F;t
F;t
), and
p(f) = dP(f)
df = 1
F;t
(f ? F;t
F;t
):
Thus
F;t+1 = s
F;t
+1Z
?1
f (f ? F;t
F;t
)s?1 (f ? F;t
F;t
)df:
6. 5
Substituting z = f? F;t
F;t
gives
F;t+1 = s
+1Z
?1
( F;tz + F;t) (z)s?1 (z)dz;
= F;t s
+1Z
?1
(z)s?1 (z)dz + F;t (s
+1Z
?1
z (z)s?1 (z)dz);
= F;t (z)s]+1
?1 + F;t s:s;
= F;t + F;t s:s: (2)
In Equation 2, s:s is the e ective selection pressure for a tournament of size s, and can be directly obtained
from Table 1. From Equation 2, it is also apparent that the change in tness between generations is given by:
F;t = F;t+1 ? F;t;
= F;t s:s: (3)
For binary tournaments (s = 2), this matches the result obtained in (Thierens and Goldberg, 1994), where
the expected increase in the average population tness for tournaments of size s = 2 was derived in a di erent
manner using the di erence between normal distributions. Their result, using the notation in this paper, was
F;t+1 = F;t + F;t
1p . This matches the result obtained using Equation 2 with a tournament size of s = 2,
as 2:2 = 1p from Table 1. Note that the order statistic model derived in this paper is generalizable to all
tournament sizes, and is not limited to s = 2.
Equation 3 shows that for tournaments of size s, the expected average population tness increase is directly
proportionalto s:s, the expected value ofthe maximalorder statisticofsize s. Table1 gives s:s, demonstrating
that increasing the tournament size will cause successively smaller increases in the expected average population
tness.
4 Tournament Selection in Noisy Environments
This section extends the model developed above to accurately predict the selection pressure in the presence
of noise. With noisy tness functions, there is a chance that the winner of a tournament might not be the
individual with the highest true tness. This section concentrates on quantifying the reduction in tournament
selection pressure due to noisy tness functions.
The model derivation in this section has three major steps. First, the relationship between an individual's
noisy tness and true tness values is determined, so that the expected true tness value of an individual
can be calculated from the noisy tness evaluation. Next, the relationship is extended to handle subsets of
individuals, so that the true tness average of a subset of the population can be estimated from the average
noisy tness value of the subset. Lastly, we use the model derived in the previous section to estimate the
average noisy tness value of a particular subset of the population - the subset consisting of noisy tournament
winners. This average noisy tness value is then plugged into the formula found in the second step to estimate
the average true tness of the winners of noisy tournaments. The selection pressure, based on the expected true
tness value of the tournament winners, is thus be determined. The result is a predictive model for tournament
selection that can handle varying noise and tournament sizes.
In a noisy environment, the noisy tness f0 of an individual is given by f0 = f +noise, where f is the
real tness of the individual, and noise is the noise inherent in the tness function evaluation. As in the last
section, the real tness of the population F is assumed to be normally distributed N( F;t; 2F;t). This section
further assumes that the noise is unbiased and normally distributed N(0; 2N). This facilitates modeling the
e ects of the noise, and is a reasonable assumption for many domains. Using these assumptions, along with
the additive property of normal distributions, gives that F0 is normally distributed N( F;t; 2F;t + 2N ).
Although the real tness value for an individual is unknown, the expected value can be determined from
the individual's noisy tness value, which is generated by a noisy tness function evaluation. As both the true
tness and the noisy tness are normally distributed, the bivariate normal distribution can be used to obtain
7. 6
the expected true tness value of F for a given noisy tness value f0 of F0. For normal random variables X
and Y, the bivariate normal distribution states that the expected value of Y for a speci c value x of X is
E(Y jx) = Y + XY
Y
X
(x? X );
where XY is the correlation coe cient for X and Y. The correlation coe cient XY can be calculated by
XY = XY
X Y
, where XY is the covariance of X and Y. The covariance between F and F0 is simply 2
F
, thus
E(Fjf0) = F +
2F
F F 0
F
F0
(f0 ? F 0 );
= F +
2F
2
F0
(f0 ? F 0 );
= F +
2
F
2F + 2N
(f0 ? F0 ): (4)
As the above formulais linear, the expected value of F for any subset R of the population can be calculated
using equation 4, with f0 set to the noisy tness mean R of the subset. Of course, the subset we are interested
in is the noisy tournament winners. The expected mean of the noisy tournament winners of tournament size s
can be derived using the same derivation as for the deterministic case:
F 0;t+1 = F 0;t + F 0 ;t s:s;
= F 0;t +
q
2F;t
+ 2N s:s:
Setting f0 to F 0;t+1 in equation 4 produces the expected true tness value of the tournament winners:
E(Ft+1j F0;t+1) = F;t+1;
= F;t +
2F;t
2F;t + 2N
( F 0
;t +
q
2F;t + 2N s:s ? F 0
;t);
= F;t +
2F;t
p
2F;t
+ 2N
s:s: (5)
As expected, equation 5 reduces to equation 1, the formula for the deterministic (noiseless) case, when the
noise variance 2N equals zero. Equation 5 is signi cant in that it predicts the convergence rate of a genetic
algorithm using tournament selection for any tournament size or noise level.
