Design Efficient Technologies for Context Image Analysis in Dialog

HCI Using Self-Configuring Novelty Search Genetic Algorithm

Evgeny Sopov and Ilia Ivanov

Institute of Computer Science and Telecommunications, Siberian State Aerospace University, Krasnoyarsk, Russia

Keywords: Image Analysis, Classification, Multiobjective Genetic Algorithm, Novelty Search.

Abstract: The efficiency of HCI systems can be sufficiently improved by the analysis of additional contextual

information about the human user and interaction process. The processing of visual context provides HCI

with such information as user identiﬁcation, age, gender, emotion recognition and others. In this work, an

approach to adaptive model building for image classification is presented. The novelty search based upon

the multi-objective genetic algorithm is used to stochastically design a variety of independent technologies,

which provide different image analysis strategies. Finally, the ensemble based decision is built adaptively

for the given image analysis problem.

1 INTRODUCTION

Modern information and communication

technologies are highly interactive. There is a

problem of designing effective HMIs (Human

machine interface, HMI), which include a contextual

analysis of the user's activities, dialog speech

interfaces, etc. Today, virtually all modern

communication and computer technologies have the

capability of capturing the user's image (using the

camera on computers and smartphones, ATMs,

vending machines services, vehicles and many

others). The user image analysis can provide HMI

with additional information to improve the

efficiency of human-computer interaction. In

particular, image analysis is able to obtain such

information as user anthropological data (gender,

age, ethnicity, etc.). This work is focused on the

particular analysis problem of estimating a person’s

age based on a photo.

In general, existing image analysis technologies

are realized using two image-processing stages: the

pre-processing stage to highlight different image

patterns and the recognition stage to solve the given

analysis problem. Both stages become a great

challenge, as the user image in real applications is

not usually well suited to automated computer

processing. Moreover, there is no a priori effective

technique for highlighting image patterns. Finally,

the practical usage of HMI leads to various image

analysis problems, which can dynamically change

over time. The design of intelligent adaptive

methods for generating and tuning effective image

analysis technologies for each particular application

seems to be a good idea.

The genetic algorithms (GA) are a well-known

and highly effective approach to solving many

search and adaptation problems (Goldberg, 1989),

and are also widely used to configure intelligent

information technologies (Shabalov et al., 2011). In

this work multi-objective GA is used to design

simultaneously pre-processing and recognition

technologies. The self-configuring version of GA is

used to eliminate the need for choosing the correct

GA structure and tuning of parameters.

In the field of statistics and machine learning,

ensemble methods are used to improve decision

making. The collective solution of multiple learning

algorithms provides better predictive performance

than could be obtained from any of the constituent

learning algorithms (Opitz and Maclin, 1999). It is

clear that similar algorithms will lead to identical

failures, so we need to support a diversity of

technologies of an ensemble. In this work the

novelty search approach is used to generate various

image analysis technologies, which represent

various analysis strategies. The advantage of the

novelty search is that the obtained solutions may be

new, previously unknown and even non-obvious

(Lehman and Stanley, 2008).

832

Sopov E. and Ivanov I..

Design Efﬁcient Technologies for Context Image Analysis in Dialog HCI Using Self-Conﬁguring Novelty Search Genetic Algorithm.

DOI: 10.5220/0005147108320839

In Proceedings of the 11th International Conference on Informatics in Control, Automation and Robotics (ASAAHMI-2014), pages 832-839

ISBN: 978-989-758-040-6

 2014 SCITEPRESS (Science and Technology Publications, Lda.)

2 RELATED WORK

The idea to obtain additional information from a

user image to adapt the dialog interaction in HMI is

not new, but has still not been sufficiently

investigated. The main reason is the problem of

discovering a generalization for the methods applied

to recognize human features. In particular, there are

the following problems in the field of estimating a

person’s age (Geng et al., 2007; Sithu et al., 2013):

 Different people grow old in different ways,

thus there is no universal methodology of age

recognition;

 Image pre-processing, feature extraction and

principal component analysis involve

operations like human features detection

(eyes, nose, mouth, eyebrows, etc.), which are

complex problems themselves;

 There is no universal technique for

highlighting image patterns for age

recognition. There exist many local

approaches based on using different filters for

image processing;

 The image of a person is obtained each time

under non-standard, non-ideal conditions

(lighting, shooting angle, excess objects and

artefacts and so on).

