Advances on cognitive automation at LGI2P / Ecole des Mines d'Als - - PDF document

Advances on cognitive automation at LGI2P / Ecole des Mines d'Alès Doctoral research snapshot 2014-2015

July 2015 Research report RR_15_01

Foreword

This research report sums up the results of the 2015 PhD seminar of the LGI2P lab of the Alès National Superior School of Mines. This annual day-long meeting gathers presentations of the latest research results of LGI2P PhD students. This year edition of the seminar took place on June 25th. PhD students presented their work for the past academic year. All presentations were be followed by extensive time for very constructive questions from the audience. The aggregation of abstracts of these works constitute the present research report and gives a precise snapshot of the research on cognitive automation led in the lab this year. I would like to thank all lab members, among which all PhD students and their supervisors, for their professionalism and enthusiasm in helping me prepare this seminar. I would also like to thank all the researchers that came to listen to presentations and ask questions, thus contributing to their thesis defense training. I wish you all an inspiring reading and hope to see you all again for next year’s 2016 edition!! Christelle URTADO

JOURNEE DE PRESENTATION DES TRAVAUX DES DOCTORANTS DU LGI2P JEUDI 25 JUIN 2015 SALLE DE CONFERENCE – SITE DE NIMES DE L’ECOLE NATIONALE SUPERIEURE DES MINES D’ALES PROGRAMME DE LA JOURNEE Début du séminaire 9h15 1A Imad BENKHALED 15 minutes / 5 minutes 9h30 1A Pierre-Antoine JEAN 15 minutes / 5 minutes 9h50 1A Massissilia MEDJKOUNE 15 minutes / 5 minutes 10h10 Pause 10h30 1A Diadie SOW 15 minutes / 10 minutes 10h45 2A Hasan ABDULRAHMAN 20 minutes / 10 minutes 11h05 3A Abderrahman MOKNI 20 minutes / 10 minutes 11h35 Déjeuner 12h05 2A Blazo NASTOV 20 minutes / 10 minutes 13h30 3A Nawel AMOKRANE 20 minutes / 10 minutes 14h00 3A Stéphane BILLAUD 20 minutes / 10 minutes 14h30 3A Nicolas FIORINI 20 minutes / 10 minutes 15h00 Pause 15h30 3A Darshan VENKATRAYAPPA 20 minutes / 10 minutes 15h45 3A Sami DALHOUMI 20 minutes / 10 minutes 16h15 Synthèse – Yannick VIMONT 16h45 Fin du séminaire 17h00 Merci de votre participation ! Plus d’infos : http://www.lgi2p.mines-ales.fr/~urtado/SeminairesLGI2P.html

Contents First year PhD students

Imad BENKHALED Page 3 Designing an interface adapted to visually impaired: colorimetric characterization of displays and generation of appropriate image Pierre-Antoine JEAN Page 7 Uncertainty detection in natural language Lori LEMAZURIER Page 11 Flexible advanced AREVA’s PWR core control design using systems engineering and advanced control methods Massissilia MEDJKOUNE Page 16 Towards a non-oriented approach for the evaluation of odor quality Diadie SOW Page 20 Between policy aspirations and capacity to act: how designing achievable goals

Second year PhD students

Hasan ABDULRAHMAN Page 26 Model of noise and color restoration: application for restoration and steganalysis Blazo NASTOV Page 32 Model verification & validation and model-based systems engineering: towards executable DSML

Third year PhD students

Nawel AMOKRANE Page 38 Toward a methodological approach of requirements engineering with formal verification applied to the computerization of small and medium-sized enterprises Stéphane BILLAUD Page 42 Interoperability: A guarentee to adequately control and evolve the system of systems Mirsad BULJUBASIC Page 46 An effcient local search for large scale combinatorial optimization problems

Sami DALHOUMI Page 50 A framework for online inter-subjects classification in motor imagery-based brain-computer interfaces Nicolas FIORINI Page 54 USI at BioASQ 2015: a semantic similarity-based approach for semantic indexing Abderrahman MOKNI Page 58 Automating the evolution management of three-level software architectures Darshan VENKATRAYAPPA Page 62 RSD-HoG :A new image descriptor

First year PhD students

Page 2

Designing an interface adapted to visually impaired: colorimetric characterization of displays and generation of appropriate image

Imad Benkhaled, Isabelle Marc, Dominique Lafon-Pham

1

Introduction

The view is probably the sense which we attach the most importance, but the disease, the aging and accidents can cause degradation or total loss of sight. Current technological developments give hope that we can provide assistance to the visually impaired to assist in the tasks of daily living, including using virtual reality. In fact, virtual reality knows a major boom in recent years in the world of video games; headsets allow immersing themselves in an entirely imaginary world. But virtual reality is used also for design in fields such as architecture, automotive, etc. We present below the context in which the work of this thesis take place, namely a design project of technical assistance for the visually impaired based on the use of virtual

• reality. Then we recall some notions about color spaces, since the way to encode the images,

which is to say to present data, will have a great influence on the efficacy of processing. We will explain why we seek to precisely control the photometric characteristics of the images displayed with the calibration of the displays. Next, we present an image processing example can provide a solution to a problem faced by people living with the diseases that concerns us.

2

Contexte

Our team is working on the implementation of technical assistance for the visually impaired which is composed of an image acquisition system (camera), an image processing system, and a display system (virtual reality glasses). We consider a visually impaired user: an environment image is captured by the camera; the image is processed to make it more easily understandable and usable by the visually impaired, and finally displayed on the screens of virtual reality glasses. The aim of this thesis is the definition of the specific image processing for each kind of vision impairment. Specifically, we try to help persons with retinitis pigmentosa and glaucoma. Other teams are working on similar projects; these include the work of E.Peli and his team uses a method of enhancing contours, the contours are detected on the images using statistical methods [1]. Veringham et al performs a semantic analysis and a segmentation of images, the identified

• bjects being highlighted by changing their color [2].

3 Physical characterization of the equipment

We currently use for our prototype Wuzix Wrap 1200VR Virtual reality glasses, equipped with LCD screens, and Logitech cameras. This choice is not final, and we need software

developed treatments for these devices are directly transferable to other devices, that is to say that the aspect of the image displayed on the monitors should not depend on the screens themselves or the cameras used. For this, we must establish a link between the computer codes defining displayed images, and the physical quantities (emission spectra) of light emitted by the screens. Furthermore, these computer codes should be the same regardless of the equipment used for capturing images.

3.1 Different color spaces

A digital image is a color bitmap pixel matrix [3]; each pixel is represented by a point in a three-dimensional vector space, called a color space. There are a large number of color spaces, and a suitable choice of the working space can help to significantly improve the effectiveness of image processing. [4] We present below those that of interest to our work. The RGB color space is used in most of the display devices: in our LCDs, the light emitted by the backlight is white, and the color of a pixel is generated by three filters controlled by the RGB coordinates of the pixel considered. This RGB space is "device dependent" because the emitted light depends really on the filters. The CIE XYZ space [5] has justly been developed to be "device independent" and thus help to overcome the variability associated with different types of hardware. There are also spaces adapted to specific treatments. We cite in the first place the spaces used in video streaming: they are based on a representation of the three color components, a first signal encoding the luminance and two others for the chrominance. For example, the YCrCb code is developed from the RGB and XYZ information: the luminance is given by Y and chroma Cr = R-Y and Cb = B-Y. Finally, spaces were defined from the functioning of the human visual system. The LMS space is built on the absorption curves of three types of cones in the retina. The MBDKL space takes into account the processing carried out below of the retina, from LMS cones responses, with a brightness signal constructed from L + M responses and two chroma signals .

3.2 Physical characterization Method

We rely on the method described by Brainard et al. [6], knowing that the hypotheses formulated in the case of CRT screens also apply to LCDs. The screen must satisfy two

• assumptions. The first is that the channels must be independent, that is to say, driving a color

channel with different values does not change the distribution and quantity of the light emitted by the other channels. The second assumption is that of constant chromaticity channels: for a given channel, the shape of the emitted light spectrum is the same regardless of the value that the driver, it is equal to a factor to that emitted by the channel for a control value of 255 (because colors are coded on 8 bits).

Page 4

The first assumption help to write the XYZ coordinates for any RGB code as:

• Where
• corresponds to a code (255, 0, 0),
• is an encoded value (0, 255, 0),
• is an

encoded value (0, 0, 255). The last term is the ambient light reflected by the screen when R = G = B = 0. The second equation is used to write;

• Finally:
• 4

Results of characterization

We work in a dark room, so the term

• is zero. We use a Spectroradiometer CS-2000

MINOLTA KONIKA that delivers for each RGB the brightness and XYZ values of the emitted light and power density for all wavelengths between 380 and 700nm. We display on the screen three series of uniform color patches by varying a single channel at a time between its max and min values. The tone response curves vr(R), vg(G) and vb(B) are obtained by the measured values. We have established the following transition matrix for

• ur virtual reality glasses:
• 5

Image processing

The retinitis pigmentosa and glaucoma patients are very easily dazzled, and also suffer from nocturnal blindness [5] [6].!We seek to manipulate images in order to make brighter the dark areas and lower the brightness dazzling zones with maintaining the same scene details.

We select the YCrCb space for the images obtained from the camera, knowing that Y represents the brightness.!We apply the following processing on the images by manipulating Y without touching the CbCr chrominance. We start by applying histogram equalization! on channel Y.! then in the resulting image a luminance limitation is applied by taking the minimum value and the maximum values specific to each patient.!The last step is the conversion of the color space YCbCr to RGB color space for displaying final images on the LCD screens of virtual reality glasses.

6 Conclusion

We realized the color calibration of our virtual reality glasses, so we can now work in any color space while controlling the light spectrum actually emitted by the

• screens. We proposed a first type of treatment to improve the Visual comfort

avoiding glare and night blindness. The first results are encouraging but it remains to validate our work through rigorous clinical trials.

References

!

[1]

• 2012.

[2] Wearable Mobility Aid for Low Vision Using Scene

• Human-Computer Interaction Vol. 15, P. 231 - 244 Issue 2, April 2003..

[3] V. D. F. H. Foley, Computer Graphics, Addison Wesley. [4]

• pp. 695706, April 2006.".

"#$!!A. Kerr, The CIE XYZ and xyY Color Spaces Douglas 2010. "%$!!D. G. P. a. T. R. D. H. Brainard, " Display characterization. Encyclopedia of imaging

science and technology, 2002.".

"&$!!A. Kaufman, Rendering, Visualization and Rasterization Hardware. "'$!!r. c. g. r. e. woods, digital image processing 2004. "($!!Robert Sève, Science de la couleur : Aspects physiques et perceptifs, Marseille,

Chalagam, 2009..

Page 6

Uncertainty detection in natural language

Pierre-Antoine Jean1, Sébastien Harispe1, Sylvie Ranwez1, Patrice Bellot2, and Jacky Montmain1

LGI2P, Laboratoire de Génie Informatique et d’Ingénierie de Production, ENS Mines Alès, Parc Scientifique G.Besse, 30035 Nîmes cedex 1, France {firstname.lastname}@mines-ales.fr1 & patrice.bellot@lsis.org2 Abstract In order to benefit from knowledge contained in numerous non structured texts, automated text analysis techniques are necessary. Among them, relation extraction techniques propose to combine various approaches to extract entities and their relations from texts. Neverthe- less, the natural language contains many uncertainty expressions, to ex- press for instance uncertain statements, “It seems to me”, “It may be” which makes extraction results uncertain. Therefore, any knowledge base enriched through such analyses must take into account these uncertain-

• ties. This information will be of major importance to further treatments:

additional knowledge inference, reasoning or decision assistance, to cite a

• few. However, how to detect the uncertainty of natural language ? In ad-

dition, how to take into account this uncertainty ? This communication is dedicated to uncertainty in the context of Information Extraction. It defines the terminology, characterizes the several sources of uncertainty and discusses strategies that can be used to capture and consider this uncertainty. Keywords: NLP, information extraction, uncertainty detection

1 Context and positioning

Nowadays, knowledge bases such as DBpedia1, Yago2 or Freebase3 contain mil- lions of entities and hundreds of millions of facts used in many applications, e.g. rules mining [1], relation extraction[2] or question answering [3]. However, since knowledge bases remain incomplete large efforts are made to extract additional knowledge from text analysis using information extraction techniques. Tradition- ally, we can divide these techniques in three main tasks [4]: (i) named entity recognition to identify and disambiguate entities (Person, Localisation, Date...) in texts, (ii) relation extraction to extract semantic connections between these entities and (iii) event extraction which aims to automatically fill out a form to associate different information to a given event. Nevertheless, a common trait of these methods is that they do not consider the uncertainty

• f natural language and the retrieved information does not precisely transcribe

1 www.dbpedia.org 2 www.mpi-inf.mpg.de/departments/databases-and-information-

systems/research/yago-naga/yago

3 www.freebase.com

the thought of the writer. For instance, in the following sentence “According to some scientists, glycopeptide antibiotics inhibit protein synthesis”, methods of relation extraction will extract the triplet (glycopeptide antibiotics - inhibit - protein synthesis). Nevertheless, this triplet loses the knowledge transmitted by the uncertainty expression on the source “According to some scientists”. This un- certainty can impact outputs of inference knowledge methods or others methods which use it [5]. More generally taking into account uncertainty in natural lan- guage is central for numerous treatments and research areas related to natural language processing, e.g. in sentiment analysis, the distinction between object- ive and subjective facts is crucial since speculation tends to correlate with sub- jectivity [6]. In question answering the management of uncertainty can improve answers reliability [7].

2 Uncertainty in natural language

Uncertainty is not an anecdotal case in natural language. For example the Bio- logical Benchmark Bioscope [8] consists of 1261 abstracts of biological articles with 11871 sentences which 2101 are uncertain (17.5%), and SFU review corpus [9] contains 17263 review sentences from many domains which 3912 are uncer- tain (22.7%). Uncertainty is expressed in a variety of explicit ways in texts. Indeed, many “specific linguistic devices modify the meaning or reflect the au- thor’s attitude towards the content of text” [10]. For example, verbs with specu- lative content (suggest, presume), adjectives and adverbs referring to uncertainty (probable, likely, possible) or auxiliaries with modal verbs used to indicate mod- ality (may, might, should). Uncertainty also allows to build numerically vague expressions (some people, experts, many), use of the passive voice with dummy subjects to avoid specifying an authority (it is said, it is claimed that), use ele- ments expressing obviousness (“Clearly” - as if the premise is undeniably true - ) or use superlatives to bring subjectivity (best, one of the largest, most prom- inent). Many publications propose a classification to organize the various forms

• f uncertainty. For example, Rubin et al. [11] propose a categorization model to

identify certainty in texts. This categorization contains four dimensions: level, perspective, focus and time to annotate the certainty of a sentence. The first di- mension requires a judgement on the explicit markers certainty levels (e.g. “might buy” is a low level of certainty and “will almost certainly have to” is an absolute level of certainty). The second dimension allows to separate the certainty point

• f view into the writer’s and the reported points of view (e.g. “French histor-

ians, estimates that...” - involves third parties - ). The third dimension, focus, distinguishes abstract (judgements, opinions) and factual (reports of states or events, evidence and know facts) information in the narrative. Finally the fourth dimension takes into account the time (for instance the future is predictions, sug- gested actions so can involve uncertainty). Another classification was proposed by Vinzce in 2014 [12], she distinguishes two mains branches: discourse-level un- certainty and semantic uncertainty. The discourse-level uncertainty concentrates

• n three aspects namely, sources (“Some have claimed that”), fuzziness (“approx-

imately 86,000 inhabitants”) and subjectivity (“most distinguished orchestras”). Whilst the semantic uncertainty concerns propositions cannot be stated for sure

Page 8

whether they are true or false, given the speaker’s current mental state (“It may be raining”).

3 Uncertainty detection in natural language

How can we detect uncertainty in natural language ? The first shared task that deals with uncertainty detection is CoNLL 2010. In this conference, two uncer- tainty detection tasks in two domains (biological publications and Wikipedia articles) with three types of submissions (closed, cross and open) were given to the participants. The aim of task 1 was to develop automatic procedures for identifying sentences in texts which contain unreliable or uncertain informa-

• tion. This task is a binary classification problem (uncertainty vs certainty). For

task 2, in-sentence scope resolvers had to be developed. These resolvers had to automatically annotate uncertainty markers and the left and right boundaries

• f their scopes [10]. The best result for task 1 with biological texts obtained

86.4 of F-measure. The approach was proposed by Tang et al. (used too for the second task). It was a cascade method composed of two components: one for detecting uncertainty markers and another one for detecting their scope [13]. The first one is a two-layer cascaded classifier. There are a CRF4 classifier and a large margin-based classifier in the first layer and a CRF classifier in the second

• layer. The second one used another specific CRF classifier to predict 3 labels,

F_scope to indicate a token at the beginning of a scope sequence, L_scope to indicate a token at the last of a scope and NONE to indicate other. A similar cascade approach was recently used by Cruz et al. [6] using SVM classifier (with differences in features). Their method was tested on SFU review corpus [9] and

• btained 92,37% of F-measure for the detection of uncertainty markers. Then the

best result for task 1 with Wikipedia texts obtained 60.2 of F-measure. Finally the best result for uncertainty scope detection in task 2 with biological texts

• nly obtained 57.3 of F-measure. Therefore these results show that uncertainty

detection can vary according to the text domains. These results indicate that the binary classification with Wikipedia texts is not a trivial task, as well as the scopes detection for uncertainty markers.