5 Validation of Model
This section tests the accuracy of the predictive model, equation 5, using a sample domain. The domain
used is the bit-counting, or onemax, domain, which works well for analysis as the variance can be determined
from the average population tness. This section uses equation 5 to predict the performance under a range
of tournament sizes and noise levels. Experiments are then run that show that the predictive model is very
accurate in determining the tournament selection pressure for di erent tournament sizes and noise levels.
5.1 Onemax Domain
The domain of interest is the onemax, which is also referred to as the bit-counting problem. The real tness
of an individual in this domain is simply the number of one bits in the chromosome. The optimal solution is
the chromosome consisting of all one bits. This population tness in this domain is binomially distributed,
and the mean and variance of the the population tness can therefore be calculated using binomialdistribution
properties. The population mean tness at generation t is given by f(t) = F;t = l p(t), where l is the
chromosomelength, andp(t)isthe percentage ofcorrect allelesinthe population. The varianceofthe population
at time t is simply 2F;t = l p(t)(1? p(t)).
8. 7
The experiments in this paper all use the following GA con guration parameters. The chromosome length
is l = 100, crossover is performed using the uniform crossover operator, and no mutation is used so as to
better isolate the selection pressure e ects. The proportion of correct alleles in the initial random population is
assumed to be p(0) = 0:5. The population size is adjusted for di erent noise levels, as described in Goldberg,
Deb, and Clark (1992). For the onemax domain, the population sizing equation reduces to N = 8( 2
f + 2
n),
with the population variance 2
f conservatively set to 2F;0 = l p(0)(1 ? p(0)) = 25. The noise variance 2n
is user speci ed for each experiment. For experiments with a non-zero noise variance 2n, a random number
generated from the noisy distribution N(0; 2n) is added to the real tness score for each individual to produce
a noisy tness score. For the noisy experiments, tournament selection is based solely on the noisy tness values
of the individuals.
5.2 Predictive Model for the Onemax Domain
This section adapts Equation 5 to determine the convergence rate of the percentage of correct alleles over time
for the onemax domain. From equation 5 the tness increase between two generation is given by:
f(t + 1)? f(t) = F;t+1 ? F;t;
=
2
F;t
p
2F;t + 2N
s:s:
For the onemax domain, F;t = l p(t) and 2F;t
= l p(t)(1? p(t)). Thus
p(t+1)? p(t) = 1
l (f(t + 1)? f(t));
= s:s
l
2
F;t
p
2F;t + 2N
;
= s:s
p(t)(1? p(t))p
l p(t)(1? p(t))+ 2N
:
Approximating the above di erence equation with a di erential equation yields
dp
dt = s:s
p(t)(1? p(t))p
l p(t)(1? p(t))+ 2N
: (6)
Although equation 6 is integrable, it does not reduce to convenient form in the general case; however, it
can be easily solved numerically for p(t), and for the noiseless case ( 2N
= 0) p(t) can be determined exactly.
Subsection 5.3 will deal with solving equation 6 for t(p). Given the initial percentage of correct alleles is
p(0) = 0:5, equation 6 can be solved exactly for p(t) in the noiseless case to yield:
p(t) = 0:5(1+ sin( s:s tp
l
)): (7)
Equations 6 and 7 together make up the predictive model for the onemax domain. Equation 6 is numerically
solved to predict p(t) for noisy domains, while equation 7 is directly used to obtain p(t) for noiseless domains.
In both equations, s:s determines the selection pressure for a tournament of size s. For noisy domains, the
term
2F;t
p 2F;t
+ 2N
causes the predicted convergence rate to decrease as the noise is increased. In the next section
we assess the accuracy of these equations for a variety of tournament sizes and noise levels.