The way to overcome these problems is to design

technology that efficiently combines the feature

extraction from the image and the recognition

system.

The list of image pre-processing filters is finite

and known. The problem is to select the proper ones

and to construct an efficient combination of filters.

The given extracted features are used in the

recognition system. There are many approaches to

complete learning and classification: statistical and

mathematical methods, machine learning, heuristics

and others. Many researchers in the field of image

analysis consider the Convolutional Neural

Networks (CNNs) as the one of the most powerful

approaches to image recognition problem solving

(Bishop, 1995; Simard et al., 2003).The CNN is

based on the idea of Deep Learning that uses many

levels of information representation for complex

relationships in data modelling. This leads to the

following advantages:

 Local perception. Only a part of the given

image is processed by each neuron.

 Weights selection. A small number of weights

is used to model a large number of links in

data.

 Generalization. CNN ignores the image noise

and finds image invariants.

As is known, the structure and parameter

adjustment of a neural network is a complex

problem, which, nevertheless, can be solved using

an adaptive stochastic procedure like a genetic

algorithm. This work is focused on a special type of

search algorithm called a novelty search algorithm.

The novelty search approach instead of searching

for a ﬁnal objective, rewards instances with

functionality signiﬁcantly different from what has

been discovered before. Thus, instead of a traditional

objective function, evolution employs a novelty

metric. The novelty metric measures how different

an individual is from other individuals, so creating a

constant pressure to do something new. The novelty

search was successfully applied to many search

problems in various domains (Lehman and Stanley,

2011).

Applying the idea of a novelty search to the

problem of CNN design, it is possible to obtain

many different recognition technologies, even those

that are not a priori obvious for an expert in a field

of CNN.

3 INTELLEGENT TECHNOLOGY

OF IMAGE ANALYSIS

The process of image analysis for each person starts

with some standard pre-processing steps like face

detection, canvas rotation/orientation, image levels

correction, transformation to grayscale (if

necessary), etc. We will assume that these operations

are well studied, universal and not in the scope of

this work.

The image after the pre-processing steps are

applied is represented in some resolution M by M

pixels (also can be viewed as a square matrix of size

M). It is obvious that a higher resolution leads to

possibly better feature extraction, but a more

complex recognition process. A set of initial images

forms a training set for CNN. Unfortunately the

initial image contains no human features in an

explicit form, thus we need to apply some operations

to extract features to be appended to the training set.

Usually special image filters are used to

highlight different image patterns. Most of them are

based on different methods to find the edges of

objects using contrast, light gradient, blurring, etc.

After a filter is applied, an additional processed

image of M by M resolution is obtained, which is

appended to the training set.

DesignEfficientTechnologiesforContextImageAnalysisinDialogHCIUsingSelf-ConfiguringNoveltySearchGenetic

Algorithm

833

As we use the GA to design the image analysis

technology, we will discuss the GA solution

representation. The binary GA is used. The

chromosome of an individual consists of two parts:

filter combination and CNN structure

representations.

It is assumed that the set of filters consists of K

items. We use the first K position of chromosome to

encode filter combination. The i

position is equal to

one if the i

filter is applied, and zero otherwise.

The number of maps and the size of kernels

define the CNN structure. Usually two types of CNN

are used: 3 layer (the sequence of Convolutional (C-

layer), Sub-sampling (S-layer) and Fully-connected

(F-layer)) and 5 layer (C-, S-, C-, S- and F-layer). As

the full-connected layer is defined by the preceding

convolution and sub-sampling, it is necessary to

define only the parameters of the convolutional and

the sub-sampling layer. Thus, the second part of the

chromosome should encode the following set of

integer parameters: (1) in a case of 3 layer CNN, (2)

in a case of 5 layer CNN.

〈

,



,



,…



,,



,



,…



〉

(1)





,



,…







,



,



,…











,





,…







,



,





,…









(2)

where CL is the number of maps of C-layer, SL is

the number of maps of S-layer, c

and s

is the size of

kernel of i

map.

