decision support for agri-food chains: a reverse engineering argumentation-based approach

Ecological Informatics xxx (2014) xxx–xxx

ECOINF-00482; No of Pages 10

Contents lists available at ScienceDirect

Ecological Informatics

j ourna l homepage: www.e lsev ie r .com/ locate /eco l in f

Decision support for agri-food chains: A reverse engineeringargumentation-based approach

Rallou Thomopoulos a,b,⁎, Madalina Croitoru b, Nouredine Tamani b

a IATE Joint Research Unit, UMR1208, CIRAD-INRA-SupAgro-Univ. Montpellier II, 2 place P. Viala, F-34060 Montpellier cedex 1, Franceb INRIA GraphIK, LIRMM, 161 rue Ada, F-34392 Montpellier cedex 5, France

⁎ Corresponding author at: IATE Joint Research Unit, UUniv. Montpellier II, 2 place P. Viala, F-34060 Montpellier

E-mail addresses: [email protected]@lirmm.fr (M. Croitoru), nouredine.tam

http://dx.doi.org/10.1016/j.ecoinf.2014.05.0101574-9541/© 2014 Elsevier B.V. All rights reserved.

Please cite this article as: Thomopoulos, R., eEcological Informatics (2014), http://dx.doi.o

a b s t r a c t
a r t i c l e i n f o
Article history:Received 24 December 2013Received in revised form 25 May 2014Accepted 27 May 2014Available online xxxx

Keywords:Decision supportKnowledge representationArgumentationReverse engineeringBackward chainingAgrifood chain control

Evaluating food quality is a complex process since it relies on numerous criteria historically grouped into fourmain types: nutritional, sensorial, practical and hygienic qualities. They may be completed by other emergingpreoccupations such as the environmental impact, economic phenomena, etc. However, all these aspects ofquality and their various components are not always compatible and their simultaneous improvement is aproblem that sometimes has no obvious solution, which corresponds to a real issue for decision making. Thispaper proposes a decision support method guided by the objectives defined for the end products of an agrifoodchain. It is materialised by a backward chaining approach based on argumentation.

© 2014 Elsevier B.V. All rights reserved.

1. Introduction

In agrifood chains, the products traditionally go through the inter-mediate stages of processing, storage, transport, and packaging, andreach the consumer (the demand) from the producer (the supply).More recently, due to an increase in quality constraints, several partiesare involved in production process, such as consumers, industrials,health and sanitary authorities, etc. expressing their requirements onthe final product as different point of views which could be conflicting.The notion of reverse engineering control, in which the demand (andnot the supply) sets the specifications of desired products and it is upto the supply to adapt and find its ways to respond, can be consideredin this case.

In this article, we discuss two aspects of this problem. First, weaccept the idea that specifications cannot be established and severalcomplementary points of view – possibly contradictory – can beexpressed (nutritional, environmental, taste, etc.). We then need toassess their compatibility (or incompatibility) and identify solutionssatisfying amaximumset of viewpoints. To this endwepropose a logical

MR1208, CIRAD-INRA-Supagro-cedex 1, France.(R. Thomopoulos),[email protected] (N. Tamani).

t al., Decision support for agrrg/10.1016/j.ecoinf.2014.05.0

framework based on argumentation and introduce amethod of decisionmaking based on backward chaining for the bread industry.

Since a joint argumentation–decision support approach is highlyrelevant to the food sector (Thomopoulos et al., 2009), the contributionof the paper is twofold. First we present a real use case of an argumen-tation process in the agrifood domain. Second we introduce the notionof viewpoint/goal in this setting based on the notion of backwardchaining reasoning and show how to use those techniques in a concreteapplication.

The main alternative method to deal with the problem is themulticriteria decision approach. However multicriteria decision aimsat evaluating several alternative options, whereas argumentation-based decision focuses on whether several options make sense togeth-er, which is a different perspective, addressed in this paper. Moreover,multicriteria decision is not connected to the backward chainingprocedure as the argumentative approach is, by construction of thearguments, as will be explained in Section 5.2.

In Section 2, we introduce the real scenario considered in theapplication. In Section 3, we motivate our technical and modellingchoices. In Section 4, the developed approach is introduced. It relies onan instantiation of a logic based argumentation framework based on aspecific fragment of first order logic. In Section 5, we explain thetechnical results that ensure the soundness and completeness of ouragronomy application method. In Section 6, some evaluation resultsare presented. Finally, Section 7 concludes the paper.

i-food chains: A reverse engineering argumentation-based approach,10

http://dx.doi.org/10.1016/j.ecoinf.2014.05.010

mailto:[email protected]




http://www.sciencedirect.com/science/journal/15749541


2 R. Thomopoulos et al. / Ecological Informatics xxx (2014) xxx–xxx

2. Scenario

The case of study considered in this paper relates to the debatearound the change of ash content in flour used for common Frenchbread. Various actors of the agronomy sector are concerned, in particu-lar the Ministry for Health through its recommendations within theframework of the PNNS (“National Program for Nutrition and Health”),the millers, the bakers, the nutritionists and the consumers.

The PNNS recommends to privilege the whole-grain cereal productsand in particular to pass to a common bread of T80 type, i.e. made withflour containing an ash content (mineral matter rate) of 0.8%, instead ofthe type T65 (0.65% of mineral matter) currently used. Increasing theash content comes down to using a more complete flour, since mineralmatter is concentrated in the peripheral layers of the wheat grain, aswell as a good amount of components of nutritional interest (vitamins,fibres). However, the peripheral layers of the grain are also exposed tothe phytosanitary products, which do not make them advisable from ahealth point of view, unless one uses organic flour.

Other arguments (and of various nature) are in favour or discreditwhole-grain bread. From an organoleptic point of view for example,the bread loses out in its “being crusty”. From a nutritional point ofview, the argument according to which the fibres are beneficial forhealth is discussed, some fibres could irritate the digestive system.From an economic point of view, the bakers fear selling less bread, be-cause whole-grain bread increases satiety — which is beneficial from anutritional point of view, for the regulation of the appetite and thefight against food imbalances and pathologies. However whole-grainbread requires also less flour and more water for its production, thusreducing the cost. The millers also fear a decrease in the quality of thetechnical methods used in the flour production.