5.3 Convergence Time for the Onemax Domain
While equation 6 is not directly solvable for p(t), it can be solved for t as a function of p:
t(p) = 1
s:s
"
p
l arctan
p
l(2p? 1)
2
p
2N +l p(1? p)
!
+ N log p
p? 1
?l ? 2 2N
+ l p?2 N
p
2N + l p(1? p)
2 2N +l p +2 N
p
2N + l p(1?p)
!
+c
#
:
(8)
9. 8
For binary alleles, at time t = 0 we can assume that half of the alleles are initially correct p = 0:5. Using
this to solve for c in equation 8 gives that c = 0. For the case where p = 1 (convergence), s = 2, and N = 0,
equation 8 reduces to t(1:0) =
p
l 2, which agrees with convergence time found in Thierens and Goldberg
(1994) for binary tournament selection. Of course, equation 8 is more general than the convergence equation
in Thierens and Goldberg (1994), as it can handle tournaments of di erent sizes and noise levels.
We are particularly interested in the time tc it takes for all alleles to converge (p = 1). For the deterministic
case, equation 8 reduces to
tc =
p
l
2 s:s
: (9)
A useful approximation of the convergence time for the noisy cases is
tc = 1
s:s
"
p
l arctan
p
l
2 N
!
+ N log (l ?1)4 2N
l + 4 2N
#
: (10)
This approximation is obtained by setting p = 1 in equation 8, except for the p
p?1 fraction in the log term.
For the p
p?1 term, we relax the convergence criterion by setting p = l?1
l , indicating that 100(l?1
l ) percent of
the bits in the population have correctly converged. Setting p = l?1
l in the p
p?1 term yields (1?l). Equation 10
is used to develop approximations for domains characterized by small, medium, and large amounts of noise.
For domains characterized by small levels of noise ( N 0), equation 10 can be approximated by
tc = 1
s:s
"
p
l arctan
p
l
2 N
!
+2 N log(2 N )
#
; (11)
as the log term is insigni cant for very small levels of noise.
A medium noise level is de ned as having the tness function noise variance N approximately equal the
initialpopulation tness variance f level, which for the onemax domainis at most
p
l=4 =
pl
2 . Approximating
equation 10 using N
pl
2 yields
tc = 1
s:s
"
p
l arctan
p
l
2 N
!
+ 2 N log(
p
2 N )
#
: (12)
For large amounts of noise ( N 1), equation 10 can be approximated by
tc = 1
s:s
l
2 N
+ N log(l ?1) ; (13)
as for small angles, arctan .
The approximationsequations for convergence with small, medium,and large amounts of noise can be used
to quickly estimate the convergence time for a GA. These are useful for the GA designer trying to gauge the
delaying e ects of noise on population convergence.
5.4 Experimental Results
In this section we assess the accuracy of our predictive model. We compare the predicted performance versus
the actual performance obtained from GA runs for varying noise levels and tournament sizes to validate our
predictive model.
To assess the accuracy of the predictive model, GA runs were made at ve di erent noise variance levels
2n = f0; 1
2
2
f; 2
f;2 2
f;4 2
fg. At each noise variance level, GA runs were made with tournaments sizes of
s = f2;3;4;5g. For each combination of noise variance and tournament size, 10 GA runs were made, and the
results were averaged. The experimental results were then compared to the results predicted from Equations 6
(noisy) and 7 (deterministic).
A sample plot is shown in gure 1a. The noise variance, given on the top line, is 50. The dashed lines in the
plot represent the predicted performance obtained using Equations 6 and 7 for tournamentsizes s = f2;3;4;5g.
The solid lines display the GA performance, averaged over 10 runs, for the given noise variance and tournament
10. 9
sizes. The dashed lines, from left to right, correspond to the predicted performance with tournament sizes
of ve (highest selection pressure), four, three, and two (lowest selection pressure). Similarly, the solid lines
correspond to the performance with tournament sizes, from left to right, of ve (highest selection pressure),
four, three, and two (lowest selection pressure).
20 40 60 80 100
Generations
50
60
70
80
90
100
%Convergence
Noise Variance: 50.
Tournament Sizes: 5,4,3,2
Predictive Model
Experimental Avg.
(a) One Crossover
20 40 60 80 100
Generations
50
60
70
80
90
100
%Convergence
Noise Variance: 50.
