Multiagent Approach for the Representation of Information in a Decision Support System

Multiagen t Approac h for the Represen tation of Information in a Decision Su pp ort System F ahem Kebair and F r ´ ed´ eric Serin Universit ´ e du Ha vre, LITIS - Lab oratoire d’Informatique, de T raitement d e l’Information et des Syst` emes, 25 rue Philipp e Leb on, 76058, Le Havre Cedex, F rance { fahem.k eb air, frederic.serin } @univ-lehavre.fr Abstract. In an emergency situation, the actors need an assistance al- lo wing them to react swiftly and efficiently . In this prosp ect, we present in th is pap er a decision supp ort system that aims to prepare actors in a crisis situation th an k s to a decision-making supp ort. The global archi- tecture of this system is presented in the first part. Then we fo cus on a part of this system whic h is designed to represent the info rmation of the current situation. This part is comp osed of a multiagen t system that is made of factual agents. Each agent carries a semantic feature and aims to represent a partial part of a situation. The agen ts d evelo p thanks to their interactions by comparing their semantic features u sing proximit y measures and according t o sp ecific ontologies . Keywords. D ecision supp ort system, F actual agent, Indicators, Multi- agen t system, Proximit y measure, Semantic feature. 1 In t ro duction Making a decisio n in a cris is situation is a complicated task. This is mainly due to the unpredictability and the rapid evolution of the environmen t sta te. Indeed, in a critic situation time and r esources are limited. Our knowledge ab out the environmen t is incomplete and uncertain, verily obsolete. Conseq uent ly , it is difficult to act and to adapt to the hostile conditions of the world. T his makes sense to the serio us need of ro bust, dynamic and intelligent planning system for search-and-rescue opera tions to cop e with the changing situation a nd to best sav e p eople [9]. The role of such a system is to pr ovide an emergency planning that a llows actor s to react swiftly a nd efficiently to a cr isis cas e. In this context, our aim is to build a system designed to help decision-makers manage cases o f crisis with an orig inal representation of info r mation. F ro m the system p oint o f vie w, detecting a crisis implies its repr e sentation, its characteri- sation a nd its comparison p erma nent ly with other cr isis stor ed in scenario s base. The res ult of this c o mparison is provided to the user as the a nswer of the global system . The idea bega n with th e sp eech in terpretatio n of hu man actors dur ing a cr isis [3], [5 ]. The goa l was to build an information, and c ommu nic ation s ystem (ICS) which enables the ma nagement of emergency situa tio ns b y interpreting asp ects communications created by the a ctors. T he n, a pr eventive vigil s ystem (P VS) [1] was designed with the mean o f some technologies used in the ICS mo delling as: semantic featur es, ontologies, and age n ts with internal v aria bles and b ehavioural automata. T he PVS aims either to preven t a crisis or to deal with it with a main int ernal g oal: detecting a crisis. Since 2003, the architecture of the PVS was redesigned with a new sp ecificity , that is the gener ic a sp ect; generic is used here with different meaning from [13 ]. A part of the globa l system, whic h is resp onsible of the dynamic infor mation representation of the current situation, was applied to the game of Risk and tested thanks to a pr o totype implemented in Ja v a [1 0]. Ho wev er, we p ostula te that some parts of the architecture and, at a deeper level, some parts of the agents were independent of the sub ject used as application. Therefor e , the ob jective at present is to connec t this part to the o ther pa rts, that we prese n t latter in this pap er, and to test the w ho le system on v arious domains, a s RoboCup Rescue [11] a nd e-lear ning. W e fo cus her e on the mo delling of the information represe ntation part of the system that we intend to us e it in a cris is manag e men t supp ort system. The pap er b egins with the pre s entation of the global system a rchitecture. The cor e o f the system is constituted b y a multiagen t system (MAS) which is structured on three mu ltiagent layers. Then, in