Beyond the polemic on the choice between two alternatives (T65 orT80), one can take the debate further by distinguishing the variouspoints of view concerned, identifying the desirable target characteris-tics, estimating the means of reaching that point. The contribution ofthis paper is showing how using argumentation can help towardssuch practical goals.

3. Motivation

In this paperwewill elicit the points of view and the desirable targetcharacteristics by themeans of interviewswith agronomyexperts. Oncethe target characteristics are identified, finding the means of reachingthem will be done automatically by a combination of reverse engineer-ing and argumentation. The reverse engineeringwill be used in order tofind the complete set of actions to take towards a given characteristic,for all characteristics. In certain cases the actions to take will beinconsistent. Argumentation will then be employed in order to identifyactions that can be accepted together.

3.1. Reverse engineering

While reverse engineering has been widely employed in otherComputer Science domains such asmulti agent systems or requirementsengineering (e.g. Brunelière et al., 2014), it is quite a novel methodologywhen applied in agronomy. In agrifood chains, the products traditionallygo through the intermediate stages of processing, storing, transporting,and packaging and reach the consumer (the demand) from the producer(the supply). It is only recently, due to an increase in quality constraints,that the notion of reverse engineering control has emerged (Perrotet al., 2011). In this case the demand (and not the supply) sets thespecifications of desired products and it is up to the supply to adaptand find its ways to respond. In what follows, starting from thedesired target criteria for the final product, the methods allowingone to identify ways to achieve these criteria (by interventionon the various stages of the supply chain) are named “reverseengineering”.

Please cite this article as: Thomopoulos, R., et al., Decision support for agrEcological Informatics (2014), http://dx.doi.org/10.1016/j.ecoinf.2014.05.0

Reverse engineering is known to be challenging from amethodolog-ical viewpoint. This is due to two main aspects. First, is the difficulty ofdefining the specifications for the expected finished product. The de-sired quality criteria are multiple, questionable, and not necessarilycompatible. The next difficulty lies in the fact that the impact of differ-ent steps of food processing and their order is not completely known.Some steps are more studied than others, several successive steps canhave opposite effects (or unknown effects), and the target criteriamay be outside of the characteristics of products. Second, reconcilingdifferent viewpoints involved in the food sector still raises unaddressedquestions. The problem does not simply consist in addressing a multi-criteria optimisation problem (Bouyssou et al., 2009): the domain ex-perts would need to be able to justify why a certain decision (or set ofpossible decisions) is taken.

3.2. Argumentation

Argumentation is a reasoning model based on the constructionand the evaluation of interacting arguments. It has been applied tononmonotonic reasoning, decision making, or for modelling differenttypes of dialogues including negotiation. Most of the models developedfor these applications are grounded on the abstract argumentationframework proposed by Dung (1995). This framework consists of a setof arguments and a binary relation on that set, expressing conflictsamong arguments. An argument gives a reason for believing a claim,for doing an action.

Argumentation theory in general (Besnard and Hunter, 2008; Dung,1995; Rahwan and Simari, 2009) is actively pursued in the literature.Some approaches combine argumentation and multi criteria decisionmaking (Amgoud and Prade, 2009).

Value based Argumentation Frameworks (Bench-Capon, 2003) havebeen proposed, where the strength of an argument corresponds to thevalues it promotes. What we call viewpoint later on in this paperwould then correspond to the notion of audience in such setting.Although intuitive, this approach is not adapted in the case of theconsidered application. Here a value can be “split” into severalaudiences: there could be contradictory goals even from the sameviewpoint. The notion of viewpoint and goals introduced in this settingalso remind those proposed by Assaghir et al. (2011).

3.2.1. Logic-based argumentationIn this paper we present a methodology combining reverse engi-

neering and logical based argumentation for selecting the actions totake towards the agronomy application at hand. The logicalinstantiation language is a subset of first order logic denoted in thispaper SRC equivalent to Datalog + − (Calì et al., 2010), ConceptualGraphs or Description Logics (more precisely the EL fragment(Baader et al., 2005) and DL-Lite families (Calvanese et al., 2007)).All above mentioned languages are logically equivalent in terms ofrepresentation or reasoning power. The reason why this applicationis using SRC is the graph based representation proper to SRC (andnot to the other languages). This graph based representation (imple-mented in the Cogui tool (Chein and Mugnier, 2009; Chein et al.,2013)) makes the language suitable for interacting with non computingexperts (Chein et al., 2013).

Here we use the instantiation of Croitoru and Vesic (2013) for defin-ing what an argument and an attack are. While other approaches suchas García and Simari (2004), Besnard and Hunter (2005) and Mullerand Hunter (2012) address first order logic based argumentation, thework of Croitoru and Vesic (2013) uses the same SRC syntax andgraph reasoning foundations. In Fig. 1 the visual interface of Cogui isdepicted: knowledge is represented as a graph which is enricheddynamically by rule application. More on the visual appeal of Cogui forknowledge representation and reasoning can be found in Chein et al.(2013).



Fig. 1. The Cogui visual graph based interface.

3R. Thomopoulos et al. / Ecological Informatics xxx (2014) xxx–xxx

4. Approach

As mentioned above, in this paper we use an instantiation of logicbased argumentation based on a specific fragment of first order logic.This subset is equivalent to Datalog + − (Calì et al., 2010), Concep-tual Graphs or Description Logics (the EL fragment (Baader et al.,2005) and the DL-Lite families (Calvanese et al., 2007)). The reasonfor which our application required this specific logic fragment is re-lated to the information capitalisation needs of the food sector. Thelong term aim is to enrich ontologies and data sources based onthese ontologies and join the Open Data movement. This entailsthat the language used by the food applications needs to be compat-ible with the Semantic Web equivalent languages as mentionedbefore.

The choice of the SRC syntax and graph reasoning mechanism isjustified by the visual appeal of this language for non computingexperts.