4 Work in progress

This Ph.D thesis focuses on uncertainty management within knowledge extrac- tion from text. At the end of this first year, we observed that no tools dedicated to uncertainty detection is available. Therefore, we plan in a first time to build a system based on Tang’s method using two types of features: lexical features (stem of the token, Part-of-Speech, dictionary of markers) and syntactic features (chunk, dependency parsing, syntactic tree). The aim is to provide an exploitable tool and to evaluate the effect of enriching the model by external data. Indeed, few approaches in CoNLL 2010 tried this strategy (for example semantic meas- ures based on Wordnet could be envisaged). This pipeline will support further

4 Conditional Random Fields: Probabilistic framework for labeling and segmenting

sequential data based on a conditional approach [14]

developments of methods and software solutions to manage uncertainty. We plan to build an hybrid approach to assemble a relation extraction method with an uncertainty detection method. This tool might as well extract relations in texts with an annotation on the scope of uncertainty. This approach will serve to enrich a knowledge base with using Information Extraction techniques while consider- ing uncertainty expressed in natural language. In this way, further treatments (recommendation, decision aid or question answering) might take into account this uncertainty.

References

• 1. Galárraga, L., Suchanek, C.T.F., Hose, K.: Amie: Association rule mining under

incomplete evidence in ontological knowledge bases. In: 20th International World Wide Web Conference. (2013) 413–422

• 2. Theobald, M., Shah, N., Shrager, J.: Extraction of conditional probabilities of the

relationships between drugs, diseases, and genes from pubmed guided by relation- ships in pharmgkb. In: Translational bioinformatics. (2009) 124

• 3. Dong, L., Wei, F., Zhou, M., Xu, K.:

Question answering over freebase with multi-column convolutional neural networks. In: Association for Computational

• Linguistics. (2015)
• 4. Ferret, O.:

Approches supervisées et faiblement supervisées pour l’extraction d’événements complexes et le peuplement de bases de connaissances. In: PhD Thesis, University Paris 11 - Paris Sud (2011)

• 5. Jean, P., Harispe, S., Bellot, P., Ranwez, S., Montmain, J.: Gérer l’incertitude

lors de l’extraction de relations et lors de l’inférence de connaissances. In: RJCIA, Rennes (30 juin 2015)

• 6. Cruz, N., Taboada, M., Mitkov, R.: A machine learning approach to negation and

speculation detection. In: Association for Information Science and Technology. (2015)

• 7. Abacha, A., Zweigenbaum, P.: Means : une approche sémantique pour la recherche

de réponses aux questions médicales. In: Traitement Automatique des Langues. (2015) 71–104

• 8. Vincze, V.: Speculation and negation annotation in natural language texts: what

the case of bioscope might (not) reveal. In: Proceedings of the Workshop on Negation and Speculation in Natural Language Processing. (2010) 28–31

• 9. Taboada, M., Grieve, J.: Analyzing appraisal automatically. In: Symposium on

Exploring Attitude and Affect in Text. (2004) 158–161

• 10. Farkas, R., Vincze, V., Mora, G., Csirik, J., Szarvas, G.: Learning to detect hedges

and their scope in natural language text. In: CoNLL. (2010) 1–12

• 11. Rubin, V., Liddy, E., Kando, N.: Certainty identification in texts: Categorization

model and manual tagging results. In: Computing attitude and affect in text: Theory and applications, Netherlands, Springer (2006) 61–76

• 12. Vincze, V.:

Uncertainty detection in natural language texts. In: PhD Thesis, University of Szeged (2014)

• 13. Tang, B., Wang, X., Wang, X., Yuan, B., Fan, S.: A cascade method for detecting

hedges and their scope in natural language text. In: CoNLL. (2010) 13–17

• 14. Wallach, H.: Conditional random fields: An introduction. In: Technical Reports.

(2004) Page 10

Flexible Advanced AREVA’s PWR Core Control Design using Systems Engineering and Advanced Control Methods

Author: Lori LEMAZURIER E-mail: lori.lemazurier@mines-nantes.fr Address: 23 av. bauffremont 78170 La Celle Saint Cloud Thesis Dates: November, 1st 2014 - October, 31st 2014 Abstract: With the rise of electrical power de- mand, Nuclear Power Plant (NPP) flexibility is vowed to concern renewable energy global settlement by accompanying the decrease of other fossil-fuel ener-

• gies. Indeed, Nuclear Power benefits from high load

capacity and variations. However, with the rise of Renewable energies, the electrical network is destabi- lized because of the electrical power variation. NPP appears to be a suited solution subject to higher flex-

• ibility. In this way, AREVA proposed an innovative

solution to gain flexibility: the Primary Loop Flow- rate Variation. Basically, it consists in controlling the flow-rate (among other controlling means) in order to adapt Core Thermal Power Production. In order to answer this problematic, this thesis will focus on Advanced Control and Systems Engineering methods. Keywords: Systems Engineering, Requirement En- gineering, Identification, LPV, UKF, EKF, Kalman Filters, Nuclear Core Control

1 Thesis Presentation

1.1 The Subject

The thesis subject was

• riginally:

Flexible Advanced AREVA’s PWR Core Control Design using Systems Engineering and Advanced Control Methods. We can already notice that one of the main challenges will be to find the best way to combine the two fields that are to be used in this thesis.

1.2 Context

1.2.1 AREVA’s Need The energy issue that the thesis raises is one of the main goals for AREVA and its partners. In fact, for its future EPR (Evo- lutionary Power Reactors), AREVA is willing to improve the performances of it plants. By performances we mean flexi- bility, effluent minimization (Boron Acid for instance), and system solicitations (Rod movements for instance). Indeed, AREVA’s strategy is to use Nuclear Power Plant (NPP) to adapt electrical network energy demand. Therefore, AREVA uses its NPPs in Load Following control mode (cf. [1]). More-

• ver, with the rise of Renewable Energy (Wind Mill Turbines

for instance), the network tends to be more and more instable which justifies the increase of performance gain strategy. Although, performance seems to be central here, AREVA wishes to improve its Design Process. In fact, the thesis aims at finding a way of applying Systems Engineering Concepts to AREVA’s Design Strategy. The main reason of this is to have different designing actors who carry different skills to work together in trust towards an innovative solution satisfying all stakeholder needs. 1.2.2 Proposed Technical Solution The actual way of controlling the Nuclear Core is to play with both Rod Positions and Boron Acid Concentration in the Primary Loop. Both impact the core reactivity and in- duce average temperature variation (Average Temperature) and axial power distribution (Axial Offset). In fact, those last two are the parameters we wish to control. However, AREVA wants to improve Core Performances. In order to do so, the thesis proposes to focus on a new mean of action: Primary Loop Flow-rate Variation. Indeed, some internal researches showed that we could have impact on Average Temperature and Axial Offset (among other parameters of interest). How- ever, Flow-rate variation may impact other systems (such as Steam Generator) and many Requirements need to be consid-

• ered. Here we can see the Systems Engineering issues emerge.

In fact, AREVA has involved in a National Systems Engineer- ing project. Initiated by Thal` es, the Clarity Project involves several partners such as Obeo, AREVA, Airbus, etc. It aims at helping the development of Capella, a Thal` es Open-Source software based on MBSE approach (Model Based Systems En- gineering). It is used to describe and represent systems ar- chitectures and some of system design processes. However, Capella does not aim at managing Requirement Engineering (RE) and Integration Validation Verification processes. 1

2 Advanced Control State of the Art

2.1 LPV-SS representation

An LPV-SS (Linear Parameter Varying - State Space) repre- sentation aims at modelling multi-variable non-linear systems. It follows the equation: ˙ x(t) = A(p)x(t) + B(p)u(t) + E1(p)e (1) y(t) = C(p)x(t) + D(p)u(t) + E2(p)e (2) Where A, B, C, D, E1 et E2 have a static dependency in p, u is the input signal, e is a white Gaussian noise, x is the state vector of the system and y is the output vector.

2.2 Kalman Filtering for Identification

Kalman algorithms use State Space representation. This al- gorithm aims at estimating the future states of a dynamical system from noisy measurements. It is constituted in two phases: the prediction and the filtering. It is a recursive esti- mator which means it only requires the actual state and actual mesures the predict the future state. The actual state of the filter is represented by: ˆ x(t) estimated state at time t and P error covariance matrix. Kalman algorithm can be used in non-linear cases and for mul- tivariable systems (cf. [5]). Since this algorithm can predict the state vector of a system, it can be used to estimate its own parameters if they are seen as exogenous signals represented by a autonomous model state. Let’s get the equation of LPV representation. By concatenating the state vector with the parameter vector p the state vector becomes: xa(t) =

• x(t)

p ⇥ (3) where p is parameter vector to identify p = (p1, ..., pn)T , n ∈ N. So that dp

dt = 0 for LTI systems or dp dt = wp Wiener

noise for LPV representations. (cf. [6] and [7]). Therefore we have: ˙ xa(t) = fa(xa(t), u(t)) + w (4) y(t) = ga(xa(t), u(t)) + v (5) where fa et ga are the model function of the augmented sys-

• tem. w model noise et v measurement noise. Kalman algo-

rithm provides two different techniques to identify non-linear systems:

• l’EKF (Extended Kalman Filter) analytic derivation of

fa et ga by calculating Jacobian matrices.

• l’UKF (Uncented Kalman Filter) statistical alternative

that uses Sigma Point method to estimate states. [6], uses the EKF algorithm for systems having several de- grees of liberty. It highlights the correlation between structure degradation and model identification. [7] compares the 2 methods (UKF and EKF) in the case

• f aerodynamic model identification. More precisely in plane

wing parameter identification in real experimental conditions. It concludes that the UKF (in its entire form) provides better performances that EKF although it show honorable results. Finally, [8] shows an application of EKF algorithm for estimat- ing Li-Ion battery charge all along its life cycle (degradation

• f estimation with battery aging).

3 Systems Engineering (SE) State of the Art

The SE considers two different systems in the core of the de- sign process (cf. [9]). The bibliography highlights:

• SOI: System of Interest SoI): the system that is to be de-

signed in regards of the different stakeholder needs and constraints. In our case we will consider the Nuclear Core Control. The system is defined in a context called environment in which it shares services with another sys-

• tem. It has a purpose (why the system exists), a mission

(what the system needs to do) and objectives (how the system need to realize it). The system is subject to con- strains (imposed by its environment) and needs (coming from stakeholders). Following the Requirement Analy- sis, the system will benefit of a Requirement Baseline (A set of Requirements) which enables the System En- gineer to design Architectures (Functional and Organic) which respect the requirements. This last also conceives the Operational Scenarios, which will be sequenced in multiple Modes (cf. [9]).

• SUTD: the system (project, resources, supplies, activi-

ties, context, facilities etc.) which enables to organize, synchronize and execute all the activities used to de- sign the SOI. The SE is an approach that highlights processes, roles and core activities which allow to de- scribe the SOI. The SPEM (Software Systems Process Engineering Meta-Model 2.0 cf. [10]) 2.0 is a standard that defines a meta-model for the SUTD. Basically, it represents all the needed concepts to design the SUTD (Activities, Roles, Tasks, Milestones, etc.). The thesis’ main goal is basically to deploy an System Engi- neering tooled method for AREVA. The practical analysis, the inclusion in Capella Project and meetings with experts drew

• ur attention on Requirement Engineering (RE) process (More

particularly Stakeholders requirement analysis et System re- quirements analysis). RE’s goal is to ”Decompose”, ”Refine” and ”Reword” upcoming and down-coming requirements so that each could be understood by all designing entities (Ar- chitects for instance) in their own model language. Basically, a requirement is SMART (Specific, Measurable, Achievable, Relevant, Traceable). Here follow several directions we are to consider:

• The boilerplates

A boilerplate is a template consisting of fixed syntax 2

Page 12

elements and placeholders for attributes, having loose limitation on form and vocabulary (cf. [11]). Their first goal would be to facilitate verification and consistency, and to limit ambiguous requirements. For instance, CE- SAR (Cost efficient methods and processes for safety relevant embedded systems) (cf. [11]) is an European Project gathering Aeronautic, Automobile and Railway competences which focuses on Requirement Engineering

• process. It is based on several concepts, in particular

the RSLs (Requirement Specification Languages) which enables to describes specific types of requirements. Here follows an example of requirement: The <stakeholder> shall be able to <capability> within <performance> of <event> while <operational condition> We can denote a difference between Requirement Pat- terns and Boilerplates in the sense that Patterns use a higher formalism degree (more constrained concepts) (cf. [12]).

• Work Skill Ontologies

Structured set of terms and concepts associated to a Work vocabulary. We are currently deepening this part.

4 Advancement

4.1 Systems Engineering

4.1.1 Meta-Model Design In order to describe all the concepts that AREVA needs to design a system we conceived a Meta-Model. It sets up all the concepts and interactions between them to deploy Systems Engineering. 4.1.2 Capella’s description Capella has been used to design the Nuclear Core Control. It uses several phase that resemble the SE processes used to design an SOI. They are presented as follows:

• Operational Analysis (OA): Stakeholder needs and Re-

quirements.

• System Analysis (SA): System Missions and Capabilities

definitions and Functional Architectures.

• Logical Architecture (LA): Function grouping and rela-

tion to Logical Components.

• Physical Architecture (PA): Physical Architecture and

relations with Functional Architecture.

• Physical Architecture (PBS): Classification of compo-

nents. Figure 1: UKF Identification of a SISO 2nd Order system. Figure (1) show an extract of the Core Control Capella model.

4.2 Advanced Control

4.2.1 Identification

• Why Identifying the system?

AREVA provides a highly precise simulator to simulate the Nuclear Core behaviour. However it is too complex to apply advanced control methods on it. Therefore,

• ur strategy is to identify the Nuclear Core system to

dispose a simpler model for Advanced Control method application.

• Core Control Description

As said in part 1.2.2, the Nuclear Core control aims at regulating the two parameters Tm (Average Tempera- ture) and AO (Axial Offset: Axial Power Distribution).

• Control Parameters

– Rod Positions and Speeds There are 36 Control Rod, combined in 9 functional

• groups. Each group has a neutron weight (related

to its composition: Boron and Carbon mostly) and has more of less effect on the AO and Tm. – Boron Concentration The Boron Concentration is regulated through a distilled water injection system which relates both water and Boron Acid flow-rates. The Boron will impact the Core reactivity and so the – ... we would like to add the Primary Loop flow-

rate

We will study the impact of flow-rate variation on the Tm, AO and others. However, it will induce be- haviour changes which might violate requirements for the system or other environment systems. Here is one of the most interesting challenges for the the- sis in terms of SE and Control.

• Core behaviour

3

– Moderator effect A neutron moderator is a medium that reduces the speed of fast neutrons, thereby turning them into thermal neutrons capa- ble of sustaining a nuclear chain reaction involv- ing uranium-235 (U235) or a similar fissile nuclide. The moderator effect is a thermal reaction. When neutron moderation warms up, molecules expand which reduces the chance to slow down fast neu- trons and so slightly drops down the temperature. Conversely when the temperature cools down. NB: the Doppler effect has the same effect as the Moderator effect but for different physical reasons (not discussed here). – Nuclear Power EPR load varying power plant work in load follow- ing modes. It mean that the turbine is guiding the power to adapt to the electrical network. There- fore, the Nuclear Core’s goal is to provide the exact amount of power appealed from the secondary loop. Through different thermal effect (Steam Generator thermal exchange and others), the Nuclear Power is not actually managed by the Nuclear Core. Only the temperature is controlled though the Rods and Boron Concentration. – Xenon Xenon is a neutron catcher and poisons the reac- tion by absorbing neutrons. Therefore, it remains less for fissile matter. U235

6.2%

− → Te135

β−

− →

1.7s I135 β−

− →

6.6h Xe135 absorption

− →

σa=2.106barns Xe136

(6) U235

0.2%

− → Xe135

absorption

− →

σa=2.106barns Xe136

(7) 4.2.2 UKF Algorithm In view of the system to identify (Multivariable system, un- known structure and non-linear behaviour a priori), we chose to use the UKF algorithm coupled with the LPV-SS represen- tation to identify the Nuclear Core. This method successfully worked in many fields such in [7] or [8]. Here follows a SISO example with bi-diagonal matrix parameterization: ˙ x(t) = Ax(t) + Bu(t) (8) y(t) = Cx(t) + Du(t) (9) where, A = ⇤− 1

T1

a21 − 1

T2

⌅ , B = ⇤1 1 ⌅ , C = c1 c2 ⇥ , D = 0 (10) On augmente alors le vecteur d’´ etat: xa(t) = ⇧ ⌥ x(t) T1 T2 c1 c2 ⌃ ⌦ ⌦ ⌦ ⌦

• (11)

The figure 2 gives a result of the identification of parameters with the parameterization above and UKF algorithm. Figure 2: UKF Identification of a SISO 2nd Order system.