The quality measure for GA individuals (or

fitness function) is directly based on a standard

metric of accuracy (the percentage of correct

predicted samples). In real-world dialog HMI

problems, it is not necessary to estimate the exact

age, but to classify some age interval. The exact age

estimation is a great challenge even for a human and

we usually divide people into groups like child,

teenager, adult under 30 y.o., 30-45 y.o., etc. Thus,

the used age intervals reduce the number of classes

in the recognition problem, give more reliable

results of prediction and make the decision-making

process closer to the human one.

The fitness evaluation implies two steps. First,

the binary solution is decoded to obtain the

structures of combinations of image filters and

CNN, i.e. an image analysis technology is designed.

Second, the learning process is performed with the

training set and the accuracy of the age estimation is

evaluated.

The weight adjustment algorithm for CNN is not

a focus of this work; any efficient technique can be

applied (e.g. genetic algorithms, evolutionary

strategies, particle swarm optimization, gravitational

search algorithm, etc.).

4 NOVELTY SEARCH GENETIC

ALGORITHM

The proposed approach suggests designing a variety

of image analysis technologies with different

properties to be applied in an ensemble. As

previously mentioned, one of the ways to improve a

collective decision in an ensemble is to include

technologies as diverse as possible.

The GA search based upon only the recognition

accuracy criteria may lead to very similar or even

the same optimal solutions. Thus to support the

solution diversity additional search criteria should be

used. In this work, we apply the idea of the novelty

search.

To implement the search for a novelty, the search

algorithm should focus on the functionality diversity

of solutions (in terms of the given problem) instead

of the objectives as in the standard approach.

The functionality of the image analysis

technologies designed in this work is based on two

things:

 The combinations of image filters applied

produce different feature sets extracted from

the original image of the person. As a result,

the same CNN is trained with different

training data.

 Different structures of CNN implement

dissimilar convolutions and sub-samplings, so

CNNs process the same image in many ways

using various links in data and unique

generalization.

Using the solution representation proposed in

section 3, we can control the functionality diversity

directly by preserving population diversity. Thus, we

can formalize the following additional search criteria

for GA (3):

















,









→

(3)

where X

is the j

-nearest neighbor of X with respect

to the distance metric dist

, dist

is the Hamming

distance for binary representation.

The number of T nearest neighbour is chosen

experimentally and its value is much less than the

binary solution dimension (chromosome length)

because of the neighbourhood structure in the binary

space.

The novelty criteria



(X) can be viewed as the

sparseness at point X. It is clear that if the average

distance to the nearest neighbours of the given point

is large then point X is in a sparse area and it is a

ICINCO2014-11thInternationalConferenceonInformaticsinControl,AutomationandRobotics

834

novel solution. Otherwise, if the average distance is

small, the point is in a dense region.

There exist some techniques for preserving

population diversity (e.g. crowding, sharing,

niching, etc.). They are widely used for finding

multiply optima (Mahfoud, 1995), but seem to be

unsuitable for a novelty search as their fitness

function is based on objectives.

We state the problem of the design of image

analysis technologies as an optimization problem

with two criteria: recognition accuracy and novelty

search. Thus, we have to use a multi-objective

technique instead of the standard GA. In this work,

we use an efficient approach called the self-

configuring coevolutionary multi-objective genetic

algorithm (SelfCOMOGA). The SelfCOMOGA was

introduced in (Ivanov and Sopov, 2013a) and its

effectiveness was investigated in (Ivanov and Sopov,

2013b).

One of the ways to develop self-configuring

search algorithms is the co-evolutionary approach

(Ficici, 2004).

Most of the works related to the co-evolutionary

multi-objective GA use the cooperative co-evolution

scheme, which implies the problem decomposition

of variables, objectives, population functionality,

etc. The fitness evaluation is based on the fitness

values of many individuals. Such a scheme shows

good performance, but requires the fine-tuning of

algorithm parameters, decomposition and

cooperation.

In the case of competitive co-evolution,

algorithms (populations) evolve independently over

some time. Then their performance is estimated.

Computational resources are redistributed to the

most effective algorithms. In addition, random

migrations of the best solutions are presented. The

competitive co-evolution scheme eliminates the

necessity to define an appropriate algorithm for the

problem as the choice of the best algorithm is

performed automatically during the run (Goh, 2007).