In a nutshell our methodology is as follows. The set of goals, view-points as well as the knowledge associated with the goals/viewpoints iselicited either by the means of interviews with the domain experts ormanually from different scientific papers. This step of the application isthe most time consuming but the most important. If the knowledgeelicited is not complete, sound or precise the outcome of the system iscompromised. Then, based on the knowledge elicited from theknowledge experts and the goals of the experts,we enrich the knowledgebases using reverse engineering (implemented using backward chainingalgorithms). Putting together the enriched knowledge bases obtained bybackward chaining from the different goals will lead to inconsistencies.The argumentation process is used at this step and the extensions yieldedby the applications are computed. Based on the extensions and the


associated viewpoints we can use voting functions to determine theapplication choice of viewpoints.

4.1. Use case real data

Expressing the target characteristics – or goals – according to variouspoints of viewconsists of identifying the facets involved in the constructionof product quality: points of view, topics of concern such as nutrition,environment, technology, etc. In addition, such viewpoints have to beaddressed according to their various components (fibres, minerals,vitamins, etc.). Desirable directions need to be laid down, and in a firststep we consider them independent one from another.

The considered sources of information include, from most formalto less formal: (1) peer reviewed scientific papers; (2) technical re-ports or information posted on websites; (3) conferences and scien-tific meetings around research projects; and (4) expert knowledgeobtained through interviews. The scientific articles we haveanalysed – with the supervision of experts in agrifood – include:Bourre et al. (2008); Dubuisson-Quellier (2006); Ginon et al.(2009); Layat (2011); and Slavin and Green (2007). Bourre et al.(2008) compare the different types of flour from a nutritionalpoint of view. Slavin and Green (2007) explore the link betweenfibre and satiety. Dubuisson-Quellier (2006); and Ginon et al.(2009) deal with consumer behaviour and willingness to pay. Theyfocus on French baguette when information concerning the levelof fibres is provided, and they base their results on statistical studiesof consumer panels. Layat (2011) provides a summary of thenutritional aspects of consumption of bread and the link with tech-nological aspects.




We also reviewed technical reports available on official websiteson health policy: the public PNNS (National Program for Nutritionand Health, www.mangerbouger.fr/pnns) (PNNS (documentsstatutaires), 2010), the European project Healthgrain (looking at im-proving nutrition and health through grains) (Dean et al., 2007;HEALTHGRAIN, 2009), as well as projects and symposia on sanitarymeasures regarding the nutritional, technological and organolepticproperties of breads (AQUA−NUP, 2009; CADINNO, 2008;DINABIO, 2008; FCN, 2009). Finally, several interviews were con-ducted to collect domain expert knowledge, in particular for technol-ogy specialists in our laboratory.

A summary of the results obtained in the baking industry is synthe-sised in Fig. 2 regarding nutritional and organoleptic aspects. Fig. 2(a)shows the main identified goals to reach for a nutritionally optimisedbread (for instance, containing a high level of soluble fibres, vitaminsand minerals, low salt, etc.), whereas Fig. 2(b) sums up the main goalsto achieve for an enjoyable bread regarding sensorial concerns (forexample, crusty, etc.).

5. Technical soundness

In this section we explain the technical results that ensure thesoundness and completeness of our agronomy application method.The section is composed of three parts. A first subsection explainsthe logical subset of first order logic language employed in thepaper. The second subsection shows how to construct argumentsand attacks in order to obtain extensions when a knowledge baseexpressed under this language is inconsistent. Last, the third sectionshows howwe used reverse engineering to complete the knowledgebase with all possible actions and how argumentation can be used inorder to select consistent subsets of knowledge which support givenactions.

5.1. The logical language

In the following, we give the general setting knowledge represen-tation language used throughout the paper.

A knowledge base is a 3-tuple K ¼ F ;R;Nð Þ composed ofthree finite sets of formulae: a set F of facts, a set R of rulesand a set N of constraints. Let us formally define what we acceptas F ;R and N .

5.1.1. Facts syntaxLet C be a set of constants and P= P1 ∪ P2…∪Pn a set of predicates

of the corresponding arity i = 1,…n. Let V be a countably infinite setof variables. We define the set of terms by T = V∪C. As usual, given i∈ {1…n}, p∈ Pi and t1,…,ti ∈ Twe call p (t1,…,ti) an atom. A fact is theexistential closure of an atom or an existential closure of aconjunction of atoms. (Note that there is no negation or disjunctionin the facts and that we consider a generalised notion of facts thatcan contain several atoms.)

• Bread, Cereal, LowSalt, ContaminantFree are examples of unarypredicates (arity 1) and IsIngredientOf is a binary predicate (arity 2).

• Wheat, oats, rye, barley are constant examples.• Cereal (wheat) is an atom.• ∃ x (Bread(x) ∧ IsIngredientOf(wheat, x)) is a fact.

Due to lack of space we do not show the full semantic definitions offacts (or rules and constraints in the following section). For a completesemantic depiction of this language please check Chein and Mugnier(2009); Chein et al. (2013); and Croitoru and Vesic (2013). It is wellknown that F′ ⊨ F (read the fact F′ entails the fact F) if and only if thereis a homomorphism from F to F′ (Chein and Mugnier, 2009).


5.1.2. RulesA rule R is a formula of the form

∀x1;…;∀xn ∀y1;…;∀ym H x1;…xn; y1;…; ymð Þ→∃z1;…;∃zk C y1;…; ym; z1;…zkð Þð Þ

where H, the hypothesis, and C, the conclusion, are atoms or conjunc-tions of atoms, n, m, k ∈{0, 1,…}, x1,…, xn are the variables appearingin H, y1,…,ym are the variables appearing in both H and C and z1,....zkthe new variables introduced in the conclusion. An example of a ruleis the following:

∀ x Bread xð Þ ∧PesticideFree xð Þ∧MycotoxinFree xð Þ→ContaminantFree xð Þð Þ:

In the following we will consider rules without new existentialvariables in the conclusion.

Reasoning consists of applying rules on the set F and thus inferringnew knowledge. A rule R = (H, C) is applicable to set F if and only ifthere exists F ′ ⊆ F such that there is a homomorphism σ from the hy-pothesis ofR to the conjunction of elements of F ′. A rule R = (H, C) isinversely applicable to a fact F if there is a homomorphism π from C to F.In this case, the inverse application of R to F according to π produces anew fact F′ such that R(F′) = F. We then say that the new fact is animmediate inverse derivation of F by R, abusively denoted R−1 (F).