4.3 How to link SE and Control

It is important to remind that the thesis aims at applying both SE and Control domains. We evoked earlier in this paper

• ne link that could be deployed in the thesis (Develop an ad-

vanced control algorithm and make sure that all Requirements are validated thanks to a proper Requirement Engineering). Another way would be to implement the models we will get from the identification into Capella or another tool to simulate architectures

References

[1] IAEA, “Non-baseload operations in nuclear power plants: Load-following and frequency control flexible operations.” No. NP-T-3.23, 2014. [2] R. T´

• th,

Modeling and Identification

• f

Linear Parameter-Varying Systems. PhD thesis, University of Pannonia geboren te Miskolc, Hongarije, 2008. [3] L. Ljung, System Identification: Theory for the User. P T R Prentice Hall, second ed., 1987. [4] D. Jones, “Estimation of power system parameters,” IEEE Transactions on Power Systems, vol. 19, no. 4, pp. 1980–1989, 2004. [5] R. G. Brown and P. Y. C. Hwang, Introduction to Random Signals and Applied Kalman Filtering. John Wiley and Sons Inc, second ed., 1992. [6] A. Corigliano and S. Mariani, “Parameter identification in explicit structural dynamics: Performance of the extended kalman filter,” Computer Methods in Applied Mechanics and Engineering, vol. 193, pp. 3807–3835, 2004. [7] G. Chowdhary and R. Jategaonkar, “Aerodynamic parameter estimation from flight data applying extended and unscented kalman filter,” Aerospace Science and Technology, vol. 14,

• pp. 106–117, 2010.

[8] S. Sepasi, R. Ghorbani, and B. Y. Liaw, “A novel on-board state-of-charge estimation method for aged li-ion batteries based on model adaptive extended kalman filter,” Journal of power sources, vol. 245, pp. 337–344, 2014.

4

Page 14

[9] AFIS, “D´ ecouvrir et comprendre l’ing´ enierie syst` eme,” 2009. [10] OMG, “Software and systems process engineering meta-model 2.0.” SPEM 2.0, 2008. [11] V. Johannessen, “Cesar - text vs. boilerplates,” Master’s the- sis, NTNU - Trondheim Norwegian University of Science and Technology, 2012. [12] A. Mitschke and al., Definition and exemplification of RSL and RMM. CESAR Project, 2010. [13] M. Esteki, G. Ansarifar, and M. Arghand, “Estimation of the xenon concentration and delayed neutrons precursors densi- ties in the pressurized-water nuclear reactors (pwr) with slid- ing mode observer considering xenon oscillations,” Anals of nuclear energy, vol. 77, pp. 1–22, 2015. [14] R. V. D. Merwe and E. A. Wan, “Sigma-point kalman filters for integrated navigation,” Proceedings of the 60th Annual Meeting of The Institute of Navigation, pp. 641 – 654, 2004. [15] J. Maciejowski, Multivariable Feedback Design. Addison Wes- ley Publishing Company, 1989.

5

Towards a non-Oriented Approach for the Evaluation of Odor Quality

Massissilia Medjkoune1, St´ ephane Cariou2 Jean-Louis Fanlo2, S´ ebastien Harispe1, and Jacky Montmain1

1 Laboratory of Computer Science and Production Engineering (LGI2P) Ecole des

mines d’Al` es, Parc scientifique G.Besse, 30035 Nˆ ımes cedex 1, France. {Firstname.name}@mines-ales.fr

2 Laboratory of Engineering for Industrial Environment (LGEI) Ecole des mines

d’Al` es, 6 avenue de Clavi` eres, 30319 Al` es cedex, France.

• Abstract. When evaluating an odor, non-specialists generally provide

descriptions as sets of terms (it smells like: lemon, sugar. . . ). The ob- jective of this work is to propose a model for merging the information expressed by such descriptions; this will make possible to evaluate odors through non-specialists descriptions. Indeed, currently, only some ori- ented approaches based on shared reference vocabularies and learning techniques exist. These protocols rely on trained specialists and domain- specific descriptors that, despite being efficient in characterizing odors, are costly and restrictive both in terms of descriptors and evaluators (since non-specialists cannot be used). This paper briefly discusses a non-oriented approach based on semantic analysis, it does not require learning a lexical field and can therefore be used to evaluate odors using non-specialists descriptions. Keywords: proximity measure, information fusion, odor quality

1 Introduction

According to several investigations, people have become increasingly sensitive to issues related to pollution and environment – ISAAC study (INSERM2007) and APHEKOM study (INVS-2011). In big cities and large industrial areas, people are today more than ever attentive to air pollution. In particular, bad smells, the second reason of complaint after noise, may have very bad effect on people judgement. Bad smells are indeed often considered to be aggressions and are generally simply perceived as risks for individual health. In this context, numerous enterprises are directly concerned (and therefore interested) by odor quality evaluation. Among others, evaluating odor quality enables them to be attentive of their image, by reducing air pollution and paying close attention to the populations located near their industrial sites, but also to improve their competitiveness by refining product odor, that is to say perceived quality.

Page 16

2 Evaluation of odor quality

Evaluation of odor quality is a subjective and complex process. Once chemical molecules which constitute the smell are captured by receivers (neurons), a cog- nitive process characterizes and evaluates the smell. The decomposition of this cognitive process underlines that evaluations are related to descriptors that may be based on olfactory relationships between odors and lived situations (happy or unhappy). This is based on these descriptors that odor quality is generally anal-

• ysed. Among the approaches proposed in the literature, the wheel of odors and
• dor fields are often used [3]. These methods use an oriented approach based on a

common referential to qualify the odors, which facilitates their characterization. Indeed, forcing evaluators to use specific descriptors facilitates understanding, interpreting and processing the results. Nevertheless, a learning phase in which valid descriptors have to be learned is required to use such methods - which prevents their use by non-specialists, implies additional training costs, and limit the number of evaluators and experiments that can be used to evaluate an odor. The aim of this work is therefore to propose an approach to evaluate an

• dor by analyzing natural language descriptions provided by non-specialists.

Our approach has to be a non-oriented and to require neither prior training nor learning of a specific lexical field.

3 Modeling

The heart of the problem is to identify descriptors that best describe the smell of an odors by analyzing natural language descriptions provided by non-specialists

• technically speaking, the aim is to summarize an odor by a set of weighted

non-ambiguous descriptors. Descriptors are assumed to be partially ordered into a taxonomy O = (, C), with C the set of descriptors (concepts). It is here considered that the terms that can be used by non-specialists to describe the

• dor are free. Therefore, contrary to descriptions provided by experts in oriented

approaches, the terms we have to deal with are not part of an implicit controlled and standardized vocabulary defining the terms commonly used for character- izing (specific) odors. Otherwise stated, it is here assumed that people provide their critics using their own words and vocabulary. In addition, this major dif- ference implies that intrinsic properties of natural language such as the potential ambiguity of terms, e.g. due to polysemy, or the distortion of senses carried by words have to be considered when extracting general knowledge from free word descriptions. For a given odor evaluation, descriptions provided by non-specialists are as- sumed to be sets of terms in which each term is associated to a specific in-

• tensity. As an example, the descriptions provided by two evaluators testing

the smell of a yoghurt could be: d1 = [(orange, 4), (cream, 3), (sugar, 1)], and d2 = [(lemon, 4), (butter, 3), (brownsugar, 1)]. The problem we face can therefore simply be formulated by the following question: how to fusion the information

expressed by k descriptions d1, d2, . . . , dk in order to characterize the odor? Defining a model for answering this need requires characterizing the seman- tics, i.e. meaning, of the terms used in the descriptions, e.g. intuitively, looking at d1 and d2, human understand that the terms “orange” and “lemon” refer to “citrus fruits” (putting aside the problem of ambiguity). This is because our knowledge about the world provides us a taxonomic organization of knowledge that we can use to derive conclusions on the basis of deductive reasoning. Indeed, in this specific case, we know that the concepts Orange and Lemon both refer to specific types of CitrusFruit, i.e. formally Orange ⇥ CitrusFruit and Lemon ⇥ CitrusFruit which implies that both descriptions d1 and d2 implicitly refer to the concept CitrusFruit, a concept that could therefore be a good candidate to characterize the odor. As we have seen, such a reasoning approach extensively relies on a taxo- nomical knowledge organization partially ordering concepts. This enables us to derive logical implications, e.g. a kind of ontological knowledge about the

• world. However, in other cases, such reasoning will not be based on logical im-

plications justified by concepts ordering (⇥). Consider the following descrip- tion d3 = [(lemon cake, 4), . . .]. Here, analyzing d1 and d3, the conclusion that CitrusFruit could be a good candidate for describing the odor could also be derived despite the fact that LemonCake CitrusFruit. This is because our reasoning abilities also exploit and take into account the semantic proximities between concepts (here LemonCake and CitrusFruit) when extracting knowl- edge from observations. In this context, one of the challenges of this project is to use, adapt and define Artificial Intelligence and Data Mining techniques for automatizing inference procedures based on semantic data, in particular sets of terms. In defining the global strategy we will use, we propose to consider that a term per se brings to mind concepts (e.g. “lemon cake” evokes Lemon) and can therefore be reduced to a vector representation traducing the degree to which a term evokes each concept. Informally the strength of evocation a term has with regard to a concept can be regarded as the semantic proximity between the term and the concept. We now briefly present a model that we propose to study for deriving such vector representations of words. They will next be used by fusion information techniques to derive the description of the odor. We first define the function σT C that will be used to compute the semantic proximity between a term and a concept: σT C : T C ⇤ [0, 1] t c ⌅⇤ σT C(t, c) with T a set of terms and C = {c0, c1, . . . , cn} the set of concepts (descriptors) defined into an ontology. We next define ProC the vector-based representation that is used to represent the strength of evocation a term with regard to each

Page 18

concepts of the ontology (C). ProC : T ⇤ R|C| t ⌅⇤ ProC(t) = (σT C(t, c0), σT C(t, c1), . . . , σT C(t, cn)) Since the set of terms T and the set of concepts C are two incomparable sets, we cannot establish the correspondence or measure the proximity semantic between terms and concepts in a straightforward manner without considering the labels (L) associated to the concepts. For this, we will first define a corre- spondence between the terms and labels (L ⇥ T) that refer to concepts, with Lc the labels associated to the concept c. This correspondence between terms and concepts is defined as the measure of proximity between terms and labels (terms) associated to concepts. The proximity between terms and concepts will be calculated using σT C, according to the measure σT T which assesses the prox- imity of two terms: σT T : T T ⇤ [0, 1]. Numerous measures for comparing terms have already been proposed in the literature [1, 2]. Then, the semantic proximity between a term t and a concept c can be calculated, by example, as the maximum of the similarity values between the term and the labels associated with the concept c: σT C(t, c) = maxl∈Lc σT T (t, l). Finally, thanks to the function ProC, to each user which provide a set of terms will be associated a set of term-vector representations. The collective evaluation will therefore be made using Information Fusion techniques to merge the sets of term-vector representations.

4 Conclusion

In this work, we propose a model that automates the cognitive pattern of assess- ing the quality of an odor from descriptions provided by non-specialists evalu-

• ators. For this, we project the terms given by the evaluators in the conceptual
• space. We are now working on how to merge the vectors of terms to find concepts

that characterize best the odor quality.

References

[1] Besanon, R.: Int´ egration de connaissances syntaxiques et s´ emantiques dans les repr´ esentations vectorielles de textes. PhD thesis, Ecole Polytechnique F´ ed´ erale de Lausanne (2001). [2] Harispe, S., Ranwez, S., Janaqi,S., Montmain,J.: Semantic Similarity from Natural Language and Ontology analysis, Synthesis Lectures on Human Language Technolo- gies (2015) 8:1, 1-254. [3] Jauber, JN.: Chap 4-1 Mthode du ”champ des odeurs” in Pollutions olfactives coordinateur ADEME DUNOD Paris mars (2005).

Between policy aspirations and capacity to act: how designing achievable goals Diadié Sow Abdelhak Imoussaten Pierre Couturier Jacky Montmain Centre de Recherche LGI2P/Ecole des mines d'Alès, Site EERIE.

Abstract

Developing new products or improving existing ones leads to the emergence of products with more features and which therefore must meet a number of interdependent objectives / criteria that may be difficult to achieve simultaneously. The variety of interactions between the criteria appears to be a central concept but yet ambiguous in the design or improvement of a product. Two types of approaches are brought face to face in our works. Approaches from the multicriteria decision community that attach to model the preferences of decision makers and approaches related to industrial engineering or software engineering which give prominence to the functional constraints that seem to govern the achievability of a set of objectives. We attempt to provide a unifying framework for building consensus between commitment to do and ability to do, inspired by

• ptimization methods of type gradient descent based on two functions characterizing the two aspects
• f the improvement project.
• 1. Introduction

!

Improving a system raises two distinct problems: (i) to be able to assess whether a system can be improved or not in accordance with a preference model in a multi-dimensional evaluation space, (ii) identify the configurations of the system able to meet in the most satisfactory way the objectives/criteria. These two distinct issues seem to cause confusion around the concept of interaction or dependence between the objectives. We propose to come back to these different modeling approaches of interactions between the objectives/criteria and then to provide a unifying approach under the name "Choosing Achievable goals for a multicriteria continuous improvement of Performance" (CAP).

• 2. Problematic

2.1 Presentation of the problem

!

We characterize a system to be designed or improved by its input parameters . The parameters values can be chosen by the system designers. Let the set of all possible values of vector . A system is then defined by a configuration or alternative . Improving a system consists to make it evolve from one configuration to another configuration constraints. Let { be the set of all objectives (goals) induced by the system requirements

• analysis. Three issues arise from such a definition of system improvement: (1) What are the

interactions between the input parameters? Which configurations are admissible, i.e. which configurations meet the functional requirements? (2) What are the relationships between the input parameters and the objectives? (3) What are the relationships between the objectives,

• del?

Two main families of approaches are used in the literature to address the above issues: MAUT type approaches that identify the most profitable improvement dimensions without worry about the feasibility of the assigned objectives; Performance Engineering approaches that emphasize the means to achieve the improvement project.

Page 20

2.2 M AUT type approaches

!

Let be the parameters of interest for the stakeholders. Such parameters are called

• utput parameters. Let be the set of these output parameters. Let us denote by N the

set of indices . To each output parameter is associated a utility function ! that quantifies the degree of satisfaction of the ith goal when output parameter takes the value . We refer to this degree of satisfaction as the performance and we denote it , i.e., ).. To finalize the preference model, the overall assessment of a system characterized by its outputs can be defined by a single consolidated criteria: ) where is an aggregation function. We assimilate hence the overall system performance to the value of performance ). 2.3 Performance Engineering approaches

!

According to [1], the aggregation of criteria should not be based on aggregation operators specified by some axiomatic result of a multi-criteria analysis of preferences, but should incorporate an explicit modeling of relationships between these criteria. For this aim, the author of [1] defines for each criteria , the set of alternatives that improve and the set of alternatives that can degrade . By reasoning on the degrees of inclusion of the sets,,,, ( ) the behavioral interactions between objectives can be established. For example if and then the objectives i and are competitors and can not be improved simultaneously. A conjunctive aggregation of two criteria would thus not make sense according to this view. By the same token, i.e., evaluate the satisfaction or the dissatisfaction of a goal, the authors of [2] propose a model that consists to decompose the final goal into sub-goals as a directed

• graph. Some of these sub-goals can be directly observed and their satisfaction or

dissatisfaction are propagated to evaluate the one of the final goal. The model of [2] is not limited to logical AND- OR relations in such a graph. The authors also define the influence relationships between goals. The alternatives are associated with tangible goals, i.e. goals on which direct actions are possible, e.g., "Reduce material costs raws Improve economics of production". These tangible goals are the sources of the influence graph [2]. To each

• bjective-node of the graph is associated a pair (sat(i); den(i)) which is the information that is

available about the satisfaction (respectively the degradation ) of the objective (more exactly

• f the performance associated with the objective/criteria) for a given alternative, i.e. for a

given state of tangible goals in the graph entry. Works in [1] and [2] are interested in the ability to do and not to the commitment to do. The directed graph structure proposed in [2] is simply "contracted" in [1] where alternatives are not explained, but just defined through the sets and for each criterion i. The models in [1] and [2] have therefore similar semantic but they use different formalisms (probabilistic propagation, versus fuzzy inclusion). In an application framework, a mathematical model semantically complete is proposed [7]. The authors focus on autonomous robots which can be adapted to different operating contexts. They are interested in both robot modes (normal, low battery, full memory, etc.), but also to functional constraints that have to be addressed by the possible configurations of the system (permissible values of input parameters). This model highlights the two dimensions that guide the decision in choosing a configuration when designing or improving a system.

The homogeneity of this formalism is attractive, but the behavioral model is masked by the expert construction of feasibility relationship.

• 3. Unifying framework

!

As mentioned in the introduction, our aim is to provide an unifying framework for building a

• makers/managers to avoid two risks: in the one hand to pursue unattainable goals, in the other

hand to focus on solutions that have already been done easily. The unifying idea in our works is to follow the steepest slope of the "gradient" of the performance to best improve the product, as far as the improvement remains achievable i.e. feasible as regards available means and resources. It is then proposed to model such a feasibility property through a qualitative expression of the relationship between the inputs parameters of the system and its objectives. d with the

• a subset of objectives has to be reached.

3.1 Degree of feasibility of a subset of goals

!