It is clear that applying the co-evolution scheme

provides the self-configuration for the given multi-

objective problem. Using different multi-objective

search approaches with fundamentally different

properties, the co-evolution algorithm is able to

improve the use of their individual advantages and

minimize the effects of the disadvantages.

The general SelfCOMOGA scheme is as follows:

Step 1. To define a set of multi-objective

algorithms included in the co-evolution.

Step 2. To perform algorithms run over some

time (called the adaptation period).

Step 3. To estimate the performance for each

algorithm over the adaptation period.

Step 4. To redistribute the computational

resources (population) and perform a new adaptation

period (go to the step 2).

We will discuss the SelfCOMOGA steps in

detail.

The first step defines the search strategies in the

form of the algorithms included. There are at least

three ways to form the algorithm set:

 A pre-defined set of algorithms with special

performance features, which can be applied to

multi-objective problem solving:

 All appropriate multi-objective algorithms;

 A random selection of algorithms from the set

of all possible algorithms combinations.

The first way requires the involvement of a priori

knowledge of the problem and appropriate

algorithms. So it is not applicable to complex real-

world problems and contradicts the idea of the self-

configuring GA. The second approach is very

expensive because of the number of all algorithm

combinations (tens and hundreds). As shown in

(Ivanov and Sopov, 2013b), the third strategy

provides acceptable efficiency on average, and it is

not necessary to involve additional information

about the algorithms and their properties. This

option provides the best concept of the self-

configuring GA.

The adaptation period is a parameter of the co-

evolution algorithm. Numerical experiments show

that the parameter value is individual for each

problem and depends on the performance criteria of

the algorithm. Moreover, the value depends on the

limitation of the computational resource (total

number of fitness evaluations). It should be noted

that the algorithm is not sensitive to the parameter

on average with sufficient computational resources.

The key point of any co-evolutionary scheme is

the performance evaluation of the individual

algorithm. Since the performance is estimated using

the Pareto concept, a direct comparison of

algorithms is not possible, so well-known

approaches cannot be used. In various studies the

following criteria are proposed: the distance

(closeness) of the obtained solutions set to the true

Pareto set and the uniformity of the set of obtained

solutions. It is obvious that in real-world problems

the first criterion is not applicable because the true

Pareto set is a priori unknown. In this work we use

the following criteria combined into two groups.

The first group includes the static criteria (the

performance is measured over the current adaptation

period):

DesignEfficientTechnologiesforContextImageAnalysisinDialogHCIUsingSelf-ConfiguringNoveltySearchGenetic

Algorithm

835

 Criterion 1 is the percentage of non-dominated

solutions. The pool of all algorithm solutions

is created and non-dominated sorting is

performed. Finally, the number of non-

dominated solutions corresponding to each

algorithm is counted.

 Criterion 2 is the uniformity (dispersion) of

non-dominated solutions. The average

variance of distances between individuals is

computed using the coordinates (for the Pareto

set) or criteria (for Pareto front).

The second group contains the dynamic criteria

(the performance is measured in a comparison with

previous adaptation periods):

 Criterion 3 is the improvement of non-

dominated solutions. The solutions of the

previous and current adaptation period are

compared. The improvement is completed if

the current solutions dominate the previous

ones, even its number has decreased.

The last stage is the redistribution of resources.

New population sizes are defined for each algorithm.

All algorithms give to the “winner” algorithm a

certain percentage of their population size. Each

algorithm has a minimum guaranteed resource that is

not distributed.

In many co-evolutionary schemes each new

adaptation period begins with the same starting

points (“the best displaces the worst”). This

approach is not suitable for multi-objective problems

due to the requirement of the diversity of Pareto

solutions. In this work, we apply the probability

sampling scheme “substitution by rank selection”.

Solutions with a high rank after non-dominated

sorting have a greater probability value to be

transferred to the best algorithms. With decreasing

rank, the probability of selection decreases linearly,

but the chance to be selected remains for all

solutions.

The co-evolutionary algorithm stop criteria are

similar to the standard GA: fitness evaluation

restriction, a number of generations with no

improvement (stagnation), etc.