Note that this technique is commonly used, for example, for back-ward chaining query answering (Baget and Salvat, 2006; Konig et al.,2012)where a query is rewritten according to the rules. The samemech-anism is also discussed by abductive reasoning algorithms (Klarmanet al., 2011) where minimal sets of facts (in the set inclusion sense) areadded to the knowledge base in order to be able to deduct a query.

Let F = Bread(bleuette) ∧ PesticideFree(bleuette) ∧ MycotoxinFree(bleuette) and R the rule ∀ x (Bread(x) ∧ PesticideFree(x) ∧ MycotoxinFree(x)→ ContaminantFree(x)).

R is applicable to F and produces by derivation the following fact:Bread (bleuette) ∧ PesticideFree(bleuette) ∧ MycotoxinFree(bleuette) ∧ContaminantFree(bleuette).

Let F = Bread(bleuette) ∧ ContaminantFree(bleuette) and R the rule∀ x (Bread(x) ∧ PesticideFree(x) ∧ MycotoxinFree(x)→ ContaminantFree(x)).

R inversely applicable to F and produces by inverse derivation thefact: F′ = Bread(bleuette) ∧ PesticideFree(bleuette) ∧ MycotoxinFree(bleuette).

Let F be a subset of F and letRbe a set of rules. A set Fn is called anR ‐ derivation of F if there is a sequence of sets (called a derivationsequence) (F0, F1, …, Fn) such that F0 ⊆ F, F0 is R − consistent, forevery i ∈ {1, …, n − 1}, it holds that Fi is an immediate derivationof Fi − 1.

Given a set {F0,…, Fk}⊆ F and a set of rulesR, the closure of {F0,… Fk}w.r.t.R, denoted ClR({F0,…, Fk}), is defined as the smallest set (with re-spect to ⊆) which contains {F0, …, Fk}, and is closed for R − derivation(that is, for every R − derivation Fn of {F0,…,Fk}, we have Fn ⊆ ClR({F0, …, Fk})). Finally, we say that a set F and a set of rules R entail afact G (and we write F , R ⊨ G) if and only if the closure of the facts byall the rules entails F (i.e. if ClR(F) ⊨ G).

5.1.3. ConstraintsA constraint is a formula ∀ x1 … ∀ xn(H(x1, …, xn) →⊥), where H is

an atom or a conjunction of atoms and n ∈ {0, 1, 2, …}. Equivalently, aconstraint can be written as ¬ (∃ x1,…, ∃ xnH(x1,… xn)). As an exampleof a constraint, consider N = ¬ (∃ x (Growth(x) ∧Decrease(x))).

Given a knowledge baseK ¼ F ;R;Nð Þ, a set {F1,…, Fk}⊆ F is said tobe inconsistent if and only if there exists a constraint N∈N such that{F1,…, Fk}⊨HN, whereHN denotes the existential closure of the hypoth-esis of N. A set is consistent if and only if it is not inconsistent. A set{F1,…, Fk} ⊆ F isR− inconsistent if and only if there exists a constraintN∈N such that ClR({F1,…, Fk}) ⊨ HN, where HN denotes the existentialclosure of the hypothesis of N.


http://www.mangerbouger.fr/pnns


Fig. 2. Nutritional (a) and organoleptic (b) goals.


Please cite this article as: Thomopoulos, R., et al., Decision support for agri-food chains: A reverse engineering argumentation-based approach,Ecological Informatics (2014), http://dx.doi.org/10.1016/j.ecoinf.2014.05.010



Let K ¼ F ;R;Nð Þ where:

• F contains the following facts:– F1 = Bread(bleuette) ∧ ContaminantFree(bleuette)– F2 = ∃ e ExtractionRate(e,bleuette)– F3 = ∃ f (FiberContent(f,bleuette) ∧ High(f))

• R consists of the following rules:– R1 = ∀ x,y (Bread(x) ∧ ExtractionRate(y,x) ∧ PesticideFree(x)

→ Decrease (y))– R2 = ∀ x,y,z (Bread(x) ∧ ExtractionRate(y,x) ∧ FiberContent(z,x) ∧

High(z) → Growth(y))– R3 = ∀ x (Bread(x) ∧ ContaminantFree(x) → PesticideFree(x)

∧ MycotoxinFree(x))• N contains the following negative constraint:– N = ¬ (∃ x (Growth(x) ∧ Decrease(x)).

K is inconsistent since (F ,R)⊨ N. Indeed, F1 and R3 allow to deducePesticideFree(bleuette). Combined to F2 and R1 we obtain Decrease(e). F3and R2 deduce Growth(e), violating the negative constraint N.

Given a knowledge base, one can ask a conjunctive query in order toknow whether something holds or not. Without loss of generality weconsider boolean conjunctive queries (which are facts). As an exampleof a query, take ∃x1cat(x1). The answer to query α is positive if andonly if F , R ⊨ α.

Answering Q, traditionally, has two different algorithmicapproaches: either forward chaining or backwards chaining. Thetwo approaches come to either (1) finding an answer of Q in theR ‐ derivations of the facts in the knowledge base or (2) computingthe inverse R ‐ derivations of the query and finding if there is amatch in the facts. We will focus on the latter approach in thefollowing.

5.2. Arguments and attacks

This section shows that it is possible to define an instantiation ofDung's abstract argumentation theory (Dung, 1995) that can be usedto reason with an inconsistent ontological KB.

We first define the notion of an argument. For a set of formulaeG ¼ G1;…;Gnf g, notation∧ G is used as an abbreviation for G1 ∧…∧Gn.

Definition 1. Given a knowledge baseK ¼ F ;R;Nð Þ, an argument a isa tuple a = (F0, F1,…,Fn) where:

• (F0,…,Fn − 1) is a derivation sequence with respect to K• Fn is an atom, a conjunction of atoms, the existential closure of anatom or the existential closure of a conjunction of atoms such thatFn − 1 ⊨ Fn.