Let T:

be the transformation that gives the value of the output parameters (attributes)

• f the system obtained from vectors of input parameter values

, where , is the restriction

to all eligible configurations satisfying the functional requirements and constraints imposed to the system. An output parameter value (or attribute value) is the response of the system configuration for the ith attribute: ) =. We denote ( ) = (

), )) the

attribute vector for the configuration . For complex systems, determining ) for

is

very difficult. Often one has to content himself by a qualitative impact model that links input parameters or actions to expected characteristics [3, 4]. For this purpose, we adopt a fuzzy model as described in [11,12]. We define for each elementary performance and each action a (a change in value of an input parameter), !the degree of belief with which the action a can affect a performance positively (respectively !the degree of belief with which the action a can affect a performance negatively). When , we say that a supports when , we say that a degrades. The qualitative relationship between actions and performance should then be extended to a configuration change (simultaneous actions on several input parameters). The difficulty is that for a configuration change and a criterion, several elementary actions of can affect positively as negatively. The resulting effect cannot be calculated objectively but can therefore be reduced to a decision problem under uncertainty. Consider a configuration change and a performance ,the effect resulting from the combination of actions in on can be based upon two sub-sets of :

)= ap: !>0} and )={ ap:

!>0}. In this bivariate context, the effect is thus modeled by the pair (! !, where !(resp ! is the degree to which supports (resp degrades). (resp !) is the result of aggregation of values {! a

} ( resp { ! a

• }.

Now consider a configuration change and a performance, the resulting degree of belief to which contributes to the improvement of is defined as: ! ! if

Page 22

! !! else. Finally we define the degree of belief to which p contributes to the improvement of by = . 3.2 Priority degree of a subset of objectives

!

The priority degree to which a subset of objectives must be achieved is borrowed here from [8]. In [8], an index (called worth index) denoted and quantifying the worth of improving vector in criteria among , and subject to the evaluation function is

• defined. The idea is to define for each an index that assesses the average benefit to cost

ratio of improving criteria in [8]:

• 1

\

( )( ) [ 1 1 , ( )]

I I I N I

H p H p p H p d

• where , is the restriction of p on B.

The subset of criteria I* which maximizes !indicates the performance values that are the most beneficial to improve. We define = the ability to realize I (regardless of the cost of such a realization). When favors coalitions with high cardinality, favors however coalitions with a small cardinality, the problem is then to find the sensible compromise. The subset of criteria I* whose improvement will be most profitable while not being too difficult to achieve, can then be given by the optimization problem:!

• This optimization problem can then be solved by a branch and bound algorithm.
• 4. Conclusion

!

to

• criteria
• improvement. We have proposed a framework to unify sometimes conflicting viewpoints

emanating from different communities. It highlights the need to consider simultaneously know-how and ambition in a project to design or improve a system. Reference

[1] Felix, R. (1994). Relationships between goals in multiple attribute decision making. Fuzzy sets and systems, 67:47-52. [2] Giorgini, P., Mylopoulos, J., Nicchiarelli, E. and Sebastiani, R. (2002). Reasoning with Goal Models, 21st Int.

• Conf. on Conceptual Modeling (ER02), Tampere, Finland, 167-181.

[3] dans les systèmes complexes, LFA'2009, Annecy, France. [4] Montmain, J., Labreuche, C., Imoussaten, A., Trousset, F. (2015). Multi-criteria improvement of complex systems, Information Sciences, 291:61-84 [5] Felix, R. (1994). Relationships between goals in multiple attribute decision making. Fuzzy sets and systems, 67:47-52. [6] Felix, R. (2008). Multicriteria Decision Making (MCDM): Management of Aggregation Complexity Through Fuzzy Interactions Between Goals or Criteria, IPMU, Malaga, Spain. [7] Fleurey, F., Delhen, V., Bencomo, N., Morin, B. and Jézéquel, J.M. (2008). Modeling and validating dynamic adaptation, 3rd , Toulouse, France [8] Labreuche, C. (2004). Determination of the criteria to be improved first in order to improve as much as possible the overall evaluation, IPMU, Perugia, Italy, pp. 609616

Second year PhD students

Page 24

MODEL OF NOISE AND COLOR RESTORATION: APPLICATION FOR RESTORATION AND STEGANALYSIS

PH.D. T HESIS SUMMARY REPORT − SECOND YEAR

Hasan ABDULRAHMAN 2,4, Marc CHAUMONT 1,2,3 and Philippe MONTESINOS 4

1 Nˆ

ımes University, Place Gabriel P´ eri, F-30000 Nˆ ımes Cedex 1, France.

2 Montpellier University, UMR5506-LIRMM, F-34095 Montpellier Cedex 5, France. 3 CNRS, UMR5506-LIRMM, F-34392 Montpellier Cedex 5, France. 4 Ecole des Mines d’Al`

es, LGI2P, Parc Scientifique G.Besse, F-30035 Nˆ ımes Cedex 1, France.

ABSTRACT Digital images, especially color images, are very widely used, as well as traded via Internet, e-mail and posting on websites. Images have a large size which allows embedding secret messages of large size, so they are a good medium for digital steganog-

• raphy. The main goal of steganalysis is to detect the presence of hidden messages in digital media.

In this report, we propose a new steganalysis method based on correlation between different channels of color images and machine learning classification. Indeed, features are extracted from the channel correlation and the co-occurrences correlation. In this report, all stego images are created with a range of different payload sizes using two steganography S-UNIWARD and WOW algorithms. To validate the proposed method, his efficiency is demonstrated by comparison with color rich model steganalysis. Index Terms— Steganalysis, color spatial rich model, channel correlation, co-occurrences, ensemble classifiers.

• 1. INTRODUCTION

Steganography is the art of secret communication, by embedding messages within media such as audio, image or video files, in a way to make difficult the detection of these hidden messages. In general, steganography requires two parts, in the first part are embedded the messages and in the second part, these messages are extracted [1]. Several steganography techniques are used for different purposes, as, for the most prominent, hide the illegal activities or secret communications [2]. There are many steganography methods that have been developed to enable secret communication. So, the real-world needs to develop methods and algorithms rapidly in order to detect the presence of hidden messages and, thus, identify this secret communication; this method is called steganalysis. Krichner et al. [3] suggested a steganalysis method to detect LSB replacement steganography in color images. These stego images are produced via processing which reveals some traces of Color Filter Array (CFA) interpolations. Here the authors have enhanced the Weighted Stego (WS) image steganalysis method [4] by replacing the cover predictor in WS with position specific predictors. This technique puzzles out the local predictability of pixels, depending on their position in the CFA interpolation to account for differences between cover (common) images and suspect images. From two different components, Goljan et al. [5] developed a color image model which is called the Spatio-Color Rich Model (SCRMQ1, where ’C’ represents the color version of the Spatial Rich Model -SRMQ1- [6], with ’Q1’ representing fixed quantization q = 1 ) produced from two different components. The first component is the SRMQ1 with a single quantization step q = 1 and dimensionality of 12753 features. The SRMQ1 feature is computed for each color channel separately, then the three features are merged to keep the same dimensionality, as for grayscale images. The second component is a collection of 3D co-occurrences with a larger threshold T = 3 and q = 1. The final dimensional features obtained from this component are 5404 features. The color rich model is built from the same pixel noise residuals as those used in the SRMQ1, but is formed across the three channels of each pixel. Embed messages inside the digital medium involves some slight changes of this medium. Research in steganalysis aims to develop some methods able to detect as much as possible these modifications. Although the real-world uses significantly color images more than the grayscale images, there is a great development for the steganalysis to handle grayscale images, but, at the same time neglecting to deal with color images [7].

Page 26

This report, is organised in five sections, Section 2 is describe the color rich model steganalysis method. In section 3 describe the proposed method by recalling the color channel correlations. Experimental results an comparisons with its tables and curves are given in section 4. The final section talk about the conclusion and future work.

• 2. COLOR RICH MODEL

The following steps describe the process by which the first set of features is computed. We used the spatial color rich model for color image steganalysis in [5] which produced 18157 features. This method extracts the noise residual from each color channel separately by the following formula: Rij = ˆ Xij(Nij) c · Xij. (1) where c ⌅ N is the residual order, Nij is a local neighborhood of pixel Xij at coordinates (i, j), ˆ Xij(·) is a predictor of c · Xij , Xij ⇧⌅ Nij, Xij ⌅ {0, ...., 255}. ˆ Xij represents a pixel value of an 8-bit grayscale cover image and its corresponding stego image. All of the submodels (Rij) ⌅ Rn1×n2 are formed from noise residual images of size n1⇥ n2 computed using high pass filters of the following form: Rij ⇤ trancT

• round

Rij q ⇥⇥ , (2) where trancT represents the truncation function with T > 0 defined for any x ⌅ R, trancT (x) = x for x ⌅ [T, T] and trancT (x) = T · sign(x) and otherwise q is the quantization step, and round is a function for rounding to an integer value. The spatio-color rich model consists of two different components. On one hand, the spatial rich model (SRMQ1) [6] with a fixed quantization q=1 and truncation T=2 yields a dimensionality of 12753 features. These features are computed from each R, G and B color channel separately. Finally, the three dimensionality features are added together to keep the same dimensionality as for grayscale images. On the other hand, from the same noise residuals (i.e. SRMQ1), the CRMQ1 builds a collection of 3D color co-occurrences, taking three color values at the same position (across the three channels of each pixel). Thus, with fixed truncation T=3 and quantization q=1, CRMQ1 produces 5404 features.

• 3. PROPOSED METHOD

Our proposition is to enrich the SCRMQ1 with an inter channel correlation which is composed of two sets of features. The first set, produced by [5], gives 18157 features. The second set gives 3000 features obtained from the correlation of different R, G, B channels gradients. In The following we gives explanation about our proposition. 3.1. Channel Correlation In this section, we introduce an inter-channel correlation measure, and demonstrate that it can be listed to first order Euclidean invariants (see Hilbert [8] for the invariant theory). Such invariants have mainly been used for stereo-matching [9]. In this paper, we show that the information provided by them can enhance steganography detection. Starting from the local correlation of red and green channels (similarly, correlation of red and blue channels) : CorrR,G(i, j, k, l) = ⇤

(i0,j0)∈Wi,j

X

(R)

i0,j0 X

(G)

k+i0,l+j0

(3) with:

• X

(R)

i0,j0 ⌅ [0, 255], being a pixel value at position (i, j) in the red channel,

• X

(G)

k,l ⌅ [0, 255], being a pixel value at position (k, l) in the green channel,

• Wi,j, being a small window centred in (i, j).

Considering (k, l) = (0, 0) and a limited development of X

(R) and X (G) around (i, j), then :

CorrR,G(i, j, 0, 0) = ⇤ h = (ii, jj) (i, j) ⇤ Wi,j

• X

(R)

i,j + ⌅X

(R)

i,j · h

⇥ X

(G)

i,j + ⌅X

(G)

i,j · h

⇥ . (4) Developing this equation leads to four terms. Three of which are constant or not informative, then there is only one informative term : ⌅X

(R)

i,j · ⌅X

(G)

i,j .

(5) If only one channel has locally been altered, the gradient in this channel is modified. Consequently, the scalar product of two channel gradients reflects the change in the cosine of the difference between the two gradient angles. Similarly, we can apply the same computation for the red and blue channel and then obtain : ⌅X

(R)

i,j · ⌅X

(B)

i,j

(6) As stated by Gouet et al. [9] (and following the Hilbert theory [8]), it is unnecessary to investigate the ⌅X

(G)

i,j · ⌅X

(B)

i,j term,

as it is already implicitly contained in the first two expressions (Eq. 5 and 6). Our proposition in this paper is to add two new features depending on the normalized correlation between red and green channels, and red and blue channels as : GRG = ⌅X

(R)

i,j · ⌅X

(G)

i,j

|⌅X

(R)

i,j | |⌅X

(G)

i,j |

(7) GRB = ⌅X

(R)

i,j · ⌅X

(B)

i,j

|⌅X

(R)

i,j | |⌅X

(B)

i,j |

(8)

• Fig. 1 illustrates our preprocessing steps to obtain normalized correlations between gradients of each channel. Note that

gradients of each channel are estimated by convolution with a [-1; 1] mask (horizontal and vertical). Then, our features are computed from these correlations GRG and GRB, by computing the co-occurrences as in rich model [6]. We used different values of the quantization q ⇤ {0.1, 0.3, 0.5, 0.7, 0.9, 1} with fixed truncation T=1. For each quantization, we

• btain 12 submodels, one symmetrized spam14h and one spam14v with 25 features each , and one of minmax22h, minmax22v,

minmax24, minmax34h, minmax34v, minmax41, minmax34, minmax48h, minmax48v, and one minmax54, with 45 features for each. All submodels are gathered in a one dimension vector to erect a dimensionality of (10 ⇥ 45 + 2 ⇥ 25) ⇥ 6 = 3000 features which are added to 18157 features of color rich models with fixed q = 1 and T = 3 in order to collect a final set of 21157 features.

!"#$%&'&()$*"& +,-$&!./00$,& 1%$$0&!./00$,& 2$3&!./00$,&

• Fig. 1: Diagram illustrating the preprocessing steps to obtain correlations between gradient in each channel.

Page 28

• 4. EXPERIMENTAL RESULTS

4.1. Building Database A raw image is a class of computer files containing untouched pixel information coming from the digital camera sensor. These files hold a large amount of meta-information about the image generated by the camera [10]. In our work, the color image database is very carefully built depending on the CFA idea. We collected raw images from two subsets which are the most standard, and have the highest number of images captured (i.e. the Dresden Image Database [11] 3500 full-resolution Nikon digital camera raw color images and the Break Our Steganographic System (BOSSbase1) 1000 Canon digital camera raw color images). In order to obtain color images in Portable Pixel Map (PPM) format of size 512×512, all images take the same CFA map

• layout. Two steps are required. The first step consists of using a demosaicking algorithm to convert raw images in demosaicked

images . The second step consists of cropping five sub-images from one image. First we use demosaicking algorithm Patterned Pixel Grouping (PPG) in the dcraw platform2 to be able to read numerous raw formats used to convert raw images in PPM and TIFF formats. We wrote a spatial code in C language under Linux to start the crop from the red channel position. Indeed, from one image, this code randomly selected the red channel position and cropped five images of size 512×512, so that all blocks share the same CFA map layout. The final number of images used 10000 PPM color images of size 512×512 to the first experimental, and then we built another six databases using different demosaicking algorithms. 4.2. Experimental Protocol Stego images are obtained using two steganography algorithms: the Spatial- UNIversal WAvelet Relative Distortion (S- UNIWARD3) steganography algorithm [12] and Wavelet Obtained Weights (WOW4) steganography algorithm [13]. As shown in [14], these two algorithms are highly adapted because they are selected for this strong resistance to steganalysis. The payload is embedded in equal proportion in the three R, G and B channels. The different payload sizes are {0.01, 0.05, 0.1, 0.2, 0.3, 0.4 and 0.5} Bit Per Channel (BPC). Two sets of features are extracted from each image, the first set is the color rich model vector which consists of 18157 features of residuals among neighboring pixels (see section 2). The second set of features consists of 3000 proposed features from correlations between channels by applying gradient derivative filters on each channel, as explained in section 3.1. Then we add the two sets of features in a one dimensional vector to get a result of 21157 features of 10000 covers and 10000 stegos which are ready to enter in the classifier. The images are randomly divided into two halves for training and testing. A random subset of images, 5000 covers and 5000 stegos is used to train and test the classifier. The result of the classification is the average testing error over 10 splits ¯

• PE. The rest of cover and stego

images were tested against the ensemble classifiers [15]. Then, the decision values were collected for each. Given the decision values, ROC curves are obtained. The area under the ROC curves was calculated as the accuracy of the ensemble classifiers. 4.3. Results and Discussion This section contains the experimental results of our proposed method. We illustrate these results in Tables 1 and 2. S- UNIWARD and WOW methods were tested with different relative payloads {0.01, 0.05, 0.1, 0.2, 0.3, 0.4, 0.5} Bits Per channel (BPC) against the method of color rich model [5]. We used the same set of payload values with the same embedding methods. Our method, achieved higher performance by registering 87.54% and 86.63% detection rates for S-UNIWARD and WOW respectively with the payload 0.5 bpc. At the same time, the color rich model method [5] is less efficient because it achieved respectively 86.14% and 85.27% detection rates on the same sample. We noted the same thing with the rest of the payload values for both (the proposed and compared method), as shown in Tables 1 and 2. Curves in Fig. 2 (a) and (b) also illustrate the comparison between the proposed method and the color rich model. As a result, the average testing error of the proposed method is less than that of the color rich model. That proves the importance of the additional 3000 features derived from the correlation between different channels proposed by our method.

1BOSSbase can be accessed at http://www.agents.cz/boss/BOSSFinal. 2dcraw code is available at http://www.cybercom.net/defin/dcraw. 3S-UNIWARD steganography method is available at http://dde.binghamton.edu/download/stego_algorithms/. 4WOW steganography method is available at http://dde.binghamton.edu/download/stego_algorithms/.