section 3, we explain the wa y we formalise the en vironment state and we ex tract information related to it, which are wr itten in the for m o f sema n tic fea tures. The la tter cons titute data that feed the system p ermanently and that carry information ab o ut the cur rent situation. The semantic features are handled b y factual agents and are compar e d the one with the other using spe cific o ntologies [2 ]. F actual agents, tha t comp ose the first lay er of the core, a re presented there- after in sectio n 4. Ea ch a gent carries a se ma nt ic feature a nd a ims to reflec t a partial pa rt of the situation. W e pr esent their structures and their behaviours inside their or ganisation using internal automaton a nd indica tors. Finally , we pr esent a sho rt view ab out the game of Risk test in which w e describ e the mo del applicatio n and the b ehaviour o f factual agents. 2 Arc hitecture of t he Decision Supp ort System The role of the de cision supp ort system (DSS) is to provide a decision- making suppo rt to the acto rs in order to assist them dur ing a crisis case. The DSS a llows also ma nagers to an ticipate the o c c ur of po ten tial inciden ts thanks to a dyna mic and a con tinuous ev aluation o f the curr ent situa tion. Ev alua tion is rea lised by comparing the curr ent situation with past situations stored in a scena r ios base. The la tter ca n b e viewed as one part of the knowledge we hav e on the spe cific domain. The DSS is co mpo sed of a cor e and three parts which a re connected to it (figure 1 ): • A s e t o f us e r-computer interfaces a nd an intelligent in terface allow the c ore to commun icate with the environmen t. The intelligen t in terface controls and manages the access to the core o f the authenticated use r s, filters entries information a nd provides ac tors with results emitted by the system; • An inside query MAS ensures the in teraction b etw een the cor e a nd w orld information. These infor mation repr e sent the knowledge the core need. The knowledge includes the scenarios, that a re stored in a scenarios base, the ontologies of the domain and the proximity measure s ; • An outside query MAS ha s as role to provide the co re with informatio n, that a re sto red in netw ork distr ibuted infor mation sy s tems. Fig. 1. General Architecture of the D SS The cor e of the decision supp or t system is made o f a MAS which is structure d on three lay ers. The latter contain sp ecific agents that differ in their ob jectives and their communications way . In a first time, the system desc rib es the semantic of the current situa tio n thanks to data collected from the environmen t. Then it analyses p ertinent informatio n ex tracted fro m the scena r io. Finally , it pr ovides an ev aluation of the cur rent situation and a decision supp ort using a dynamic and incremental case - base rea s oning. The three lay ers of the co re are: • The lowest lay er: factual agents; • The intermediate layer: synthesis age nts; • The hig hest layer: prediction agents. Information are coming from the en v ir onment in the form o f s e ma nt ic fea- tures without a prio ri knowledge of their imp or tance. The ro le of the fir s t layer (the low est o ne) is to deal with these data tha nks to factual ag ent s and let emer- gence detect some subsets o f all the infor mation [7]. Mo re prec is ely , the set of these agents will enable the app ear ance of a glo bal behaviour thanks to their int eractions and their individual op erations. The s ystem will extr act therea fter from this b ehaviour the p er tinen t infor mation tha t r epresent the salient facts of the situa tion. The role of the s ynthesis agents is to deal with the agents emer ged fro m the first layer. Syn thesis agents aim to create dyna mically factual agents clusters according to their evolutions. Each cluster r epresents an o bserved scenario. The set of these scenar ios will b e compared to past ones in o rder to deduce their po tent ial consequences. Finally , the upp er layer, will build a co ntin uous and incre mental pr o cess o f recollection for dynamic situations. This layer is c o mpo sed of pr e diction agents and has as g oal to ev a luate the deg ree of re semblance b etw een the current sit- uation a nd its asso ciate sce na rio