This definition, following the definition of Croitoru and Vesic (2013)is a straightforward way to define an argument, since an argumentcorresponds to a derivation.

To simplify the notation, fromnowon, we suppose thatwe are givena fixed knowledge baseK ¼ F ;R;Nð Þ and do not explicitly mention F ,R norN if not necessary. Let a = (F0,…, Fn) be an argument. Then, wedenote Supp (a) = F0 and Conc (a) = Fn.

Arguments may attack each other, which is captured by a binaryattack relation Att ⊆ Arg(F) × Arg(F).

Definition 2. LetK ¼ F ;R;Nð Þ be a knowledge base and let a and b betwo arguments. The argument a attacks argument b, denoted (a,b) ∈Att, if and only if there exists φ ∈ Supp (b) such that the set {Conc (a),φ} is R − inconsistent.

This attack relation is not symmetric. To see why, considerthe following example. Let F = {p(m), q(m), r(m)}, R = ∅, N ¼∀x1 p x1ð Þ∧q x1ð Þ∧r x1ð Þ→⊥ð Þf g . Let a = ({p(m), q(m)}, p(m) ∧q(m)), b = ({r(m)}, r(m)). We have (a,b) ∈ Att and (b,a) ∉ Att.This will ensure that the naive extension is different, at least in


theory, from the preferred, stable, etc. semantics. However, in ourapplication they all entail the same information as shown later on.

Definition 3. Given a knowledge baseK ¼ F ;R;Nð Þ, the correspond-ing argumentation framework AFK is a pair A ¼ Arg Fð Þ;Attð Þ whereArg(F) is the set of all arguments that can be constructed from F andAtt is the corresponding attack relation as specified in Definition 2.

Let E⊆A and a∈A. We say that E is conflict free if and only if thereexists no arguments a, b ∈ E such that (a,b) ∈ Att. E defends a if andonly if for every argument b∈A, if we have (b,a)∈ Att then there existsc ∈ E such that (c,b) ∈ Att.

E is admissible if and only if it is conflict free and defends all itsarguments. E is a complete extension if and only if E is an admissibleset which contains all the arguments it defends. E is a preferredextension if and only if it is maximal (with respect to set inclusion)admissible set. E is a stable extension if and only if it is conflict-freeand for all a∈A∖E, there exists an argument b ∈ E such that (b,a) ∈ Att.

E is a grounded extension if and only ifℰ is a minimal (for set inclu-sion) complete extension.

For an argumentation framework AS ¼ A;Attð Þ we denote by Extx(AS) (or by Extx A;Attð Þ) the set of its extensions with respect to seman-tics x. We use the abbreviations c, p, s, and g for respectively complete,preferred, stable and grounded semantics.

An argument is sceptically accepted if it is in all extensions, credu-lously accepted if it is in at least one extension and rejected if it is notin any extension.

Based on this definition of arguments and attacks in Croitoruand Vesic (2013) was also shown that the rationality postulates ofCaminada andAmgoud (2007) are respected. This instantiation respectsthe direct, indirect consistency as well as the closure.

5.3. Formalising the use case

In this subsection we formalise the notions presented in Section 4.Let K ¼ F ;R;Nð Þ be a consistent knowledge base. This is the

knowledge base that all actors share and agree upon. In this paperwe assume that the rules and negative constraints are common toeverybody.

The goals of the different actors can be seen as a set of existentiallyclosed conjuncts. We denote them by G1, G2, …, Gn.

Let Gi be a goal and K the knowledge base. K is consistent and Kdoes not entail Gi. We compute the inverse R − derivations of Gi

(where R is the set of rules of the knowledge base). We add all ofthe R−1(Gi) to the facts. We thus obtain a new knowledge base Ki

which differs fromK solely by its facts set (which now also includesR−1(Gi)): K ¼ F∪R−1 Gið Þ;R;N� �

. We also impose that Ki isconsistent.

Given G ¼ G1;G2;…;Gnf g, the goals correspond to a set of view-points V (there exists a function κ : G→2V). This function can assign agoal to one or more viewpoints and each viewpoint can be associatedwith one or more goals. Given a goal Gi, the (set of) viewpoint(s) asso-ciated with this goal is denoted by κ(Gi). Similarly, given a viewpoint vi,the set of goals associated with it is denoted by κ−1(vi).

Example 1. Let the set of viewpointsV= {nutrition, sanitary, organoleptic}and G consisting of the following goals: G1 = ∃ x (Bread(x)∧ LowSalt(x)),G2 = ∃ x (Bread(x) ∧ ContaminantFree(x)), G3 = ∃ x (Bread(x) ∧Crusty(x)), G4 = ∃ x (Bread(x) ∧ TraceElementRich(x)).

We have κ(G1) = κ(G4) = nutrition, κ(G2) = sanitary and κ(G3) =organoleptic. Conversely κ−1 (nutrition) = {G1, G4}, κ−1 (sanitary) ={G2} and κ−1 (organoleptic) = {G3}.

The rules will correspond to the set of sufficient conditions neededfor the goal Gi. In the context of our practical application this is illustrat-ed in Fig. 3 (with respect to nutrition goals).



Fig. 3.Ways to reach nutritional goals.


Example 2. To reach the goal G1 = ∃ x (Bread(x) ∧ LowSalt(x)), theknowledge base K contains the following rule: ∀ x,y (Bread(x) ∧SaltAdjunction(y,x) ∧ Decrease(y) → LowSalt(x)).