5 EXPERIMENTAL RESULTS

The proposed approach was investigated with a

widely used dataset of human facial images labelled

with their true age. The dataset is collected by

researchers of the University of Michigan and

provided by the University of Texas in Dallas (the

full data set can be found in the Internet in

http://agingmind.utdallas.edu/facedb). The goal is to

design an image analysis technology for automatic

age recognition. Either this problem may be viewed

as a classification problem (with the finite number of

age groups) or as a regression problem (the age of a

person is a real number).

The data contains more than 1000 different facial

images of humans of different genders and ethnic

groups, and the people also have different emotions.

The original resolution was 640 by 180 pixels. A

face detector is applied to every photo to select the

informative region of the image. Colour images are

converted to grayscale. After the feature extraction

filters are applied, the resolutions of all images are

reduced to 100 by 100 pixels for the CNN training.

We use the following encoding of image analysis

technologies designed using the Self-configuring

Novelty Search Genetic Algorithm (4):









,…,



,



,…,





(4)

where X is a GA binary solution, 



∈



0,1



. l is the

number of bits encoding the filters applied, m is the

number of bits encoding the CNN structure. The

total length of the chromosome is n=l+m.

We start with the list of filters which are used to

highlight image patterns. The list, based on studies

in (Zhou and Chellappa, 2004), contains the

following: Logarithmic, Canny, Zerocross, Prewitt,

Sobel, Roberts and LBP.

Thus the value of l is equal to 7 (total number of

all possible filters combinations is 2



 128). The

position is equal to one, if the i

filter is applied,

and zero otherwise. The solution with all l positions

equal to zero corresponds to the case that no filters

are applied; the training set contains the initial image

only.

The example of an arbitrary initial image taken

from the data set is presented in Figure 1. The

examples of the same image after face detection and

filter are applied are in Figure 2 and Figure 3.

Figure 1: The example of an initial image.

ICINCO2014-11thInternationalConferenceonInformaticsinControl,AutomationandRobotics

836

Figure 2: The example of canny filter applied.

Figure 3: The example of zerocross filter applied.

A new image, obtained after a filter is applied

forms new data in the training set. There the smallest

size of the training set is in the case of no filter being

used, the largest one is in the case of all filters being

applied.

The next step is that we encode the structure of

the CNN that are trained with the given extracted

features. We use the CNN with 5 layer of the

following order: C-, S-, C-, S- and F-layer.

Starting with the l+1 position of the solution

chromosome, we use the following encoding:

 5 bits to encode the number of maps of the

first convolutional layer. The max number is



32.

 6 bits 32 times (summary 6∙32  192 bits)

to encode the kernel size for each map. If the

number of maps after decoding is less than 32,

the corresponding parts of size encoding are

not used in the CNN building.

 5 bits 32 times (160 bits) to encode the

subsampling kernels for each convolutional

layer map.

 4 bits to encode the number of maps of the

second convolutional layer (maximum 16

maps for each map after subsampling).

 2 bits 16 times (32 bits) – convolutional

kernels sizes.

 2 bits 16 times (32 bits) – subsampling kernels

sizes.

As result, the total length of the CNN encoding

part is 425 bits. The total chromosome length is 432

bits.

The initial parameters for the SelfCOMOGA are:

 The list of multi-objective search algorithms

contains Vector Evaluated Genetic Algorithm

(VEGA), Fonseca and Fleming’s Multi-

objective Genetic Algorithm (FFGA), Niched

Pareto Genetic Algorithm (NPGA), Non-

dominated Sorting Genetic Algorithm

(NSGAII) and Strength Pareto Evolutionary

Algorithm (SPEA) (Tan et al., 2001; Zhoua et

al., 2011). All algorithms use the self-

configured parameters as performed in

(Semenkin and Semenkina, 2012).

 Total population size is 400 individuals.

 The initialization is random.

 The crossover operators: one-point, two-point

and uniform.

 The probability of mutation operator: low,

mean and high.

 The maximum size of the Pareto set (for

SPEA): 50 individuals.

 The niching parameter (niche radius) (for

FFGA and NPGA): 1, 3 and 5.

 The size of a comparison set (for NPGA): 3, 5

and 7.

The different image analysis technologies

discovered by GA are applied in the form of

ensemble with averaging of individual predictions.