contin uously . Each prediction ag ent will b e asso ciated to a scenario that will bring it clos e r, from semantic po in t of view, to other scena rios for which we know a lready the co ns equences. The res ult of this compariso n constitutes a s uppo rt information that can help a mana ger to make a go o d decision. Fig. 2. Architecture of the Core 3 En vironment Study and Creation of Seman tic F eatures 3.1 Situation F ormalisation T o for malise a situation means to create a forma l sys tem, in an attempt to capture the essential features of the real-world. T o realise this, w e mo del the world as a collection of o b jects, where each one holds s o me prop erties. The aim is to define the environmen t ob jects following the ob ject pa radigm. Therefore, we build a structural a nd hierar chical form in or der to give a meaning to the v arious relatio ns that may exist b etw een them. The dynamic change of these ob jects states and more still the in teractions that could b e entrench ed betw een them will provide us a snapshot descriptio n o f the environment. In our co ntext, information are decomp os ed in atomic data where each one is assoc iated to a given o b ject. 3.2 Seman ti c F eatures A semantic feature is a n elementary piece of information coming from the envi- ronment a nd which r epresents a fa c t that o ccurred in the world. E ach semantic feature is related to a n ob ject (defined in section 3.1 ), and a llows to define all or a part of this ob ject. A semantic feature has the following form: (key , ( q ual if ication, v al ue ) + ), where key is the describ ed ob ject and ( q ual if ication, v al ue ) + is a set of co uples for med by: the qualification o f the ob ject and its asso ciated v alue. As example of a semantic feature rela ted to a phenomenon ob ject: (phenomenon#1, type, fire, lo cation, #4 510, time, 9:3 3). The ob ject de- scrib ed by this seman tic feature is phenomenon#1, and ha s as qualifications: t yp e, lo cation, a nd time. The modelling o f seman tic features makes it p ossible to obtain a homog e- neous structure. This homogeneity is of pr ima ry imp or tance because it allows to establis h compariso ns b etw een these data. The la tter are mana ged b y fac- tual ag ent s, where each one ca rries one sema ntic feature and of which b ehaviour depe nds o n the type of this infor mation. According to FIP A co mm unicative acts [6], the a g ents must shar e the same language a nd vocabular y to communicate. The use o f s e ma nt ic features in com- m unications pr o cess implies to define an o nt ology . Inside the r e pr esentation lay er (the first lay er of the sy stem), agents evolv e by compar ing their semantic fea tur es. These comparisons allow to establish se- mantic distances b etw een the agents, and are computed thank s to pr oximit y measures. W e distinguish thre e t yp es of proximities: time proximit y , s patial proximit y and sema n tic proximity . The globa l proximit y multiplies these three pr oximities together. The measurement of a semantic proximit y is related to ont ologies . Whereas time proximit y and spatial proximit y are computed according to sp ecific functions. Proximities computation provides v alues on [ − 1 , 1] and is asso cia ted to a scale. The reference v alue in this scale is 0 that mea ns neutra l r elation b etw een the tw o co mpared semantic features. Otherwise, we can define the scale as follow: 0.4=Quiet Close, 0.7= Clo se, 0.9=V ery Close, 1=E qual. Negative v alues mirro rs po sitive o nes (repla c ing clos e by different). 4 F actual Agen ts 4.1 Presen tatio n and Structure of a F actual Agen t F actual a g ents a re hybrid agents, they are b oth cognitive a nd reactive agents. They hav e therefore the fo llowing c haracter istics: r eactivity , pro activeness and so cial a bility [14 ]. Such an ag ent repre sents a fea ture with a s emantic character and has also to formulate this character feature, a b ehaviour [4]. This be haviour ensures the a g ent activity , proa ctiveness and communication functions. The r ole o f a fac tua l a gent is to manage the sema ntic fea tur e tha t it carries inside the MAS. The a g ent m ust develop to acquir e a dominating place in its organis ation and cons equently , to make pr ev ail the semantic category