Let us now consider the set of goals G ¼ G1;G2;…;Gnf g and theinitial knowledge base K ¼ F ;R;Nð Þ . As described above wecompute the n knowledge bases, corresponding to each goal: Ki ¼F∪R−1 Gið Þ;R;N� �

for each i = 1,…,n. We consider the union of allthese knowledge bases:

Kagg ¼ F ∪i¼1;…;n

R−1 Gið Þ;R;N� �

Example 3. Let K ¼ F ;R;Nð Þ where:

• F = {Fi} = {CurrentExtractionRate(T65)}• R contains the following rules:- R1–∀ x,y (Bread(x)∧ ExtractionRate(y,x)∧ Decrease(y)→ Digestible(x))

– R2 = ∀ x,z (Bread(x) ∧ SaltAdjunction(z,x) ∧ Decrease(z) → LowSalt(x))

– R3 = ∀ x,y (Bread(x) ∧ ExtractionRate(y,x) ∧ Growth(y) → TraceElementRich(x))

– R4=∀ x,y (Bread(x)∧ ExtractionRate(y,x)∧ Decrease(y)→ PesticideFree(x))

• N contains the following negative constraint:– N = ¬(∃ x (Growth(x) ∧ Decrease(x)))

Let the goal set G as follows:

• G1 = ∃ p (Bread(p) ∧ Digestible(p)), where κ(G1) = nutrition• G2 = ∃ p (Bread(p) ∧ LowSalt(p)), where κ(G2) = nutrition• G3 = ∃ p (Bread(p)∧ TraceElementRich(p)), where κ(G3) = nutrition• G4 = ∃ p (Bread(p) ∧ PesticideFree(p)), where κ(G4) = sanitary.

Then:

• K1 ¼ F 1;R;Nð Þ where F 1 = F ∪ R−1(G1) contains the followingfacts:– F1 = CurrentExtractionRate(T65)– F2 = Bread(p) ∧ ExtractionRate (τ,p) ∧ Decrease(τ)


• K2 ¼ F 2;R;Nð Þ where F 2 = F ∪ R−1(G2) contains the followingfacts:– F1 = CurrentExtractionRate(T65)– F3 = Bread(p) ∧ SaltAdjunction(s,p) ∧ Decrease(s)

• K3 ¼ F 3;R;Nð Þ where F 3 = F ∪ R−1(G3) contains the followingfacts:– F1 = CurrentExtractionRate(T65)– F4 = Bread(p) ∧ ExtractionRate(τ,p) ∧ Growth(τ)

• K4 ¼ F 4;R;Nð Þ where F 4 = F ∪ R−1(G4) contains the followingfacts:– F1 = CurrentExtractionRate(T65)– F2 = Bread(p ∧ ExtractionRate(τ,p) ∧ Decrease(τ)

Finally Kagg ¼ F∪i¼1;…;nR−1 Gið Þ;R;NÞ�where F ∪ i = 1,…,n R−1

(Gi) = {F1, F2, F3, F4}.

As observed in the previous example, it may happen that Kagg isinconsistent (and it does so even for goals belonging to the same view-point). We then use argumentation, which, by the means of extensionswill isolate subsets of facts we can accept together (called extensions).Furthermore, the extensions will allow us to see which are the view-points associated to each maximal consistent subset of knowledge (bymeans of the function κ). A choice procedure then has to be used (seeexample below).

The argument framework we can construct from the aboveknowledge base is A;Attð Þ where A contains the following:

• a = ({F2}, F2, R1(F2)) where R1(F2) = Bread(p) ∧ ExtractionRate(τ,p)∧ Decrease(τ) ∧ Digestible(p).

• b = ({F4}, F4, R3(F4)) where R3(F4) = Bread(p) ∧ ExtractionRate(τ,p ∧ Growth(τ) ∧ TraceElementRich(p).

• c = ({F2}, F2, R4(F2)) where R4(F2) = Bread(p) ∧ ExtractionRate(τ,p) ∧ Decrease(τ) ∧ PesticideFree(p).

• d = ({F3}, F3, R2(F3)) where R2(F3) = Bread(p) ∧ SaltAdjunction(s,p) ∧ Decrease(s) ∧ LowSalt(p) and Att = {(a,b), (b,a), (b,c),(c,b)}.

In this argumentation system defined we now obtain:

• Extstable A;Attð Þ ¼ Extsemi−stable A;Attð Þ ¼ Extpreferred A;Attð Þ ¼ a; c; df g;ffb; dgg.



Fig. 4. Demonstrator screenshot showing two sets of possible actions.


Starting from the extensionsExtx A;Attð Þ, the proposed decision sup-port system functions as follows: for every extension ε∈Extx A;Attð Þ:

• Consider the facts occurring in the arguments of ε;• Identify the knowledge bases Ki where these facts occur;• Obtain the goals Gi which are satisfied by the extension;• Using the κ function to obtain the viewpoints corresponding to thesegoals;

• Show domain experts the set of goals, and compatible viewpointscorresponding to the given extension.

This method allows us to obtain a set of options equal to the cardi-nality of Extx A;Attð Þ. For taking a final decision several possibilitiescan be considered and presented to the experts:

• Maximise the number of goals satisfied;• Maximise the number of viewpoints satisfied;• Use preference relations of experts on goals and/or viewpoints.

In the previous example (please recall that the goals G1 and G2 areassociated with the nutritional viewpoint while G4 is associated withthe sanitary viewpoint) we have:

• The first extension {a,c,d} is based on the facts F2 and F3 obtained fromK1, K2 and K4 that satisfy the goals G1, G2 and G4.

• The second extension {b,d} is based on F3 and F4 obtained fromK2 andK3 satisfyingG2 andG3 both associatedwith the nutritional viewpoint.

One first possibility (corresponding to the extension {a,c,d}) consistsof accomplishing F2 and F3 and allows to satisfy the biggest number ofgoals and viewpoints.

The second possibility (corresponding to the extension {b,d}) con-sists of accomplishing F3 and F4. It would satisfy two goals and one view-point. It could be considered though if the goal G3 (not satisfied by thefirst option) is preferred to the others.


6. Evaluation

The evaluation of the implemented system was done via a seriesof interviews with domain experts. The above knowledge and reason-ing procedures were implemented using the Cogui knowledge repre-sentation tool (Chein et al., 2013), with an extension of 2000 lines ofsupplemental code. Three experts have validated our approach: two re-searchers in food science and cereal technologies of the French nationalinstitute of agronomic research, specialists respectively of the grain-to-flour transformation process and of the breadmaking process, and oneindustrial expert — the Director of the French National Institute ofBread and Pastry.