Embedding Proposed method Color rich model Payload Average testing error Detection Rate% Average testing error Detection Rate % 0.01 0.4830 ± 0.0010 51.70 0.4841 ± 0.0017 51.59 0.05 0.4010 ± 0.0032 59.90 0.4045 ± 0.0026 59.55 0.1 0.3203 ± 0.0023 67.97 0.3298 ± 0.0016 67.02 0.2 0.2370 ± 0.0031 76.30 0.2498 ± 0.0026 75.02 0.3 0.1808 ± 0.0034 81.92 0.1947 ± 0.0023 80.53 0.4 0.1470 ± 0.0025 85.30 0.1599 ± 0.0015 84.01 0.5 0.1246 ± 0.0023 87.54 0.1386 ± 0.0023 86.14

Table 1: The comparison between the steganalysis of the proposed method with the color rich model in different relative

payloads for the S-UNIWARD steganography method: the average testing error for the proposed is less than for the color rich model method. Embedding Proposed method Color rich model Payload Average testing error Detection Rate% Average testing error Detection Rate % 0.01 0.4836 ± 0.0017 51.64 0.4850 ± 0.0014 51.50 0.05 0.4042 ± 0.0027 59.58 0.4092 ± 0.0032 59.08 0.1 0.3317 ± 0.0034 66.83 0.3397 ± 0.0023 66.03 0.2 0.2502 ± 0.0041 74.98 0.2654 ± 0.0025 73.46 0.3 0.1918 ± 0.0013 80.82 0.2081 ± 0.0029 79.19 0.4 0.1574 ± 0.0021 84.26 0.1783 ± 0.0034 82.17 0.5 0.1307 ± 0.0015 86.63 0.1473 ± 0.0025 85.27

Table 2: The comparison between the steganalysis of the proposed method with the color rich model in different relative

payloads for the WOW steganography method: the average testing error for the proposed is less than for the color rich model method.

0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 Embedding payload Probability of error PE Proposed method Color rich model

(a) S-UNIWARD

0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 Embedding payload Probability of error PE Proposed method Color rich model

(b) WOW

• Fig. 2: Detection error ¯

PE as a function of the payload for (a) S-UNIWARD and (b) WOW steganography methods illustrates good detectability with our proposed features.

Page 30

• 5. CONCLUSION AND FUTURE WORK

Here we proposed and implemented a new method for color image steganalysis. The proposed method estimates correlations between the gradients of red, green and blue channels. Then these correlations are incorporated in the rich model using a co-occurrence matrix in order to obtain 3000 features. These features are added to those that are obtained from the color rich model in order to build a vector of a total of 21157 features. All feature vectors are fed to the ensemble classifiers. We used a quantization step with a set of values that differed from the color rich models. The ensemble classifiers is used to detect steganographic messages. Images for cover and stego are collected using the Dresden database and BOSSbase 1.0. Multiple comparisons are made between the proposed method with the color rich model using WOW and S-UNIWORD steganography methods in different payloads, reflecting the efficiency of the proposed features. Our future work will focus on developing a new steganalysis method for digital color images, enhance the feature vector and calculate a new correlation between all channels of the color image. REFERENCES [1] J. Fridrich, Steganography in digital media: principles, algorithms, and applications, Cambridge University Press, 2009. [2] C. Hosmer and C. Hyde, Discovering covert digital evidence, In Digital Forensic Research Workshop (DFRWS), 2003. [3] M. Kirchner and R. Bohme, Steganalysis in technicolor boosting ws detection of stego images from cfainterpolated covers, In IEEE International Conference on Acoustics, Speech and Signal Processing, pp. 39823986, 2014. [4] J. Fridrich and M. Goljan, On estimation of secret message length in lsb steganography in spatial domain, In International Society for Optics and Photonics in Electronic Imaging, pp. 2334, 2004. [5] M. Goljan, J. Fridrich, and R. Cogranne, Rich model for steganalysis of color images, In IEEE Workshop on Information Forensic and Security, GA, 2014. [6] J. Fridrich and J. Kodovsky, Rich models for steganalysis of digital images, In IEEE Transactions on Information Forensics and Security, vol. 7, no. 3, pp. 868882, 2012. [7] J. Fridrich and M. Long, Steganalysis of lsb encoding in color images, In IEEE International Conference on Multimedia and Expo (ICME), pp. 12791282 vol.3, 2000. [8] D. Hilbert, Theory of algebraic invariants, Cambridge University Press, 1993. [9] V. Gouet, P. Montesinos, and D. Pel, A fast matching method for color uncalibrated images using differential invariants, In BMVC, pp. 110, 1998. [10] L. Yuan and J. Sun, High quality image reconstruction from raw and jpeg image pair, In IEEE International Conference

• n Computer Vision (ICCV), pp. 21582165, 2011.

[11] T. Gloe and R. Bohme, The dresden image database for benchmarking digital image forensics, In ACM, pp. 15841590, 2010. [12] V. Holub, J. Fridrich, and T. Denemark, Universal distortion function for steganography in an arbitrary domain, In EURASIP J. Information Security, pp. 113, 2014. [13] V. Holub and J. Fridrich, Designing steganographic distortion using directional filters, In IEEE International Workshop

• n Information Forensics and Security (WIFS), pp. 234239, 2012.

[14] V. Holub and J. Fridrich, Digital image steganography using universal distortion, In ACM Workshop on Information Hiding and Multimedia Security, pp.5968, 2013. [15] J. Kodovsky, J. Fridrich, and V. Holub, Ensemble classifiers for steganalysis of digital media, In IEEE Transactions on Information Forensics and Security, vol. 7, no. 2, pp. 432444, April 2012.

Model Verification & Validation and Model Based Systems Engineering: towards executable DSML

Blazo Nastov1, Vincent Chapurlat1, Christophe Dony2 and François Pfister1

1LGI2P, Parc Scientifique G. Besse, 30000 Nîmes, France,

firstname.lastname@mines-ales.fr

2LIRMM, 161 rue Ada, 34392 Montpellier, France,

firstname.lastname@lirmm.fr

1 Dynamic semantics and MBSE

Achieving Verification & Valicationd (V&V) objectives in field of Model Based Systems Engineering (MBSE) require first and above all, profound comprehension and understanding of all Domain Specific Modeling Lan- guages (DSML) concepts. For this mater, the abstract and concrete syntaxes alone are insufficient, due to the ambiguous meaning one concept might have if observed and sometimes handled by different experts. It is therefore es- sential to explicit in precise manner the DSML semantics. The literature highlights two different kinds of semantics; one describing only the meaning, independently of the behavior, static semantics, and the other, describing the behavior, dynamic semantics [Com08]. Static semantics are implicitly and partially given through the abstract and concrete syntaxes. Dynamic semantics however, have to be additionally explicated by formally defining a behavior for each evolving concept. Note that concepts that have associated behavior are referred as “evolving concepts”, taking part into the DSML’s dynamic semantics. There are several tools and approaches for defining static and dynamic se-

• mantics. Among the principle effective solutions are: EMF/Ecore for ab-

stract syntaxes, Sirius (graphical) or Xtext (textual) for concrete syntaxes, Kermeta for operational semantics, ATL for translational semantics and OCL for axiomatic semantics. Nevertheless, such tools and approaches are related to software engineering programming languages (e.g. imperative, object-

• riented or aspect-oriented languages) and are still unfamiliar to MBSE ex-
• perts. Indeed, dynamic semantics is to be modeled (not programmed) and

formalized with minimal efforts from an expert by assisting him and au- tomating the DSML design and V&V process as much as possible. This is, for instance, possible by using formal behavioral languages such as 1

Page 32

Statecharts, Petri Nets, or Finite Automata. The dynamic semantics of a DSML, i.e. the behavior of all DSML’s concepts, is then represented by a single behavior model, an instance of the used formal behavioral language. This one must be, furthermore, checked for well-constructiveness, coherence and conformity to its metamodel, and simulated for development and vali- dation purposes by a simulation dedicated tools. However, several questions remain still an open issue: (1) Formal behav- ioral languages also have dynamic semantics which unfortunately, may vary. For instance, the literature highlights 21 different dynamic semantics for the Statecharts language. (2) Dynamic semantics of a DSML are given via single behavioral model, in contrary to the previously argued approaches and pro- gramming languages, both defining the behavior for each DSML’s (evolving)

• concept. (3) Additional efforts are requested for modeling the interactions

between the dynamic semantics of different DSMLs (by synchronization or by composition). This may be impossible if different behavioral languages are used (e.g. Statecharts for one DSML and Petri Nets for the other). (4) Some properties are described only through the interactions between the dy- namic semantics of DSMLs. For instance, the behavior of a DSML’s concept Function cannot / should not be in execution state if its physical component (a concept from another DSML) is in a breakdown state.

2 Executable DSMLs

As a response, we propose a formal approach for modeling dynamic semantics

• f a DSML as a set of behavioral models associated to the evolving concepts.

The approach is formalized in an extension of the metamodeling language Ecore by composing it with a metamodel of dynamic semantics. This allows DSML designers to build dynamic semantics of the DSML that achieve and control all interactions between behavioral models. Simulation and model checking techniques can be then applied as proposed in the approach. The behavior of an evolving concept can be described by using discrete or continuous approach, giving respectively so-called discrete events, continu-

• us or hybrid (i.e. mixing discrete and continuous hypothesis) model. For

each type of behavioral models, a common base of information, an execution scheduler and a formal language, are requested. We focus here, however, only

• n discrete events behavioral models, proposing an extension of the Finite

Sequential Machine called Interpreted Sequential Machine (ISM) [LCMC97]. Discrete Events Behavioral Models: Behavioral models build using

• ur language are composed of four interconnected parts called respectively

Input Interpreter (II), Output Interpreter (OI), Control Part (CP) and Data Part (DP) as illustrated in Figure 1. The CP is a graph of states and tran- sitions and the DP holds the model data. The II interprets inputs data 2

Figure 1: The components (modules) of an ISM model (gathered into the set I) available in the Black Board (BB) and data from DP in order to evolve the CP by the means of firing conditions associated to each transition of the CP. The OI is an interface that interprets the evolution

• f the CP by updating the values of the output (gathered into the set O) and

the values of the data from DP. The Black Board (BB) is a common base

• f information where all behavioral models of a DSML write their outputs

(O) and read inputs (I). It can also be seen as the part of our approach that enables the information exchange between ISMs models. The Global Clock (GC) is a clock input, scheduling the execution of all behavioral models of a DSML according to a logical and periodical clock. The time instants of a logical clock are named Logic Time Units (LTU), in contrary to Physical Time Units i.e. seconds, minutes, etc. respecting temporal hypothesis pre- sented in the next. Stability management: modeling critical, parallel or distributed sys- tems requires the management of stability of created models. A model is stable if succeeding its evolution cycle’s completion, taking into account the same inputs, the model cannot evolve in another state i.e. its current state does not contain an outgoing fireable transition. Otherwise, the model is unstable and its current state is named “transient” state. Managing this sta- bility involves a model evolution taking into account a double scaled internal and external time, modeled by two different logic clocks. Executable Metamodeling Language: existing metamodeling lan- guages can be extended in order to integrate the notion of dynamic seman- tics (behavior). As illustration, Figure 2 shows the composition between a metamodeling language and the dynamic semantics metamodel. The re- sulting language is furthermore promoted to the M3 (meta-meta-modeling layer) substituting the previous metamodeling language. The composition process is designed to ensure that the existing and already defined DSMLs, (conforms to the previous metamodeling language,) remains fully compatible 3

Page 34

with the new metamodeling language. Figure 2: Promoting to the M3 layer a new Executable Metamodeling lan- guage As a result, Figure 3 illustrates the metamodel of our Executable meta- modeling language. Using our metamodeling language, experts can design DSML along with their dynamic semantics. Created models by such DSML can interoperate and be simulated. Figure 3: The metamodel of an Executable metamodeling language.

References

[Com08] Benoit Combemale. Approche de métamodélisation pour la sim- ulation et la vérification de modèle–Application à l’ingénierie des procédés. PhD thesis, Institut National Polytechnique de Toulouse-INPT, 2008. [LCMC97] Mireille Larnac, Vincent Chapurlat, Janine Magnier, and Ben- jamin Chenot. Formal representation and proof of the inter- preted sequential machine model. In Computer Aided Systems Theory—EUROCAST’97, pages 93–107. Springer, 1997. 4

Third year PhD students

Page 36

!"#$%&'$'()*+"&","-./$,'011%"$/+'"2'3)45.%)6)7*8' 97-.7))%.7-'#.*+':"%6$,';)%.2./$*."7'011,.)&'*"'*+)' <"615*)%.=$*."7'"2'>6$,,'$7&'()&.56?>.=)&'97*)%1%.8)8' '

!"#$%&!'()*+#,%-!./,0%,1!23#45+&#1-!',,%67/8%!2)5+9/8-!:3)(#8!7#(9)&#/8!#,;! <)388/,%!=#33)5>! 7?@AB-!! :%#(!@CDE!F!@,1%+)4%+#9&%!CG81%(!H!D+I#,/J#1/),!E,I/,%%+/,I!K!

@'''<"7*)A*'$7&'1%"1"8$,'

Requirements engineering (RE) is a major step in any project to enhance its chances to succeed. For many reasons, this step is particularly crucial when it is question of (re)engineering information systems to support the business activities of Small and <%;/5(!8/J%!E,1%+4+/8%8!LC<E8M>!! D5+! +%8%#+03! /8! ;+/N%,! 9G! #,! /,;581+/#&! ,%%;! )O! #,! @:68%+N/0%! %,1%+4+/8%! ,#(%;! =ECP7@C-!$3)8%!Q)9!/8!1)!0)(451%+/J%!#&&!)+!4#+1!)O!C<E8!#,;!4+)N/;%!13%(!$/13! #;#41%;!)+/I/,#&!8)O1$#+%-!;%8/I,%;!#+)5,;!13%/+!958/,%88!1)!(re)engineer their @,O)+6 (#1/),!8G81%(8!L@CM>!@,!O#01! 13%!&#11%+!0#,!9%!8%%,!#8!13%! ;+/N%!9%&1!)O!/,O)+(#1/),! 9%1$%%,!;%0/8/),!#,;!)4%+#1/),#&!8G81%(8!4+)N/,I!#;;%;!N#&5%!)O!13%!%,1%+4+/8%>!This is why RESULIS need to carry a special effort to 5,;%+81#,;!#,;!3#+,%88!13%!$#G!13%! C<E!)4%+#1%8!951!/1!&#0*8!)O!(%#,8!1)!O)+(#&&G!I#13%+!#,;!N#&/;#1%!13%!+%R5/+%(%,18! )O!13%!N#+/)58!81#*%3)&;%+8!L958/,%88!%S4%+18-!;%0/8/),!(#*%+8!#,;!%,;!58%+8M>! E84%0/#&&G! 13#1! 13%G! 585#&&G! 3#N% different! 05&15+%8! #,;! N)0#95&#+/%8! 0)(4#+/,I! 1) project management stakeholders (requirements engineers, designers, engineers). Also, stakeholders do not have the skills to use requirements elicitation tools

• r modeling languages. :3/8!/,;50%8!;/OO/05&1/%8!/,!(#,#I/,I!13%!#01/N/1/%8!)O!%&/0/1#6

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

Page 38

!"""#$"%&'()*+,-"%,."/0*+1+2%'+),"%33*)%2("

Business process models serve traditionally as the main source of requirements [3], but still this functional view of the enterprise is not sufficient to grasp all IS require-

• ments. The actors in charge of the tasks have to be known along with their role and

position in the enterprise or its environment. The information used as inputs and out- puts of the tasks have to be traced and the resources needed for the achievement of the tasks are relevant information too. For this, Enterprise Modeling (EM) approaches [4, 5, 6] allow to represent, understand and engineer the structure and behavior of the enterprise providing various modeling languages, frameworks and reference models. Therefor we consider that classical requirement models should be linked to enterprise models as the latter represent the concerns that have to be addressed when building an

• IS. Based on that we defined a! "#$#%&'! '($'#)*+,-! .(/#- [7] inspired by the

ISO19440 standard [1] to build the abstract syntax of our RE authoring languages. It covers the classical functional, informational, organizational, and resources views to which were integrated the requirements of all internal and external stakeholders. This ,--(0!.(/#-&$"!*1#!0,2!%#3+&%#.#$*4!%#-,*#!*(!*1#!(*1#%!.(/#-&$"!5&#046!,44#44&$"! *1#!.,*'1&$"!-#5#-!7#*0##$!4*,8#1(-/#%!,$/!424*#.!%#3+&%#.#$*4!,$/!)%(5&/#!9+4*&:&; ',*&($!:(%!/#4&"$!/#'&4&($4<!

• they can be transferred in confidence to the design team in a well-written, verified and

validated requirements repository. So, a systematic RE process provided with model- e- ments are reflected by the different views (functional, resources, organization and information) of the enterprise model where stakeholders express their current situa- tion (e.g. current activities description) and their expectations (e.g. being able to per- form new activities with data and resources, taking into account constraints and exist- ing applicatie- ments defied with collaboration with the design team to decide what has to be kept or created. Our approach of requirements authoring is close to what is done in real interviews while ofThe stakeholders progressively build the different views of the enterprise model answering a series of predefined questions using a proposed set of boilerplates [8]. To begin with, the questions are addressed to decision makers (managers or heads of department). Next, the questions are intended for enterprise members to represent their roles and roles on which they have good

• knowledge. The requirements authoring process consists of the steps described in the

following:

Step 1- Framing by defining enterprise context, business domains and partners in-

teractions: business domains and environment of the enterprise are captured. The scope of the enterprise part to be modeled is defined and limited by domains; a do- main is a functional area representing a business of the enterprise. For each domain, relevant inputs and outputs that are exchanged with external partners are defined. The exchanges are later detailed in the functional view in step 3.

Step 2- Organization and resources views definition: this step is concerned with

the definition of the structure and resources of the enterprise. It provides a first look

• f the distribution of the responsibilities among the human resources according to

their workstation (position) in the different organization cells main resources achievement of tasks in each workstation are also listed.