To estimate the designed CNN performance we

have divided the given age range into the following

7 intervals:

 18-23 years old,

 23-27 years old,

 25-36 years old,

 32-49 years old,

 46-60 years old,

 58-70 years old,

 >70 years old.

The training set is randomly divided into two

samples (80% and 20%) to perform the training and

test. The error is a standard accuracy.

The result of the SelfCOMOGA run is a set of

designed technologies, which are effective with

respect to two criteria: accuracy and novelty

(diversity). The result Pareto front set is shown in

Figure 4.

As we can see, the SelfCOMOGA provides a

rather uniform distribution over the Pareto set, so we

obtain a good variance of image analysis

technologies. The Pareto set contains the age

estimation technologies that both have good

accuracy, but their structures are similar and there is

poor accuracy with a unique structure. As we

previously assumed, the novel solutions provide a

DesignEfficientTechnologiesforContextImageAnalysisinDialogHCIUsingSelf-ConfiguringNoveltySearchGenetic

Algorithm

837

completely different analysis strategy and may

improve total efficiency when applied in an

ensemble with more accurate solutions.

Figure 4: Pareto front set distribution.

Also we have compared the proposed approach

with the following algorithms:

 7 algorithms are the combination of a filter

and manually (experimentally) tuned CNN.

 The averaging ensemble mentioned above 7

algorithms.

 7 algorithms + ensemble, where the CNN is

substituted with Support Vector Machines

(SVM). SVMs are widely and successfully

used for many complex classification and

recognition problems in machine learning

(Byun and Lee, 2002).

The results of numerical experiments are shown

in Table 1. The algorithms are grouped: self-

configuring, CNN-based and SVM-based.

Table 1: Age estimation accuracy.

Algorithm

Training

sample

Test

sample

Self-configuring

technology design

85.37 81.03

Logarithmic+CNN 72.45 68.26

Canny+CNN 70.32 66.88

Zerocross+CNN 72.11 65.63

Prewitt+CNN 60.87 56.17

Sobel+CNN 58.90 54.46

Roberts+CNN 58.42 55.04

LBP+CNN 59.93 58.13

7 CNN ensemble 78.87 71.23

Logarithmic+SVM 64.15 63.74

Canny+ SVM 71.97 64.13

Zerocross+ SVM 70.00 66.21

Prewitt+ SVM 56.37 54.02

Sobel+ SVM 56.57 52.11

Roberts+ SVM 52.39 51.34

LBP+ SVM 53.54 50.87

7 SVM ensemble 70.55 68. 43

As we see, the ensemble approach for both CNN and

SVM shows slightly better performance than the

best algorithm in a group (CNN or SVM). The self-

configuring approach shows significantly better

results. We have applied the ANOVA (Wilcoxon-

Mann-Whitney) tests to confirm the statistical

significance of the difference in results.

6 CONCLUSIONS

The self-configuring novelty search genetic

algorithm for the design of image analysis

technologies is proposed in this work. We have

demonstrated the efficiency of the approach with

respect to the particular problem of age recognition.

The approach can be implemented for any other

image analysis problem raised in the field of dialog

HCI as we do not apply specific information about

the problem.

The computational costs of the approach are

higher than of the standard algorithms, but the

results are performed automatically in an adaptive

way for a certain problem.

So we can recommend the proposed approach for

the design of image analysis when the problem is

poorly-studied, no a priori information on the

problem has been presented or there are no

appropriate experts in the field of computer image

analysis.

In further investigations, we intend to examine

the approach with other particular problems and to

integrate the approach with dialog HCI systems to

provide the design of technologies for certain HCI

requirements.

ACKNOWLEDGEMENTS

Research is performed with the financial support of

the Ministry of Education and Science of the

Russian Federation within the federal R&D

programme (project RFMEFI57414X0037).

The authors express their gratitude to Mr. Ashley

Whitfield for his efforts to improve the text of this

article.

REFERENCES

Bishop, C.M. (1995). Neural Networks for Pattern

Recognition. Oxford University Press.

ICINCO2014-11thInternationalConferenceonInformaticsinControl,AutomationandRobotics

838

Byun, H., Lee, S.W. (2002). Applications of Support

Vector Machines for Pattern Recognition: A Survey.