Thefirstmeetingdealtwith the delimitation of theproject objectivesand addressed fundamental questions such as: Is it possible to uniquelydefine a “good” bread? Which scenarii of “good bread” should be con-sidered? How could they be defined from a nutritional, sanitary, senso-rial and economic point of view? Which are the main known ways toachieve them?

Then a series of individual interviews constituted the elicitationphase. Each expert gave more arguments which were complementingeach other. In parallel, the writing of specifications for the demon-strator and the definition of the knowledge base structure wereconducted.

In the following plenary meeting the real potential of the approachwas shown. The experts were formulating goals and viewpoints theywere interested in and the Cogui system together with the argumenta-tion extension was yielding the associated possible propositions. Fig. 4shows a screenshot of the demonstrator answers for a two-goalquery: a nutritional goal (high fibre content) and an organoleptic goal(crusty bread). Two sets of compatible actions are proposed, somechoices (such as increasing or decreasing the extraction rate) beingincompatible for both goals, and thus separated in the two alternativesets.




Four scenarii were more specifically evaluated. These scenariiconcern four kinds of consumers: obese (fibre preference), peoplewith iron deficiency (micronutrient preference), people withcardiovascular disease (decreased salt preference) and vegetarians(limited phytic acid), which produces different sets of goals. Foreach scenario, the system proposes several outputted recommenda-tions. The audience for decreasing salt tips the balance in favour of arecommendation for the T80 bread, while the audience for decreas-ing phytic acid pushes to specify recommendations towards a natu-ral sourdough bread or a conservative T65 bread. Other audiencesare in favour of a status quo. The results were considered asexplanable by experts, but not obvious, since many considerationshad to be taken into account.

Two interests of the approach were more particularly highlighted.They concern cognitive considerations. Firstly, experts were consciousthat the elicitation procedure was done according to their thought pro-cesses, that is, in a forwardwaywhich ismore natural and intuitive. Thesystem was thus able to restitute the knowledge in a different mannerthan the experts usually do. Secondly, from a problem that could initial-ly seem simple, the experts realised that it covered a huge complexitythat a human mind could hardly address alone. The tool is currentlyavailable to them under restricted access.

The knowledge modelling task can be a very time-consuming step.As presented in Section 4.1, several sources of information were used,from peer reviewed scientific papers and technical reports, to confer-ence meetings and expert interviews. On the one hand, expert inter-views appeared to be the least expensive ones in terms of time. A one-day period allows both eliciting knowledge through an interview andformalising it in the software system — which constitutes the longestpart of the work. However, this relatively short time hides a strongprerequisite: having already a clear view of the case study, a synopsisof the questions to ask the expert and an implemented knowledgemodel. On the other hand, websites, technical reports and scientificarticles are more costly to analyze. For instance, the critical reading ascientific paper of the domain may require a one-day period on itsown, for a discerning reader. However they allow one to grasp the insand outs of the question.

During the evaluation step, the experts raised the question of theimportance attached to the different pieces of knowledge modelled inthe system. Moreover, in some cases experts may hesitate on the rele-vance of some facts or rules. A possibility would thus be to adopt apreference-based argumentation system, as proposed in several workssuch as Amgoud and Cayrol (2002); Amgoud et al. (2000); Bench-Capon (2003); Bourguet et al. (2013); and Kaci and van der Torre(2008), able to take into account different levels of importance amongarguments.

7. Conclusion

Even if argumentation based decision making methods appliedto the food industry were also proposed by Bourguet (2010) andBourguet et al. (2013), this paper addresses a key point in the contextof current techniques used by the food sector and namely addressing re-verse engineering. Also, in this approach, an argument is used here as amethod computing compatible objectives in the sector. This case studyrepresents an original application and an introspective approach in theagronomy field by providing an argumentation based decision-supportsystem for the various food sectors. It requires nevertheless the very ex-pensive task of knowledgemodelling. Such task, in its current state can-not be automated. It strongly depends on the quality of expert opinionand elicitation (exhaustiveness, certainty, etc.). The current trend fordecision-making tools includes more and more methods of argumenta-tion as means of including experts in the task of modelling and thedecision-making processes. Another element to take into account, notdiscussed in this paper, is the difficulty of technologically (from an


agronomy viewpoint) putting in place the facts of each option. Model-ling this aspect in the formalism is still to be studied.

References

Amgoud, L., Cayrol, C., 2002. A reasoning model based on the production of acceptablearguments. Ann. Math. Artif. Intell. 34, 197–216.

Amgoud, L., Prade, H., 2009. Using arguments for making and explaining decisions. Artif.Intell. 173, 413–436.

Amgoud, L., Parsons, S., Perrussel, L., 2000. An argumentation framework based oncontextual preferences. Proceedings of the International Conference on Formal andApplied and Practical Reasoning, pp. 59–67.

AQUA−NUP, 2009. http://www.inra.fr/inra_cepia/vous_recherchez/des_projets/france/aquanup.

Assaghir, Z., Napoli, A., Kaytoue, M., Dubois, D., Prade, H., 2011. Numerical informationfusion: lattice of answers with supporting arguments. ICTAI, pp. 621–628.

Baader, F., Brandt, S., Lutz, C., 2005. Pushing the el envelope. Proc. of IJCAI 2005.Baget, J.F., Salvat, E., 2006. Rules dependencies in backward chaining of conceptual graphs

rules. Conceptual Structures: Inspiration and Application, 14th International Confer-ence on Conceptual Structures. Springer, pp. 102–116.

Bench-Capon, T.J., 2003. Persuasion in practical argument using value-based argumenta-tion frameworks. J. Log. Comput. 13, 429–448.

Besnard, P., Hunter, A., 2005. Practical first-order argumentation. Proc. of AAAI, pp. 590–595.Besnard, P., Hunter, A., 2008. Elements of argumentation. The MIT Press.Bourguet, J.R., 2010. Contribution auxmethodes d'argumentation pour la prise de decision.

(Thèse de doctorat) Application a l'arbitrage au sein de la filiere cerealiereUniversitéMontpellier II, Montpellier, France.