Step 3- Function and Information views definition: according to each business

domain, single roles are defined where a role is the function that an enterprise mem- ber plays while intervening in the enterprise activities. Indeed, in SMEs the enterprise members may plays different distinguish roles. Here, the focus is on what and how the work is done while playing a single role, without having to picture the whole execution of the business processes. The role is defined in terms of tasks. Each task is delineated by a set of actions that can be triggered by notifications (receiving some kind of enterprise object or being notified by a state change). Informational Input and

• utput are defined and more precision is given for the used resources. The business

rules that have to be respected during the achievement of the task are also specified, linking the rule to its application. The generic conceptual model representing the abstract syntax of the boilerplate based textual language we proposed is also a basis for consistency analysis and verifi-

• cation. We perform verification over the textual IS requirements models, by first

providing guided requirements authoring with such boilerplates. We defined a formal definition of their syntax (the language's grammar), which supports the users in the process of entering syntactically correct inputs. We use Natural Language Processing (NLP) techniques as well for part of speech tagging to check whether the grammatical formulation of the requirements corresponds to the boilerplates. Nonetheless the well- formedness of each requirement boilerplate does not guaranty the overall consistency and the absence of omission, imprecision and duplication. To check the existence of such defaults we defined a set of verification rules (technically defined into model queries and constraints) where the requirements given by one stakeholder or different stakeholders are crossed within the same view of the enterprise model or inter-views.

• stakeholders may describe differently the same tasks. Throughout these mechanisms

we address six requirements problems: complexity, ambiguity, inconsistency, duplica- tion, omission and imprecision. Furthermore, !"#$%"#&'!#!'%()&*#'&#+$()&*#,-"#'.# $/0$&1"/#.'%+$2#3"14&)5,"-#678#3'#0"%).9#34"#'0"%$22#1'&-)-3"&19#'.#$#-"3#'.#:,-)&"--# %,2"-# Restitutions through +'/"2# 3%$&-.'%+$3)'&-# .%'+# 34"# 3";3,$2# +'/"2-# 3'# -)+<2"# *%$<4)1$2#&'3$3)'&-#!)22#:"#-"332"/#3'#offer global views of the enterprise and favor expert reviews, visual restitutions are provided from the gathered textual IS require-

• ments. The different interactions with external partners are depicted in a context dia-

gram and the organizational structure of the enterprise and the allocated resources are presented in organigrams. Business processes are derived in a bottom-up way from the local role-based descriptions of the tasks achieved by enterprise members or ex- ternal stake-holders. This is assured by means of a mechanism that uses notifications to trace the workflow throughout the enterprise structures. Such restitutions provide awareness among enterprise members of the impact of the tasks they achieve and the use of their productions in the overall business processes of the enterprise. At all events, our aim is to enhance e- quirements so that they can be transferred in confidence to the design team in a well- written, verified and validated requirements repository. Functional System Require-

Page 40

ments are mainly extracted from the definition of the tasks achieved within the differ- ent roles. These requirements define the features that the system must propose and the constraints that must be considered are defined in terms of information, resources, internal organization and external relationships.

!"""#$%&'()*$%"

In the scope of the thesis we focus on requirements authoring and verification objec-

• tives. Authoring and verifying requirements needs the involvement of enterprise

members, as they are the most likely to define the enterprise structure, behavior and

• expectations. A requirements authoring and verification approach suitable to that

context is proposed. It is compliant with Enterprise Modeling reference frameworks and favoring the autonomy of SME stakeholders. Verification, arbitration and restitutions are used to enhance the quality of the Stakeholder Requirements and to provide with confidence System Requirements to the development team. The benefit of our work is to be a trade-off between the auton-

• my of the SMEs stakeholders that requires a simple formalism and the level of for-

malization needed to operate and automatize the requested support for these engineer- ing activities. !

+,-,.,%&,)"

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`C!"<<`#!!! `# G93K-2*,! ]?! 5H2.6-;2+! M?! 536-S/0! G#V#?! V29S3;2/0! @?! ! Q2HH36! E#! E3:,;/*1! 8-29,4 J3-K0?!9,+H3:0!2*:!;2*1621,0!83-!739.6+,-/a/*1!%92;;!2*:!E,:/6940/a,:!)*+,-.-/0,0^!

• ,T/,J!2*:!.-3.302;#!$*+,-3.,-2S/;/+F!83-!)*+,-.-/0,!%F0+,90!2*:!G..;/72+/3*0!Y$4)%GZ?!

``A`#!G;S/?!I-2*7,?!D>"=#!! A# G93K-2*,!]?!5H2.6-;2+!M?!536-S/0!G#V#?!V29S3;2/0!@?!!Q2HH36!E#! Q,X6/-,9,*+0!G64 #!$IG5!%F9.30/69! 3*!$*83-92+/3*!53*+-3;!/*!E2*6827+6-/*1!Y$]5&EZ#!&++2J2?!52*2:2?!D>"B#! <# R32*1! @#! %?! $+3H! %?! &F292! b?! E/F2a2K/! b?! P! %2+3! ]#! M2;/:2+/*1! +H,! 53*0/0+,*7F! 38! %.,7/8/72+/3*!Q6;,0#!D>"B!

Interoperability: A guarentee to adequately control and evolve the System of Systems (SoS) [Internal report - LGI2P]

St´ ephane Billaud, Nicolas Daclin, and Vincent Chapurlat

LGI2P, Laboratoire de G´ enie Informatique et d’Ing´ enierie de Production, ENS Mˆ ınes d’Al` es, Parc Scientifique G.Besse, 30035 Nˆ ımes cedex 1, France {firstname.lastname}@mines-ales.fr

• Abstract. A System of Systems (SoS) presents a number of functional

and non-functional characteristics that need to be satisfied all over its life

• cycle. Interoperability is one of the non-functional characteristics, which

is from our point of view, essential in order to guarantee the control of the SoS, its behavior and the fulfillment of its global mission(s). Moreover, this interoperability ensures the behavior of the SoS to deal with some risky situations and with potential local or global deficits during the evo- lution of the SoS. In this report, we present an overview of the work that has been done to propose a general matrix to determine the impact of interoperability, which is divided into sub characteristics/requirements (Compatibility, Interoperation, Autonomy and Reversibility), on the so- called analysis perspectives of the SoS: Performance, Integrity and Sta-

• bility. The evaluation of this impact is based on the use of a simulation

technique which turns to be adequate face to the dynamic aspect of the

• SoS. Thus, a set of indicators are derived and formalized

Keywords: System of Systems (SoS), Simulation, Performance, Stabil- ity, Integiry, Adaptability, System Engineering, System of Systems En- gineering (SoSE), Interoperability

1 Introduction

A System of Systems (SoS) is a set of heterogeneous systems called sub- systems, assembled together to achieve a global mission, that a system alone cannot fulfil, while maintaining the operational and managerial indepen- dency (autonomy) of each of the subsystems. The subsystems are able to communicate together and to adapt locally and globally face to the evolu- tion of the SoS and to any change in their environment [1] [2] [3]. It is admitted that the SoS Engineering (SoSE) carefully focuses on choosing, assembling and interfacing existing systems to build their SoS [4]. The subsystems are selected and involved according to some constraints particularly their capacity of be- ing interoperable. In this context, and as demonstrated by the literature and the System Engineering domain, interoperability is required to coordinate such

Page 42

large multi-disciplinary subsystems and to accommodate the inevitable evolu- tion of the SoS. Therefore, interoperability has to be fully considered when the subsystems decide to set up their SoS. Moreover, a strong linkage exists between the interoperability and other non- functional SoS requirements: analysis perspectives (Stability, Integrity and Per- formance) [4]. Therefore and in response to this challenge, the original aim of this report is to present briefly the impact of the interoperability on the so- called analysis perspectives by proposing an impact matrix. With respect to the state-of-the-art and to the best of our knowledge, the requested interoperability and interaction between heterogeneous and socio-technical subsystems have not been yet handled before and the achieved work characterizes the novelty of the

• approach. It is evidently a new challenging area and there is research directions

towards discovering it [5]. This report focuses first on presenting the reasons behind considering the interoperability as a crucial characteristic of the SoS. Secondly. we expose a matrix to assist the engineers, designers and managers involved in SoSE process in choosing their subsystems prior the assembling and understanding the impact

• f the interoperability over the SoS analysis perspectives. The impact is evaluated

through a simulation technique. Therefore, a set of indicators has been defined in order to concertize the existing relationship between the interoperability and the SoS analysis perspectives

2 Interoperability versus SoS characteristics

The term of interoperability has been defined recurrently in the literature in a way to provide a better understanding of its various aspects and levels [6]. However, we will not delve into the debate of choosing a convenient and rele- vant definition since all definitions share mostly the same concepts. Thus, and in attempt to reach to a consensus, we define the interoperability as the ability

• f connected, autonomous, loosely coupled, compatible and heteroge-

neous systems to coexist, to interoperate and to exchange data and services to/from other systems while continuing their own logic of operation and con- serving their autonomy. This definition of the interoperability evokes a number of characteristics which are consistent with the SoS characteristics presented in the previous sec- tion (See section 1): – The autonomy of a subsystem is the given possibility to continue to act and make decisions in order to ensure its own mission(s) independently of other subsystems. – The reversibility is the given ability to a subsystem to achieve its mission, after breaking an alliance with other subsystems inside a SoS. It is the abil- ity to satisfy its initial requirements (e.g. initial level of performance) after leaving the SoS. The reversibility and autonomy are consistent here with the notion of op- erational independency and managerial independency of the SoS.

– Interoperability is the action of coupling subsystems together. These sub- systems are seen as loosely coupled. On the one hand, this kind of coupling enhances the connectivity inside a SoS, where subsystems are capable of building links among their interfaces and destroying them dynamically [7] and on the other hand, it enhances the evolutionary development of the SoS [3] when it becomes possible to easily remove, modify or add SoS entities to the SoS. – The interoperability considers the heterogeneity (diversity) between the sub-

• systems. This diversity is essential for the SoS since it can only achieves its

global mission by leveraging the diversity of its subsystems [7]. . . . It is obvious that the interoperability is essential to guarantee a good control

• ver the SoS. Maintaining a good interoperability helps the preservation of the

SoS characteristics (constituent subsystems autonomy, enriched connectivity and commitment to diversity of SoS manifestations and behavior).

3 Interoperability’s impact matrix

Interoperability has been studied in multiple fields, it is one of the key con- cern in the SoS and it is important factor for the success of collaboration and

• interaction. The literature presents approaches to measure and evaluate the in-
• teroperability. These approaches are mainly based on maturity measurement

and most of them focus on the interoperability of a particular kind of systems (e.g. information systems). A few models integrate the organizational aspects of

• interoperability. Moreover, none of these models have been presented in large

systems or organizations like the SoS and [6] did not present any evidence of that. In our previous work performed, we focused on evaluating the impact of inter-

• perability on the SoS analysis perspectives [8]. The interoperability is divided

into four basic sub requirement of interoperability: Compatibility, Interopera- tion, Autonomy and Reversibility. A set of indicators is given to express these sub requirements. In order to allow the interoperability, these indicators consider three barriers: Conceptual barriers (they concern the semantic and syntactic dif- ference of exchanged information. They are mainly related to modeling at high level of abstraction.), Technological barriers (they concern the incompatibilities

• f information technologies (protocols, coding, architecture and platforms etc.))

and Organizational barriers (they concern the definition of authority, trust, in- tellectual property, responsibility to allow to the interoperability to take place under good conditions. These barriers are related to the organizational structures and management techniques used in multiple enterprises). Table 1 shows how the variation in the interoperability sub requirements impacts the analysis perspectives of the SoS [8].

Page 44

• Fig. 1. Matrix of the subsystems interoperabilitys impact on the SoS analysis perspec-

tives (S: Stable, I: Increase, D: Decrease).

4 Conclusion and Perspectives

This report illustrated the importance of the interoperability of the subsystems in the control and evolution of the System of Systems. After careful analysis, we re- alized that there exists a strong linkage between the interoperability and the SoS characteristics and between the interoperability and some other non-functional characteristics of the SoS (analysis perspectives). The significance of this report lies in its ability awareness about the need to consider the interoperability prior the assembling of the subsystems. An impact matrix of the interoperability on the analysis perspectives has been proposed, it is a first crucial step towards an effective System of Systems Engineering. The proposed matrix will serve to allow the evaluation of the impact of the interoperability on the analysis perspectives which will be achieved through the simulation.

References

• 1. Boardman, J., Sauser, B.:

System of Systems - the meaning of of. IEEE/SMC International Conference on System of Systems Engineering (2006) 118–123

• 2. Jamshidi, M.: System of systems engineering: innovations for the twenty-first cen-
• tury. John Wiley & Sons (2011)
• 3. Maier, M.W.: Architecting principles for systems-of-systems. Systems Engineering

1(4) (1998) 267–284

• 4. Bilal, M., Daclin, N., Chapurlat, V.:

Collaborative Networked Organizations as System of Systems: a model-based engineering approach. IFIP AICT, pro-ve (2014)

• 5. Camarinha-matos, L.M.: Collaborative networks: A mechanism for enterprise agility

and resilience. Enterprise Interoperability VI (2014) 1–8

• 6. Ford, T., Colombi, J., Graham, S., Jacques, D.: A Survey on Interoperability Mea-
• surement. Twelfth International Command and Control Research and Technology

Symposium (937) (2007)

• 7. Stevens Intitute Of Technology, Castle Point On Hudson, Hoboken, N..: Report On

System Of Systems Engineering. (2006)

• 8. Billaud, S., Daclin, N., Chapurlat, N.: Interoperability as a key concept for the

control and evolution of the System of Systems (SoS). Enterprise Interoperability (2015) 53–63

I = {i1, . . . , in} wj j = 1, . . . , n C UB UB − 1

Page 46

hard28 m m − 2 m m − 2 m TC B = {b1, b2, . . . , bm−2} IB ⊆ I B Ib b ∈ B ITC TC B TC b ∈ B Swap(p, q) p b ∈ B q TC Pack(b) b ∈ B TC b P ⊆ Ib ∪ITC Pack Swap Swap(p, q) p q

Swap Swap(p, q) (p, q) ∈ {(0, 1), (1, 1), (2, 1), (1, 2), (2, 2), (2, 3), (3, 2)} Swap(0, 1) shift b ∈ B (p, q) ∈ {(0, 1), (1, 1), (2, 1), (1, 2), (2, 2)} Swap(p, q) B O(l × C) l w(TC) 2C Pack C1 C2 w1(TC) w2(TC)

w1(TC) C1 w2(TC) C2

Page 48

m m + 2 28 1

A Framework for Online Inter-Subjects Classification in motor imagery-based Brain-Computer Interfaces

Sami Dalhoumi, Gérard Dray, Jacky Montmain

1 Parc Scientifique G. Besse, 30035 Nîmes, France.

!"#$%&'()"#$%*+$,"%'-#.%'&/)0

• Abstract. Inter-subjects classification and online adaptation techniques have

been actively explored in the brain-computer interfaces (BCIs) research com- munity during the last years. However, few works tried to conceive classifica- tion models that take advantage of both techniques. In this thesis, we propose an

• nline inter-subjects classification framework for motor imagery-based BCIs.

Inter-subjects classification is performed using a weighted average ensemble in which base classifiers are learned using data recorded from different subjects and weighted according to their accuracies in classifying brain signals of target

• subject. Online adaptation is performed in

a semi-supervised way. The effectiveness of this framework is demonstrated us- ing publically available electroencephalography (EEG) data sets. Keywords: Brain-Computer Interfaces, Inter-Subjects Classification, Online Adaptation, Weighted Average Ensembles.

1 Introduction

A brain-computer interface (BCI) is a communication and control technology that allows or hemodynamic activity patterns into commands for an external device [1]. This technology was originally meant to allow patients with severe neuromuscular disabilities to autonomously interact with their environ-

• ment. Depending on the modality of interaction, BCIs can be classified as either ex-
• genous or endogenous [2]. Exogenous BCIs rely on brain activity patterns that are

elicited spontaneously in response to external stimuli, while endogenous BCIs are based on the voluntary induction of different brain states by the user. One popular paradigm for voluntarily inducing different brain states is motor imagery, i.e., the imagination of the execution of a movement with a particular limb such as the left hand or the right hand. Motor imagery (MI)-based BCIs can offer a natural way of interaction for the user but they are difficult to set-up because self-regulation of brain rhythms is not a straightforward task. For this reason, a long calibration phase is need- ful for user-system co-adaptation before every use of the BCI. During this phase, the user interacts with the BCI in a cue-based mode which allows him to learn self- regulating his brain rhythms and the system to create a classification model.