Tournament Sizes: 5,4,3,2
Predictive Model
Experimental Avg.
(b) Two Crossovers
Figure 1: E ects of Multiple Crossovers
Figure 1 compares the e ects of performing one crossover versus two when the noise variance is equal
to the tness variance. While the predictive model slightly overestimates the performance of GAs using
one crossover ( gure 1a), it accurately estimates the performance of the experiments using two crossovers
( gure 1b). This is a result of crossover decreasing the correlation between alleles (Thierens and Goldberg,
1994), and the tendency of crossover to `normalize' the population tness distributions, making our model
assumption of a normal population distribution more accurate. As done in Thierens and Goldberg (1994), we
perform two crossovers per generation in our experiments, so after the usual procedure of tournament selection
and recombination, we randomly shu e the population and again recombine the population using crossover.
This has the bene cial e ects of reducing the correlation between alleles (Thierens and Goldberg, 1994), and
`normalizing'the population tness distribution. However, for domainscharacterized by high allele interaction,
this could reduce the overall performance of the GA, as there would be a higher disruption rate for the evolving
building blocks.
Figure 2 summarizes our experimental results. Figure 2a plots the deterministic case, where the noise
variance is zero, for a variety of tournament sizes. Figure 2b plots the experiments where the noise variance
is 25, for a variety of tournament sizes. Figure 2c takes a di erent view, in that it plots the results of using a
xed tournament size (s = 2) for a variety of noise levels. These gures demonstrate that our model is very
accurate for predicting GA performance in the onemax domain for a wide range of tournament sizes and noise
levels.
5.5 Discussion of Results
This subsection discusses the general accuracy of the model, and how selection pressure a ects the accuracy
of the model. The accuracy of the approximation convergence equations is also discussed.
The model proved to be very accurate over a wide range of noise levels and tournament sizes. On many
experiments,the predicted andexperimentalresults were practicallyidentical. However, the modelismarginally
less accurate in domains characterized by high selection pressure. This is primarily a result of the high
selection pressure causing the tournament selection process to generate a non-normal (skewed) distribution,
which violates the model assumption of a normally distributed population. For our experiments, high selection
pressure was caused by high tournament sizes (s = 5). Interestingly, higher levels of noise actually reduces the
tournament selection pressure, making the model more accurate. For our experiments, the highest selection
pressure was for s = 5 and N = 0. The results of this experiment are shown in gure 2a, in the upper left
11. 10
20 40 60 80 100
Generations
50
60
70
80
90
100
%Convergence Noise Variance: 0
Tournament Sizes: 5,4,3,2
Predictive Model
Experimental Avg.
(a) Variance: 0
50 100 150 200 250
Generations
50
60
70
80
90
100
%Convergence
Noise Variance: 25.
Tournament Sizes: 5,4,3,2
Predictive Model
Experimental Avg.
(b) Variance: 25
20 40 60 80 100
Generations
50
60
70
80
90
100
%Convergence
Tournament Size: 2.
Noise Var: 0,12.5,25,50,100
Predictive Model
Experimental Avg.
(c) Tournament Size: 2
Figure 2: Onemax Experiments for Various Tournament Sizes and Noise Levels
two lines (predicted and experimental results) of the plot. This demonstrates that even with high selection
pressure, the model is still fairly accurate.
Crossover has a `normalizing' e ect on the population tness. When the selection pressure is high, the
matingpoolselected is non-normal(skewed). Performing multiplecrossovers per generation has a `normalizing'
e ect on the tness distribution of the new o spring generated through crossover from the mating pool, which
in turn makes the model more accurate (see gure 1). The experiments performed in this paper used two
crossovers per generation so as facilitate comparison of results with those obtained in Thierens and Goldberg
(1994). For very high selection pressures, the model accuracy can be increased by simplyincreasing the number
of crossovers performed per generation. However, this increased accuracy does not come free, for performing
multiple crossovers per generation in domains characterized by high allele interaction will retard building block
growth. This slows the GA convergence rate, and would thus make the model less accurate. For the onemax
domain, which has no allele interaction, multiple crossovers only increases the model accuracy.
These experiments also veri ed the accuracy of our approximation equations for the convergence time.