Bourguet, J.R., Thomopoulos, R., Mugnier, M.L., Abécassis, J., 2013. An artificialintelligence-based approach to deal with argumentation applied to food qualityin a public health policy. Expert Syst. Appl. 40, 4539–4546.

Bourre, J.M., Bégat, A., Leroux, M.C., Mousques-Cami, V., Pérandel, N., Souply, F., 2008.Valeur nutritionnelle (macro et micro-nutriments) de farines et pains français.Med. Nutr. 44, 49–76.

Bouyssou, D., Dubois, D., Pirlot, M., Prade, H., 2009. Decision-making process—conceptsand methods. Wiley.

Brunelière, H., Cabot, J., Dupé, G., Madiot, F., 2014. Modisco: a model driven reverse engi-neering framework. Inf. Softw. Technol. 56, 1012–1032.

CADINNO, 2008. Information, choix, consommateurs responsables: des leviers pour undéveloppement durable? http://www.melissa.ens-cachan.fr/IMG/pdf/Colloque_CadInno_FR.pdf.

Calì, A., Gottlob, G., Lukasiewicz, T., Marnette, B., Pieris, A., 2010. Datalog+/−: afamily of logical knowledge representation and query languages for newapplications. LICS, pp. 228–242.

Calvanese, D., De Giacomo, G., Lembo, D., Lenzerini, M., Rosati, R., 2007. Tractable reason-ing and efficient query answering in description logics: the dl-lite family. J. Autom.Reason. 39, 385–429.

Caminada, M., Amgoud, L., 2007. On the evaluation of argumentation formalisms. Artif.Intell. 171, 286–310.

Chein, M., Mugnier, M.L., 2009. Graph-based knowledge representation and reasoning—computational foundations of conceptual graphs. Advanced Information andKnowledge ProcessingSpringer.

Chein, M., Mugnier, M.L., Croitoru, M., 2013. Visual reasoning with graph-based mecha-nisms: the good, the better and the best. Knowl. Eng. Rev. 28, 249–271.

Croitoru, M., Vesic, S., 2013. What can argumentation do for inconsistent ontology queryanswering? Proc. of SUM 2013 (to appear).

Dean, M., Sheperd, R., Arvola, A., Lampila, P., Lahteenmaki, L., Vassalo, M., Saba, A.,Claupein, E., Winkelmann, M., 2007. Report on consumer expectations ofhealth benefits of modified cereal products. Technical Report. University ofSurrey, UK.

DINABIO, 2008. Proceedings of dinabio développement et innovation en agriculturebiologique. http://www.inra.fr/ciag/revue_innovations_agronomiques/volume_4_janvier_2009.

Dubuisson-Quellier, S., 2006. De la routine à la délibération. les arbitrages desconsommateurs en situation d'achat. Réseaux 135 (136), 253–284.

Dung, P.M., 1995. On the acceptability of arguments and its fundamental role innonmonotonic reasoning, logic programming and n-person games. Artif. Intell. J.77, 321–357.

FCN, 2009. Fibres, céréales et nutrition. http://www.inra.fr/content/view/full/24670029.García, A.J., Simari, G.R., 2004. Defeasible logic programming: an argumentative approach.

Theory Pract. Log. Program. 4, 95–138.Ginon, E., Lohérac, Y., Martin, C., Combris, P., Issanchou, S., 2009. Effect of fibre information

on consumer willingness to pay for french baguettes. Food Qual. Prefer. 20, 343–352.HEALTHGRAIN, 2009. http://www.healthgrain.org.Kaci, S., van der Torre, L., 2008. Preference-based argumentation: arguments supporting

multiple values. Int. J. Approx. Reason. 48, 730–751.Klarman, S., Endriss, U., Schlobach, S., 2011. Abox abduction in the description logic alc. J.

Autom. Reason. 46, 43–80.Konig, M., Leclere, M., Mugnier, M.L., Thomazo, M., 2012. A sound and complete backward

chaining algorithm for existential rules. In: Krotzsch, M., Straccia, U. (Eds.), WebReasoning and Rule Systems. volume 7497 of Lecture Notes in Computer Science.Springer, Berlin Heidelberg, pp. 122–138.

Layat, T., 2011. Place du pain dans l'équilibre alimentaire. Prat. Nutr. 7, 45–50.Muller, J., Hunter, A., 2012. An argumentation-based approach for decision making. Proc.

of ICTAI, pp. 564–571.


http://refhub.elsevier.com/S1574-9541(14)00056-9/rf0005







http://www.inra.fr/inra_cepia/vous_recherchez/des_projets/france/aquanup

http://www.inra.fr/inra_cepia/vous_recherchez/des_projets/france/aquanup























http://www.melissa.ens-cachan.fr/IMG/pdf/Colloque_CadInno_FR.pdf

http://www.melissa.ens-cachan.fr/IMG/pdf/Colloque_CadInno_FR.pdf




















http://www.inra.fr/ciag/revue_innovations_agronomiques/volume_4_janvier_2009

http://www.inra.fr/ciag/revue_innovations_agronomiques/volume_4_janvier_2009






http://www.inra.fr/content/view/full/24670029





http://www.healthgrain.org














Perrot, N., Trelea, I., Baudrit, C., Trystram, G., Bourgine, P., 2011. Modelling and analysis ofcomplex food systems: state of the art and new trends. Trends Food Sci. Technol. 22,304–314.

PNNS (documents statutaires), 2010. http://www.sante.gouv.fr/htm/pointsur/nutrition/pol_nutri4.htm.

Rahwan, I., Simari, G., 2009. Argumentation in artificial intelligence. Springer.


Slavin, J., Green, H., 2007. Diatery fibre and satiety. Br. Nutr. Found. 32 (1), 32–42.Thomopoulos, R., Charnomordic, B., Cuq, B., Abécassis, J., 2009. Artificial intelligence-

based decision support system to manage quality of durum wheat products. Qual.Assur. Saf. Crops Food 1, 179–190.





http://www.sante.gouv.fr/htm/pointsur/nutrition/pol_nutri4.htm

http://www.sante.gouv.fr/htm/pointsur/nutrition/pol_nutri4.htm







decision support for agri-food chains: a reverse engineering argumentation-based approach

Documents