Page 50

The accuracy of the system depends on the capacity of the classification model to decode brain activity patterns of the user during a feedback phase (self-paced interac- tion mode). In order to bring MI-based BCIs out of the lab, many research groups have focused

• n conceiving new machine learning techniques that allow reducing calibration time

without decreasing classification accuracy of the system. Among these techniques, inter-subjects classification [3-5] and online adaptation [6-7] of have been actively explored during the last years. Inter-subjects classification consists of incorporating labeled data recorded from different BCI users in the learning process of a new user. When performed correctly, inter-subjects classification allows capturing information that generalize across users and extend to new users. Online adaptation consists of updating parameters of the prediction model during self-paced interaction mode. Since class labels are not provided, different online adaptation approaches have been investigated in order to increase classification performance of MI-based BCIs. Ap- proaches based on error-related potentials (ErrPs) are the most promising ones. These potentials are elicited in response to a feedback that contradicts users intention and have been successfully used in MI-based BCIs. Combining inter-subjects classification and online adaptation techniques may al- low bringing MI-based BCIs technology out of the lab. Inter-subjects classification allows creating classification models that rapidly achieve high performance without the need of long calibration phase. Online adaptation ensures optimality of the predic- tion model. Few approaches have been attempted in this direction. Vidaurre et al., [3] proposed a machine learning framework in which subject-independent spatial filters and corresponding classifier were initially learned using manually selected EEG sig- nals from a large motor imagery data set of different subjects and then updated using data from a new subject. Although it was effectively used in a MI-based BCI, this framework uses a naïve approach for creating the subject-independent classification

• model. Two ensemble frameworks based on dynamic classifiers weighting were pro-

posed in [4] and [5], respectively. Dynamic classifiers weighting is usually used for combining predictions of multiple learners in non-stationary environments but it may perform poorly because it does not take into consideration the stochastic dependence between time-contingent feature vectors.

2 Contribution

The baseline of the proposed framework is an accuracy-weighted ensemble (AWE) in which base classifiers are learned using data from different subjects and weighted using a small calibration set recorded from a new BCI user. This static ensemble may be efficient for performing inter-subjects classification when the contributions of base classifiers to the ensemble are accurately weighted. But because of non-stationarity of EEG signals between calibration and feedback phases and during feedback phase, accuracies of base classifiers in predicting class labels of EEG signals recorded from the new subject may change and consequently their contributions to the ensemble

must be reconsidered. Since true class labels are not provided during feedback phase, we reinforced by error-related potentials to update base classifiers weights in online fashion. n ensemble classification model was proposed in [8] for online fMRI data classification. Because fMRI data suffers from a very high feature to instance ratio, the author used a random subspace ensemto differen This apoutperformed a static random subspace en- semble and an adaptive random subspace ensemble in which base classifiers are up- dated using their own predictions (naïve labeling). However, the main drawback of this approach is that base classifiers may converge to the same prediction model which jeopardizes the diversity of the ensemble and consequently its accuracy. In our case, the situation is different. Spatial filters and base classifiers are sup- robustare learned using enough labeled data from other subjects. However, their weights are subject to uncer- tainty because they are learned using a small labeled set from the new subject. So these weights should be updated rather than base classifiers. This allows tracking non- stationarity within the same session and does not compromise the diversity of the ensemble since base classifiers are unchanged. As the predictions of the ensemble are weights may lead to error accumulation and consequently degrades the accuracy of the BCI. If an additional source of information is available, it should be used to mini- mize uncertainty. In BCI applications, such information could come from error- related potentials (ErrPs). Fig. 1. gives an illustration of the proposed framework.

• Fig. 1. Illustration of the proposed online inter-subjects classification framework for MI-

based BCIs Page 52

3 Results

The proposed framework was evaluated using two publically available EEG data sets. Preliminary results were published in a peer-reviewed international conference in biomedical engineering [9]. Results of a thorough study are submitted to a peer- reviewed international conference in computer science [10].

4 Conclusion

In this work, we highlighted the need of conceiving universal classification models that combine different machine learning techniques such as knowledge transfer and

• nline adaptation of prediction models in order to create every-day-use BCI systems.

In this direction, we proposed an online inter-subjects classification framework for motor imagery (MI)-based BCIs and we assessed its performance using real EEG data sets.

References

• 1. McFarland, D.J., Wolpaw, J.R.: Brain-computer interfaces for communication and control.

Communications of the ACM, vol. 54, issue 5, pp. 60-66 (2011)

• 2. Nicolas-Alonso, L.F., Gomez-Gil, J. :Brain Computer Interfaces, a Review. Sensors, vol.

12, pp. 12111279 (2012)

• 3. Vidaurre, C., Sannelli, C., Muller, K.R., Blankertz, B.: Machine-learning based co-

adaptive calibration: a perspective to fight BCI illiteracy. Lecture Notes in Computer Sci- ence Volume 6076, 2010, pp 413-420 (2010).

• 4. Tu, W., Sun, S.: A subject transfer framework for EEG classification. Neurocomputing,
• vol. 82, pp. 109--116 (2011)
• 5. Liyanage, S.R., Guan, C., Zhan, H., Ang, K.K., Xu, J., Lee, T.H.: Dynamically weighted

ensemble classification for non-stationary EEG processing. Journal of Neural Engineering,

• vol. 10, no. 3, p. 036007 (2013 )
• 6. Llera, A., Van Gerven, M.A.J., Gomez, V., Jensen, O., Kappen, H.J.: On the use of in-

teraction error potentials for adaptive brain computer interfaces. Neural Net- works, vol. 24, no 10, pp. 11201127 (2011)

• 7. Zeyl, T.J., Chau, T.: A case study of linear classifiers adapted using imperfect la-

bels derived from human event-related potentials. Pattern Recognition Letters,

• vol. 37, pp. 5462 (2014)
• 8. Plumpton, C.O.: Semi-supervised ensemble update strategies for on-line classifi-

cation of fMRI data. Pattern Recognition Letters, vol. 37, pp. 172-177 (2014)

• 9. Dalhoumi, S., Dray, G., Montmain, J., Derosière, G., Perrey, S.: An adaptive ac-

curacy-weighted ensemble for inter-subjects classification in brain-computer in-

• terfacing. In: 7th International IEEE EMBS Neural Engineering Conference

(2015)

• 10. Dalhoumi, S., Dray, G., Montmain, J., Perrey, S.: A framework for online inter-

subjects classification in endogenous brain-computer interfaces. In 22nd Interna- tional Conference on Neural Information Processing (2015) (submitted).

USI at BioASQ 2015: a semantic similarity-based approach for semantic indexing

Nicolas Fiorini1,∗, Jacky Montmain1, Sylvie Ranwez1, Vincent Ranwez2

1 LGI2P research center from the ´

Ecole des mines d’Al` es, Site de Nˆ ımes, Parc scientifique G. Besse, 30035 Nˆ ımes cedex 1, France {name.surname}@mines-ales.fr

2 UMR AGAP, Montpellier SupAgro/CIRAD/INRA, 2 place Pierre Viala,

Montpellier Cedex 1 34060, France ranwez@supagro.inra.fr

• Abstract. The need of indexing biomedical papers with the MeSH is

incessantly growing and automated approaches are constantly evolving. Since 2013, the BioASQ challenge has been promoting those evolutions by proposing datasets and evaluation metrics. In this paper, we present

• ur system, USI, and how we adapted it to participate to this challenge

this year. USI is a generic approach, which means it does not directly take into account the content of the document to annotate. The results lead us to the conclusion that methods that solely rely on semantic annotations available in the corpus can already perform well compared to NLP-based approaches as our results always figure in the top ones. Keywords: MeSH indexing, semantic similarity, kNN approach

1 Introduction

The task 3a of the BioASQ challenge consists in indexing new biomedical papers with MeSH concepts. The need of document indexed by terms of a thesaurus like the MeSH has already been emphasized several times in Information Retrieval [1]. This task has been historically done by experts, later helped by automated

• methods. Nowadays, this task gets more and more importance with the increas-

ing number of papers to annotate. BioASQ aims at evaluating the indexing systems on two criteria: effectiveness and efficiency. The first one is assessed by using common metrics in text classifi- cation: the F-measure and the LCA-F [2]. Although the speed of the systems is not directly measured, when a test set is released, the participants have a short amount of time — 21 hours — to send their results. The litterature shows that several investigations have been made to accom- plish this task. Some rely on concept extraction based on the text. This means that the system does some Natural Language Processing (NLP) and aims at finding lemmas in the text that can be mapped on the thesaurus [3, 4]. Those methods can be enriched with other processes such as Machine Learning (ML). It has been extensively used during the last years and many approaches have been

Page 54

tested: learning-to-rank [5], gradient boosting [6], or reflective random indexing [7]. So far, hybrid methods — those based on both NLP and ML — produced the best result [8]. We previously presented USI (User-oriented Semantic Indexer), a method that does not do any NLP nor ML [9]. This paper briefly introduces USI and shows how we adapted and tested it for this challenge. We also investigate on how generic methods may or may not contribute to the future of semantic indexing.

2 Method

USI is a generic method that aims at indexing entities of any type — be it text files, audio or video samples, genes, etc. — as long as there is an existing collection of entities that are already annotated. It is based on a k Nearest Neighbors (kNN) approach and it proceeds in two steps: – identifying the neighbors, i.e documents in the corpus that are close to the

• ne to annotate,

– using this neighborhood for annotating this document. This kind of approach has already been used in the litterature [5, 8] as part

• f a bigger process mixing NLP and ML. The aim of USI is to get rid of the

content processing so that it can be applied on any document. When the neighborhood has been set, USI selects the set of concepts in the annotations of the neighbors that is the most semantically similar with every

• ther neighbor. Formally, for a given document to annotate, let us define its

neighbors as K with k = |K|. K contains documents d that are respectively annotated by a set of concepts Ad each. The list of all those annotations for the neighborhood forms a family of set AK. In order not to look for the solution in the whole thesaurus, we define a search space A0 such that A0 = [

Ad∈Ak

Ad. (1) An optimal solution A∗ ⊆ A0 has to be found. To do so, USI follows an

• bjective function defined as:

A∗ = argmax

A⊆A0

• f(A)

, f(A) = 1 k X

Ad∈AK

sim(A, Ad) − µ|A|, (2) with µ ∈ [0; 1]. sim(A, Ad) represents the semantic similarity of A with Ad. This semantic similarity can be any one of the numerous semantic similarity mea- sures of the litterature. We chose to use an indirect groupwise measure called Best Match Average (BMA) [10], a composite average based on pairwise similar- ity values that can be efficiently updated when removing a single concept from the testing solution A. µ is a concision constraint and represents the decrease of the similarity with the neighbors that is allowed for the removal of one concept

• f A. For the challenge the value of µ has been empirically optimized on the

training sets (µ = 0.01). The BioASQ challenge is a fantastic opportunity for us to test USI with several variations in a real case context. Indeed, USI has parameters for which we have no or poor idea of the value they should take. They are (i) the size

• f the neighborhood, (ii) the semantic similarity measure and an associated IC

metric, and (iii) the impact of integrating baseline results. We thus propose several systems presented in table 1.

System name Description USI 10 neighbors Default version of USI where 10 neighbors are selected USI 20 neighbors Default version of USI where 20 neighbors are selected USI abstract “USI 10 neighbors” where semantic similarity is chosen using an abstract framework USI baseline “USI abstract” integrating the provided baselines

Table 1: Description of the systems submitted to BioASQ 2015.

3 Results

Once a test set has been published, each paper of this set is manually annotated by experts to constitute the gold standard annotation. This takes a long time and results are not final as of June 16th, 2015. The table 2 shows the — provisional — best results of all our systems com- pared to the baselines on batch 1 week 2 test set. The main evaluation metrics are a flat measure (micro F-measure, or MiF) and a hierarchical one (LCA-F) [14]. It appears that the most elaborated version of USI ranks first among the

• thers. It also performs better than all the baselines.

System name MiF MiR MiP LCA-F LCA-R LCA-P USI baseline 0.5624 0.5417 0.5847 0.4677 0.4765 0.4895 Default MTI 0.5618 0.5460 0.5786 0.4788 0.4754 0.5186 MTIFL 0.5552 0.5008 0.5230 0.4672 0.4391 0.5377 USI abstract 0.5159 0.4951 0.5384 0.4416 0.4364 0.4882 USI 10 neighbors 0.5114 0.4883 0.5368 0.4430 0.4375 0.4864 USI 20 neighbors 0.4846 0.4725 0.4973 0.4450 0.4457 0.4820 MeSH Now BF 0.4361 0.4332 0.4391 0.3817 0.3987 0.3992 BioASQ baseline 0.1083 0.0865 0.1451 0.1541 0.1398 0.2850 Table 2: Results obtained (as of June 16th, 2015) on batch 1 week 2 with all USI sys- tems compared to the baselines. Bold values represent the best scores among these systems. Page 56

4 Conclusions

Every year, the BioASQ challenge reveals novel and powerful methods. The most elaborated version of USI, USI baseline, ranked in the top results of all test sets. However, this approach does not directly take document contents into account nor does any NLP. The results show that our generic method can be applied to specific use cases with a bit of adaptations. Given an information retrieval system, USI can be an easy-to-set-up solution for annotating document with terms from a thesaurus.

References

• 1. Lu, Z., Kim, W., Wilbur, W.J.: Evaluation of query expansion using MeSH in
• PubMed. Information retrieval 12, 69-80 (2009)
• 2. Kosmopoulos, A., Partalas, I., Gaussier, E., Paliouras, G., Androutsopoulos, I.: Eval-

uation measures for hierarchical classification: a unified view and novel approaches. CoRR abs/1306.6802, (2013)

• 3. Jonquet, C., Shah, N.H., Musen, M. A.: The open biomedical annotator. Summit
• n Translational Bioinformatics, 2009, 56 (2009)
• 4. Aronson, A.R., Lang, F.-M.: An overview of MetaMap: historical perspective and

recent advances. Journal of the American Medical Informatics Association 17, 229- 236 (2010)

• 5. Huang, M., N´

ev´ eol, A., Lu, Z.: Recommending MeSH terms for annotating biomed- ical articles. Journal of the American Medical Informatics Association 18, 660-667 (2011)

• 6. Delbecque, T., Zweigenbaum, P.: Using co-authoring and cross-referencing informa-

tion for MEDLINE indexing. AMIA Annual Symposium Proceedings, 147 (2010)

• 7. Vasuki, V., Cohen, T.: Reflective random indexing for semi-automatic indexing of

the biomedical literature. Journal of Biomedical Informatics, 43(5), 694–700 (2010)

• 8. Mao, Y., Wei, C.H., Lu, Z.: NCBI at the 2014 BioASQ challenge task: large-scale

biomedical semantic indexing and question answering. Working Notes for CLEF 2014 Conference, 1180, 1319-1327 (2014)

• 9. Fiorini, N., Ranwez, S., Montmain, J., Ranwez, V.: USI: a fast and accurate approach

for conceptual document annotation. BMC Bioinformatics, 16, 83 (2015)

• 10. Schlicker, A., Domingues, F.S., Rahnenf¨

uhrer, J., Lengauer, T.: A new measure for functional similarity of gene products based on Gene Ontology. BMC Bioinformatics, 7, 302 (2006)

• 11. Sayers, E.: E-utilities Quick Start. Entrez Programming Utilities Help. (2010)
• 12. Harispe, S., Sanchez, D., Ranwez, S., Janaqi, S., Montmain, J.: A framework for

unifying ontology-based semantic similarity measures: A study in the biomedical

• domain. Journal of biomedical informatics, 48, 38-53 (2014)
• 13. Harispe, S., Ranwez, S., Janaqi, S., Montmain, J.: The Semantic Measures Li-

brary and Toolkit: fast computation of semantic similarity and relatedness using biomedical ontologies. Bioinformatics, 30(5), 740-742 (2014)

• 14. Tsatsaronis, G., Balikas, G., Malakasiotis, P., Partalas, I., Zschunke, M., Alvers,

M.R., Weissenborn, D., Krithara, A., Petridis, S., Polychronopoulos, D., Almirantis, Y., Pavlopoulos, J., Baskiotis, N., Gallinari, P., Arti´ eres, T., Ngonga Ngomo, A.- C., Heino, N., Gaussier, E., Barrio-Alvers, L., Schroeder, M., Androutsopoulos, I., Paliouras, G.: An overview of the BIOASQ large-scale biomedical semantic indexing and question answering competition. BMC bioinformatics, 16, 138 (2015)

Automating the evolution management of three-level software architectures

Abderrahman Mokni+, Marianne Huchard*, Christelle Urtado+, Sylvain Vauttier+, and Huaxi (Yulin) Zhang‡

+LGI2P, Ecole Nationale Sup´

erieure des Mˆ ınes Al` es, Nˆ ımes, France *LIRMM, CNRS and Universit´ e de Montpellier 2, Montpellier, France

‡ Laboratoire MIS, Universit´

e de Picardie Jules Verne, Amiens, France {Abderrahman.Mokni, Christelle.Urtado, Sylvain.Vauttier}@mines-ales.fr, huchard@lirmm.fr, yulin.zhang@u-picardie.fr

1 Introduction

Software evolution [1] is becoming more and more challenging due to the in- creasing complexity of software systems and their importance in everyday life. While component reuse has become crucial to shorten large-scale software sys- tems development time, handling evolution in such systems is a serious issue. As witnessed by Garlan et al. in their recent study [2], the difficulty of reuse lies essentially on architectural mismatches that arise due to several changes that affect software. A famous problem is software architecture erosion [3,4]. It arises when modifications of the implementation of a software violate the design prin- ciples captured by its specification architecture. Such erosion leads to software degradation and shortens its lifetime. Increasing confidence in reuse-centered, component-based software systems lies on resorting out multiple issues. First, software architectures must support change at any step of component- based development to meet new user needs, improve component quality, or cope with component failure. Second, the change impact must be handled locally (at the same abstraction level) to avoid architecture inconsistencies and propagated to the other abstraction levels to avoid incoherence between the architecture descriptions, notably erosion. Third, the evolution activity must be tracked to enable monitoring, commitment and/or rollback and versioning.