Table 2 presents the average convergence time of the experiments for a variety of noise levels when the tourna-
ment size is two, as well as the convergence times predicted by the exact models (equation 8 and 9), and the
small, medium and large noise approximations (equations 11, 12, and 13). For the GA experiments, conver-
gence was de ned as the rst generation in which the experimental average of the average population tness
was over 99% converged. The exact model for the noisy case (equation 8) also used p = :99 convergence for
the noisy cases, as it evaluates to in nity if p = 1:0, while the deterministic model (equation 9) was used for
the noiseless case. The approximation equations all estimate the time until absolute convergence p = 1:0.
12. 11
The table shows that the exact convergence equations (equations 8 and 9) predict the experimental results
quite well. The small approximation equation turns out to be fairly accurate at N =
pl
2 = 12:5, but it was
designed for smaller amounts of noise. For lower noise levels, it should be more accurate than the medium
approximation. The medium approximation, being designed around a noise level of N =
pl
2 = 25 is fairly
accurate for all noise levels up to 100. At the high noise level of 100, the large approximation model is the
most accurate approximation. These results indicate that the approximation equations do very well as a quick
estimate of the convergence time for GAs utilizing tournament selection.
Noise Exper. Exact t Approximate tc
2N tc (p = :99) Small Med. Large
0 28.0 28.8 28.8 28.8 NA
12.5 40.0 39.9 42.8 38.3 55.6
25.0 49.0 49.9 56.5 50.2 60.4
50.0 64.0 65.4 79.9 70.9 72.4
100.0 87.0 88.4 118.2 105.5 93.3
Table 2: Convergence Times for s = 2.
6 Future Research
This section describes a number of important areas for future research:
Applying newfound understanding of noise for other selection schemes
Modeling other GA operators by utilizing the noise component of the model to account for their e ect
on convergence
Testing the model in more complex domains
Using the model to answer basic performance questions for a GA
Applying the model to help tune GA con guration parameters
Determining appropriate sample sizes for tness function employing sampling, so as to maximize per-
formance within a given environment
This research is important in that it has furthered our understanding of noise and its delaying e ect on
convergence. The model has proved accurate at predicting the convergence rate of a GA utilizing tournament
selection for a variety of noise levels and tournament sizes in the onemax domain. The approach taken in
section 4, where the deterministic tournament selection model is extended to handle noise, shows promise for
adapting other deterministic selection models to handle noise.
Within this paper, the noise component was considered to be produced by noise present in the tness
functions. However, there is no reason why the noise input for the developed model can not include noise
introduced from other GA components. The noise input indirectly indicates the degradation of the mating pool
tness as compared to the mating pool selected with no noise present. The noise input can thus be used to
account for other degradations of matingpool tness fromother GA mechanismsbesides noisy tness functions.
Other GA mechanisms that also introduce noise could be included in the noise component, such as di erent
mutation mechanisms and mutation rates, and tournament selection used with or without replacement. This
would increase the model's predictive accuracy for a wider range of GA con gurations.
As discussed in Subsection 5.5, the use of order statistics has proved very accurate in predicting tournament
selection pressure for the onemax domain. We would like to extend our model to handle other domains that
have di erent characteristics than the onemax domain. The onemax domain is characterized by equal allele
weighting (each bit is equally important),and no allele interaction (no higher order building blocks). We would
like to extend our model to handle more complex domains, including domino-like" domains, where the alleles
have unequal weighting, and domains characterized by high allele correlation. However, the current model is
13. 12
still of use for these types of domain,in that it can provide an lower bound on the convergence rate. Extending
our model to directly handle more complex domains will increase the accuracy of the predicted convergence
rates.
One of our model's strengths is that it predicts the distribution of the population tness over successive
generations. Order statistics can be used not only to predict both the increase in tness between generations,
but also to predict the population tness variance in the next generation. As the population tness mean
and variance can be accurately modeled over time using order statistics, our model can be used to answer
questions relating to population tness distribution. The model could thus be applied to answer PAC-like
performance questions like What is the probability that a solution of quality Y will be present at generation
X," or At what generation are we Z percent con dent that a solution of at least quality Y will be present
in the population." Answers to these questions could be used to determine how long a GA should run before
a solution of acceptable quality is likely to be be produced. This would allow for a GA designer to set GA
stopping criteria that achieves a desired solution quality.
This model should be very helpful in determining appropriate settings for many GA parameters. While
GAs with generic parameter settings are good at nding good solutions in a reasonable amount of time,
their performance can be improved by tuning the parameter settings for a speci c domain. As discussed
above, our model can be used to determining appropriate stopping criteria for the GA. It can also be used
to design a GA that has a desired convergence rate for a given domain, by applying the model to determine
the appropriate tournament size for achieving a speci ed convergence rate. It could even be used to design
custom" tournament that achieves a given selection pressure (i.e., a tournament where the best 2 out of 5
competitors are selected for the mating pool).