In: Proceedings of the First International Workshop

on Pattern Recognition with Support Vector Machines.

pp. 213-236.

Ficici, S.G. (2004). Solution Concepts in Coevolutionary

Algorithms. A Doctor of Philosophy Dissertation /

S.G. Ficici. - Brandeis University.

Geng, X., Zhou, Z.-H., Smith-Miles, K. (2007). Automatic

age estimation based on facial aging patterns. In: IEEE

Trans. On PAMI 29(12), pp. 2234-2240.

Goh, C.K. (2007). A competitive-cooperation

coevolutionary paradigm for multi-objective

optimization. In: 22nd IEEE International Symposium

on Intelligent Control, Singapore. pp. 255-260.

Goldberg, D. E. (1989). Genetic Algorithms in Search,

Optimization and Machine Learning, Addison-

Wesley Publishing Co., Inc., Reading, Mass.

Ivanov, I., Sopov, E. (2013a). Self-configured genetic

algorithm for multi-objective decision making support.

Vestnik SibSAU (scientific journal of Siberian State

Aerospace University) 1(47). pp. 30-35. - In Russian,

abstract in English.

Ivanov, I., Sopov, E. (2013b). Investigation of the self-

configured coevolutionary algorithm for complex

multi-objective optimization problem solving. Control

systems and information technologies, 1.1 (51). pp.

141-145. - In Russian, abstract in English.

Lehman, J., Stanley, K.O. (2008). Improving Evolvability

through Novelty Search and Self-Adaptation. In:

Proceedings of the 2011 IEEE Congress on

Evolutionary Computation (CEC 2011). IEEE, pp.

2693-2700.

Lehman, J., Stanley, K.O. (2011). Abandoning objectives:

Evolution through the search for novelty alone. Evol.

Comp. 19(2), pp. 189-223.

Mahfoud, S.W. (1995). A Comparison of Parallel and

Sequential Niching Methods. In: Proceedings of the

Sixth International Conference on Genetic Algorithms.

pp. 136-143.

Opitz, D., Maclin, R. (1999). Popular ensemble methods:

An empirical study, Journal of Artificial Intelligence

Research 11, pp. 169–198.

Semenkin, E.S., Semenkina, M.E. (2012). Self-

configuring Genetic Algorithm with Modified

Uniform Crossover Operator. Advances in Swarm

Intelligence. Lecture Notes in Computer Science 7331.

– Springer-Verlag, Berlin Heidelberg. pp. 414-421.

Shabalov, A., Semenkin, E., Galushin, P. (2011).

Automatized Design Application of Intelligent

Information Technologies for Data Mining Problems.

In: Joint IEEE Conference “The 7th International

Conference on Natural Computation & The 8th

International Conference on Fuzzy Systems and

Knowledge Discovery", FSKD’2011, pp. 2596-2599.

Simard, P.Y., Steinkraus, D., Platt, J.C. (2003). Best

Practices for Convolutional Neural Networks Applied

to Visual Document Analysis. In: International

Conference on Document Analysis and Recognition

(ICDAR), IEEE Computer Society.

Sithu, U., Shyama, D., Imthiyas, M.P. (2013). Human age

prediction and classification using facial image.

International Journal on Computer Science and

Engineering, 5(05).

Tan, K.C., Lee, T.H., Khor, E.F. (2001). Evolutionary

Algorithms for Multi-Objective Optimization:

Performance Assessments and Comparisons. In:

Proceedings of the 2001 IEEE Congress on

Evolutionary Computation Seoul, Korea. pp. 979-986.

Zhou, S.K., Chellappa, R. (2004). Illuminating Light

Field: Image-Based Face Recognition Across

Illuminations and Poses. In: Proc. Sixth IEEE Int’l

Conf. Automatic Face and Gesture Recognition, pp.

229-234.

Zhoua, A., Qub B.Y., Li, H., et al. (2011). Multiobjective

evolutionary algorithms: A survey of the state of the

art. Swarm and Evolutionary Computation, 1 (2011).

pp. 32-49.

DesignEfficientTechnologiesforContextImageAnalysisinDialogHCIUsingSelf-ConfiguringNoveltySearchGenetic

Algorithm

839