2 Overview of Dedal

Dedal [5] is a novel architectural model and ADL that covers the three main steps of component-based development by reuse: the specification, the imple- mentation and deployment. The idea of Dedal is to build a concrete software ar- chitecture (called architecture configuration) from suitable software components stored in indexed repositories. Candidate components are selected according to an intended architecture (called architecture specification) that represents an ab- stract and ideal view of the software. The implemented architecture can then be

Page 58

instantiated (the instantiation is called architecture assembly) and deployed in multiple contexts. Figure 1 illustrates the three abstraction levels of Dedal on an example of Home Automation Software. The software enables to manage the light of build- ings in function of the time through an orchestrator (component role Home- Orchestrator). The specified functionalities are turning on/off the light (compo- nent role Light), controlling its intensity (component role Intensity) and getting information about the time (component role Time).

• Fig. 1. Example of a Dedal model: Home Automation Software

3 Summary of ongoing research

3.1 Dedal formal basis Dedal formalization is crucial to enable the verification and validation of the derived architectural models as well as evolution management. In [6], we propose a formalization of Dedal that comprises two kinds of typing rules. The first kind is about intra-level rules. These rules define the relations between components

• f the same abstraction level such as compatibility and substitutability. The

second kind is about inter-level rules. These rules define the relations between components of different abstraction levels (cf. Figure 1 for inter-level relations). For instance, the realization rule checks whether a (or a set of) component class(es) realizes a (or a set of) component role(s). This rule is also used to search for candidate concrete components in repositories to implement a specified software architecture. Inter-level and intra-level rules are generalized to support the architectural level. First, using intra-level rules, we can check architecture

consistency at any abstraction level. Second, using inter-level rules, it is possible to check coherence between architecture descriptions at different abstraction

• levels. For instance, we can check whether a configuration implements all the

desired functionalities documented in its specification. In [7], we propose a set

• f evolution rules that enable to evolve Dedal models. The proposed rules allow

the manipulation (addition, deletion and substitution) of architectural elements (e.g. components, connections) at the three Dedal abstraction levels. 3.2 An evolution mangement model for Dedal In [8], we proposed an evolution management model to deal with change in three- level software architectures. The evolution management model encompasses all the elements involved in the evolution management activity. The interest of the evolution management model is twofold. First, it simplifies the evolution management task by separating the evolution management concern from the architectural modeling concern. Second, it helps understanding and analyzing the architectural change by representing changes as first class entities at the same level as the architectural model. The evolution management model consists

• f three main parts:

– The architectural model: is the target of change. Since the objective is to manage evolution throughout all the stages of component-based develop- ment, the target is architectural models derived from Dedal. – The architectural change: gathers all the necessary information about the change such as its origin and abstraction level. – The evolution manager: is the system that captures the change request and tries to find a solution to evolve the architectural model with respect of all the pre-defined conditions. 3.3 The formal evolution approach The evolution management approach has the following objectives: – To capture architectural change at any of the three main stages of component- based software lifecycle (i.e. specification, implementation, runtime). – To control the impact of change where it is initiated by reestablishing the architecture consistency if altered. – To propagate the change to the other abstraction levels in order to restore the global coherence of the architecture descriptions if altered. Evolution management starts when a change is requested somewhere in the architectural model (For instance, a component class addition is requested in the configuration architecture). Evolution manager captures change and initiates it by running adequate evolution rules on the corresponding formal model. It then runs a resolution algorithm that attempts to find a sequence of evolution rules leading to a consistent state of the architecture. The same algorithm is applied

Page 60

to the other levels to restore global coherence if it is necessary. If a solution is found, the evolution manager generates the corresponding evolution plan that could then be committed by user. In the other case (i.e. failure), the evolution manager rejects the change request. 3.4 Tooling: DedalStudio We developed an eclipse-based tool called DedalStudio to support evolution man- agement in Dedal. DedalStudio is equipped with a graphical modeling environ- ment that supports the three Dedal architecture representations. These repre- sentations are translated into B formal models when a change is requested. The resulted formal models enables the analysis of architectural change and the gen- eration of evolution plans through the evolution manager system.

4 Conclusion and future work

We proposed a formal approach to automate evolution management in Dedal. This approach contributes to component-based software engineering by support- ing the evolution thorough the whole component-based development process. As a future work, we aim to improve the performance of the evolution manager and test it on real case studies.

References

• 1. Mens, T., Demeyer, S.: Software Evolution. Springer (2008)
• 2. Garlan, D., Allen, R., Ockerbloom, J.: Architectural mismatch: Why reuse is still

so hard. IEEE Software 26(4) (July 2009) 66–69

• 3. Perry, D.E., Wolf, A.L.: Foundations for the study of software architecture. SIG-

SOFT Software Engineering Notes 17(4) (October 1992) 40–52

• 4. de Silva, L., Balasubramaniam, D.:

Controlling software architecture erosion: A

• survey. JSS 85(1) (January 2012) 132–151
• 5. Zhang, H.Y., Zhang, L., Urtado, C., Vauttier, S., Huchard, M.: A three-level compo-

nent model in component-based software development. In: Proc. of the 11th GPCE Conf., Dresden, Germany, ACM (Sept. 2012) 70–79

• 6. Mokni, A., Huchard, M., Urtado, C., Vauttier, S., Zhang, H.Y.: Towards automating

the coherence verification of multi-level architecture descriptions. In: Proc. of the 9th ICSEA, Nice, France (Oct. 2014) 416–421

• 7. Mokni, A., Huchard, M., Urtado, C., Vauttier, S., Zhang, H.Y.: Formal rules for

reliable component-based architecture evolution. In: Formal Aspects of Component Software - 11th International FACS Symposium revised selected papers, Bertinoro, Italy (Sept. 2014) 127–142

• 8. Mokni, A., Huchard, M., Urtado, C., Vauttier, S., Zhang, H.Y.: An evolution man-

agement model for multi-level component-based software architectures. In: To ap- pear in proc. of the 27th SEKE, Pittsburgh, USA (June. 2015)

Intermediate Report on RSD-HoG (A new image descriptor)

Darshan Venkatrayappa, Philippe Montesinos, and Daniel Diep

Ecole des Mines d’Ales, LGI2P, Parc Scientifique Georges Besses, 30035 Nimes, France {darshan.venkatrayappa,philippe.montesinos,daniel.diep}@mines-ales.fr

• Abstract. In this report we propose a novel local image descriptor called

RSD-HoG. For each pixel in a given support region around a key-point, we extract the rotation signal descriptor(RSD) by spinning a filter made

• f oriented anisotropic half-gaussian derivative convolution kernel. The
• btained signal has extremums at different orientations of the filter. These

characteristics are combined with a HoG [1] technique, to obtain a novel descriptor RSD-HoG. For evaluation, we have used the standard Oxford data set. Extensive experiments on the images in this dataset demon- strates that our approach performs better than many state of the art de- scriptors. Keywords: Image descriptor, Half Gaussian kernel, Feature matching, Rotation signal descriptor, HoG.

1 Introduction

In computer vision, problems related to object matching, tracking, panorama generation, image classification, structure and motion estimation are effectively addressed by the popular approach of image representation by a set of local image descriptors. The main purpose of local image descriptor is to capture the geometry of a support region around a key-point. In addition to this, the image descriptor should be invariant to certain image transformations such as rotation, brightness, blurring and scale changes. Scanning through the computer vision literature, one can come across the term image matching pipeline. The image matching pipeline has four stages. In the first stage, key-points or regions are selected using the popular detectors such as LoG [2], DoG [3] or Harris Affine [5]. This is followed by the extraction of features or feature description from the support region around the key-point. Next, various post processing steps such as normalization [3], quantization [6] and dimensionality reduction [4] is applied. The final block involves matching the descriptor using different distance measures such as euclidean distance [3].

Page 62

2 Darshan Venkatrayappa, Philippe Montesinos, and Daniel Diep

! ! ! ! ! ! "

" " " " "

(a) (b)

!"#"$ !"#"% !"#"& !"#"'

"

"#"' "#"& "#"% "#"$ " %("

' &

)*+,-./*0.123456.2222271.5-..+8 9:36*/;2<:4,/*=4

6 6 6

>" '?" &@"

6

(c)

• Fig. 1. (a) A thin rotating Gaussian derivative half-filter. (b) Half Gaussian kernel

applied to a keypoint xp, yp. (c) Extrema of a function Q(xp, yp, θ). µ = 10, λ = 1 and ∆θ = 2o. Note that the initial orientation of the filter is vertical, upwardly directed and steerable clockwise.

2 Methodology

2.1 Image-processing As shown in Fig 1(a). at pixel (x, y), a derivative kernel is applied to obtain a derivative information ϕ(x, y, θ) in a function of orientation θ ⇤ [0; 360⇥[ : ϕ(x, y, θ) = Iθ ⇥ C · H(y) · x · e

• 

 x2

2λ2 + y2 2µ2

 

(1) Where Iθ corresponds to a rotated image of orientation θ, C is a normalization coefficient, (x, y) are pixel coordinates, and (µ, λ) the standard deviation of the anisotropic Gaussian kernel. Only the causal part of this filter along the Y axis is used. This is obtained by cutting the kernel in the middle, in an operation that corresponds to the Heaviside function H. As in [9] we have chosen to rotate the image instead of the filter, there by reducing the algorithmic complexity by making use of a recursive Gaussian filter [10]. 2.2 Rotational Signal Descriptor(RSD) RSD is obtained by rotating the above described filter around a key-point. Fig. 1(c) shows a sample RSD obtained by applying the Gaussian derivative half filter at the the pixel location (xp, yp) in steps of 2⇥. We compute the gradient ⌅⇧I⌅ and the two angles for the key-point/pixel at location (xp, yp) by considering the global extrema of the function Q(xp, yp, θ). The two angles θ1 and θ2 define a curve crossing the pixel(an incoming and outgoing direction). Two of these global extrema are combined to maximize ⌅⇧I⌅, i.e :

Intermediate Report on RSD-HoG (A new image descriptor) 3

• Fig. 2. Affine region is normalized to a square of size 41x41. On each pixel of the patch,

RSD is generated and the two angles θ1 and θ2 are extracted. These angles are binned separately.

         ⇥⇤I⇥ = max

θ∈[0,360[Q(xp, yp, θ)

min

θ∈[0,360[Q(xp, yp, θ)

θ1 = arg max

θ∈[0,360[

(Q(xp, yp, θ)) θ2 = arg min

θ∈[0,360[

(Q(xp, yp, θ)) (2) 2.3 RSD-HoG construction Prior to feature extraction, the image is smoothed with a Gaussian filter. The framework of the RSD-HOG extraction is illustrated in Fig. 2. On each pixel of the normalized patch, we apply the rotating semi Gaussian filter to obtain the

• RSD. From this RSD, we extract two angles and a magnitude for each pixel as

explained above. Then we bin the two angles separately as in Eq.3 and Eq.4. The image patch is divided in to 16 blocks. Since the image patch is of size 41x41, most of the blocks are of size 10x10 ( blocks on the extreme right and extreme bottom are of size 11x11 ). Each of these block contributes 8 bins to the final descriptor. We fuse the two intermediate descriptors to form the final descriptor as in Eq.5. The intermediate descriptors in Eq.3 and Eq.4 alone can be used as descriptors. But, fusing these two descriptors as in Eq.5 results in a more robust description. The performance of two intermediate descriptor and the final descriptor for the boat dataset(Rotation changes) can be seen in the first row of Fig.3. RSD HoG Theta1 = {θ1bin1, θ1bin2, θ1bin3, θ1bin4.....θ1bin128} (3) RSD HoG Theta2 = {θ2bin1, θ2bin2, θ2bin3, θ2bin4.....θ2bin128} (4) RSD HoG = {θ1bin1, θ1bin2, ...θ1bin128, θ2bin1, θ2bin2, ...θ2bin128} (5)

Page 64

4 Darshan Venkatrayappa, Philippe Montesinos, and Daniel Diep

3 Experiments, Discussions and Results

We evaluate and compare the performance of our descriptor against the state of the art descriptors on the standard dataset using the standard protocol provided by Oxford group. Recall versus precision curves as proposed by [11] are used to evaluate our descriptor. Our descriptor has 4 different parameters. The rotation step (∆θ) of the filter is fixed to 5. Increasing the rotation step results in loss

• f information. We have fixed the number of bins to 8 per block, resulting in a

8∗16 = 128 bins for 16 blocks. Increasing the number of bins results in the same performance but, increases the dimensionality of the descriptor. Filter Height (µ) is fixed to 6. Filter Width (λ) is fixed to 1. All the parameters are chosen empirically. The performance of RSD-HoG is compared against SIFT-OXFORD [3], SIFT- PATCH, GLOH [7], DAISY [8] and PCA-SIFT. First, we compare the perfor- mance of RSD − HoG (Eq.5) with the intermediate descriptors RSD − HOG − Theta1 (Eq.3)and RSD − HOG − Theta2 (Eq.4). From the first two graphs in the first row of Fig.3 it is clear that RSD−HoG performs better than the two in- termediate descriptors. This is same for all the images with different transforma-

• tions. Due to lack of space we restrict ourselves to graphs for rotational changes.

For rotation changes, when RSD − HOG is compared with other descriptors it can be seen that our descriptor performs better than other descriptors. This can be seen from graphs (c) and (d) of Fig.3. For variations in viewpoint (graphs (e) and (f)), it can be seen that RSD − HoG performs similar to other descriptors. Last two graphs in the second row of Fig.3 represent Recall vs Precision curves for the Blur changes. From the graphs it is clear that RSD − HoG outperforms other descriptors. Based on graphs in the third row of Fig.3, we conclude that the performance of RSD − HoG is superior to the performance of other descriptors when it comes to handling variations in brightness and compression.

4 Conclusion

This paper proposes a new image descriptor called RSD-HoG. On the standard dataset provided by the Oxford group, RSD-HoG outperforms other state of the art descriptors. Currently, high complexity and the dimension of the descriptor are a major drawback. In the future we would like to reduce both the complexity and dimension of our descriptor. In the future, we would like to experiment with other variations of the anisotropic half Gaussian kernel. We would also like to focus on the real time implementation of our descriptor using parallel programming techniques.

References

• 1. Dalal, N., Triggs, T.: Histograms of Oriented Gradients for Human Detection.

In:Computer Vision and Pattern Recognition(CVPR), pp. 886-893. (2005)

Intermediate Report on RSD-HoG (A new image descriptor) 5

0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 1 − precision Recall RSD−HOG−Theta1 RSD−HOG−Theta2 RSD−HOG 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 0.1 0.2 0.3 0.4 0.5 0.6 0.7 1 − precision Recall RSD−HOG−Theta1 RSD−HOG−Theta2 RSD−HOG

• (a) boat 1-3

(b) boat 1-4 (c) boat 1-3 (d) boat 1-4

• 0.1

0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 1 − precision Recall RSD−HOG−INTERLACE DAISY GLOH PCA SIFT−OXFORD SIFT−PATCH

• (e) Graff 1-3

(f) Graff 1-4 (g) BIKE 1-3 (h) BIKE 1-4

• (i) Leuven 1-3

(j) Leuven 1-4 (k) Compression 1-3 (l) Compression 1-4

• Fig. 3. Recall vs Precision curves for Oxford dataset using nearest neighbour matching

method

• 2. Lindeberg, T.: Feature Detection with Automatic Scale Selection. In:International

Journal of Computer Vision(IJCV), vol. 30, pp. 79-116. (1998)

• 3. Lowe,

D.G.:Distinctive Image Features from Scale-Invariant Keypoints. In:International Journal of Computer Vision(IJCV), vol. 60, pp. 91-110. (2004)

• 4. Ke, Y., Sukthankar, R.: PCA-SIFT:A More Distinctive Representation for Local

Image Descriptors. In:Computer Vision and Pattern Recognition(CVPR), pp. 506-

• 513. IEEE Computer Society, (2004)
• 5. Mikolajczyk, K., Schmid, C.: Scale & Affine Invariant Interest Point Detectors.

In:International Journal of Computer Vision(IJCV), vol. 60, pp. 63-86. (2004)

• 6. Chandrasekhar, V., Takacs, G., Chen, D.M., Tsai, S.S., Reznik, Y.A., Grzeszczuk,

R., Girod, R.: Compressed Histogram of Gradients: A Low-Bitrate Descriptor. In:International Journal of Computer Vision(IJCV), vol. 96, pp. 384-399. (2012)

• 7. Mikolajczyk, M., Schmid, C.: A performance evaluation of local descriptors. In:IEEE
• Trans. Pattern Anal. Mach. Intell(PAMI), vol. 27, pp. 1615-1630. (2005)
• 8. Tola, E., Lepetit, V., Fua, P.: DAISY: An Efficient Dense Descriptor Applied to

Wide Baseline Stereo. In:IEEE Trans. Pattern Anal. Mach. Intell(PAMI), vol. 32,

• pp. 815-830. (2010)
• 9. Montesinos, P., Magnier, B.: A New Perceptual Edge Detector in Color Images. In:

ACIVS, pp. 209-220. (2010).

• 10. Deriche, R.: Recursively implementing the gaussian and its derivatives. In: ICIP,
• pp. 263-267. (1992).
• 11. Mikolajczyk, K., Schmid, C.: A Performance Evaluation of Local Descriptors. In:

IEEE Trans. Pattern Anal. Mach. Intell, pp. 1615-1630. (2005) Page 66

Page 68