In some domains, a GA designer is faced with a range of possible tness functions, all with di erent noise
and run-time performance characteristics. The model can be applied to help select a tness function that
achieves an acceptable solution in an acceptable amount of runtime for a given domain. Some tness functions
are based on sampling, with the sampling tness function's noise and run-time performance characteristics
directly controlled by the sample size. Our model, in conjunction with sampling theory being used to predict
the noise from a given sample size, should be able to determine the appropriate sample size needed for the GA
to achieve a given convergence rate.
7 Conclusions
Tournament selection is an important selection mechanism for GAs. It is simple to code, easy to implement
on non-parallel or parallel architectures, robust in the presence of noise, and has adjustable selection pressure.
This paper has developed a model that works under a wide range of noise levels and tournament sizes to
accurately predict the convergence rate of a GA utilizing tournament selection. The model has been veri ed
using the onemax domain, and shown to be accurate for predicting the convergence rate under a wide range of
noise levels and tournament sizes.
The paper has discussed a number of immediate practical uses of the model. It can be used to correctly
set various GA control parameters for a given domain, including tournament sizes and mutation rates. The
model can determine appropriate stopping criteria for achieving a desired solution quality. The model can be
used to answer basic performance questions, such as What is the probability that a solution of quality Y will
be present at generation X?" In addition, the model can also be used to determine appropriate sample sizes
for the class of tness functions that employ sampling so as to maximize GA performance.
This research has several long term rami cations. Through the study of one selection scheme, tournament
selection, this paper has opened the door for understanding noiseand its delayinge ect for allselection schemes.
The approach used in this study mayalso be appliedto predict the convergence delays resulting frominherently
noisy operators such as selection, crossover, and mutation.
8 Acknowledgments
This work was supported under NASA Grant No. NGT 9-4. This e ort was also sponsored by the Air Force
O ce of Scienti c Research, Air Force Materiel Command, USAF, under grant numbers F4960-94-1-0103 and
14. 13
F49620-95-1-0338. The U.S. Government is authorized to reproduce and distribute reprints for Governmental
purposes notwithstanding any copyright notation thereon.
The views and conclusions contained herein are those of the authors and should not be interpreted as
necessarily representing the o cial policies or endorsements, either expressed or implied, of the Air Force
O ce of Scienti c Research or the U.S. Government.
References
Balakrishnan, N. & Cohen, A. C. (1991). Order statistics and inference: estimation methods (pp. 51{54).
Boston: Harcourt Brace Jovanovich.
David, H. A. (1981). Order statistics (2nd ed.). New York: John Wiley & Sons, Inc.
Goldberg, D. E. & Deb, K. (1991). A comparative analysis of selection schemes used in genetic algorithms.
Foundations of Genetic Algorithms, 1, 69{93. (Also TCGA Report 90007).
Goldberg, D. E., Deb, K., & Clark, J. H. (1992). Genetic algorithms, noise, and the sizing of populations.
Complex Systems, 6, 333{362.
Goldberg, D. E., Deb, K., & Thierens, D. (1993). Toward a better understanding of mixing in genetic
algorithms. Journal of the Society of Instrument and Control Engineers, 32(1), 10{16.
Harter, H. L. (1961). Expected values of normal order statistics. Biometrika, 48, 151{165.
Muhlenbein, H. & Schlierkamp-Voosen, D. (1993). Predictive models for the breeder genetic algorithm: I.
Continuous parameter optimization. Evolutionary Computation, 1(1), 25{49.
Teichroew, D. (1956). Tables of expected values of order statistics and products of order statistics for samples
of size twenty and less from the normal distribution. Annals of Mathematical Statistics, 27, 410{426.
Thierens, D. & Goldberg, D. (1994). Convergence models of genetic algorithmselection schemes. In Davidor,
Y., Schwefel, H.-P.,& Manner, R. (Eds.), Parallel Problem Solving from Nature- PPSN III (pp. 119{129).
Berlin: Springer-Verlag.
Tietjen, G. L., Kahaner, D. K., & Beckman, R. J. (1977). Variances and covariances of the normal order
statistics for sample sizes 2 to 50. Selected Tables in Mathematical Statistics, 5, 1{73.