which it represents. F or this, the factual agent is des ig ned with an implicit goal that is to g ather a round it as m uch frie nds a s possible in or der to build a cluster. In other words, the purp o se of the ag ent is to add p er manently in its a cquaintances net work a gr e at num b er of s emantically close ag ent s. The cluster formed by these agents is recog nized by the sys tem a s a s c e nario of the current situation and for which it can bring a p otential cons equence. A cluster is formed only when its agents ar e enoug h stro ng and c onsequently they ar e in an a dv anced state in their automato n. Ther efore, the goal o f the fa c tua l a gent is to reach the a ction state, in which is supreme and its information ma y b e regarded by the system as r elev ant. Fig. 3. Stru ct ure of a F actual Agent An internal automaton des crib es the b ehaviour and defines the actions of the agent. Some indicators and a n acqua intances netw ork allow the automaton op eration, that means they help the age nt to pr ogres s inside its automaton and to e xecute actions in o rder to reach its go al. These c ha racteristics e xpress the proactiveness of the agent. The a cquaintances netw ork contains the addresse s of the friends ag ent s and the enemies agents used to send messa ges. This netw o rk is dynamically con- structed a nd p ermanently up dated. Agents are friends (enemies) if their seman- tic pr oximities a re strictly p ositive (negative). 4.2 F actual Agent Beha viour Beha viou ral Automaton The internal b ehaviour of a fac tual ag ent is de- scrib ed by a generic augmented transition netw ork (A TN). The A TN is made o f four sta tes [3] (quoted ab ov e) linked by tra nsitions: • Initialisation state: the agent is created and enters in activities; • Delib er ation state: the agent sear ches in its ac quaintances allies in order to achiev e its go als; • De cision state: the agent try to co ntrol its enemies to b e re info r ced; • A ction state: it is the state-g oal of the factual agent, in which the latter demonstrates its streng th by acting and liquida ting its enemies. Fig. 4. Generic Automaton of a F actual Agent A TN tra nsitions a re s tamp e d by a set of conditions and a sequence of actions. Conditions are defined as thresholds using internal indica tors. The agent must v alidate thus one of its o utg oing curr e n t state tr a nsitions in order to pa ss to the next s tate. The actions of the agents may be an enemy agg ression or a friend help. The choice of the actions to p erfo r m dep end b oth o n the type of the age n t and its p osition in the A TN. F actual Agen t Indicators The dynamic measurement of a n a g ent b e haviour and its state progr ession at a given time ar e given thanks to indicator s. Thes e characters are significant para meters that des crib e the activities v ar iations of each age n t and its s tructural evolution. In other words, the agent state is sp ecified by the set of these sig nificant characters that allow b oth the description of its current situation and the prediction of its future b ehaviour [4] (quoted ab ov e). F actual agent has five indicators , which ar e ps e udo Position (PP), pseudoSp eed (PS), pseudoAcceler ation (P A), sa tisfactory indicator (SI) a nd constancy indi- cator (CI) [8]. The “pseudo” prefix means that thes e indica tors ar e not a real mathematical s p eed or accele ration: we chose a constant in terv al o f time of one betw een tw o evolutions o f se ma nt ic features. PP repr esents the current po sition of an agent in the agent representation spac e. PS ev aluates the P P evolution sp eed a nd P A means the PS ev olution estimatio n. SI is a v alua tion o f the suc- cess of a factual agent in rea ching and staying in the delibera tio n state. This indicator measur e s the sa tisfaction deg ree of the agent. Whereas, CI repre sents the tendency of a given factual agent to trans it b oth from a sta te to a different state and fro m a state to the sa me s tate. This allows the stability mea surement of the ag ent b ehaviour. The co mpute of these indicators is accor ding to this formulae wher e valPr ox- imity dep ends on the catego ry of a given applicatio n factual agents: P P t +1 = valPoximity P S t +1 = P P t +1 − P P t P A t +1 = P S t +1 − P S t PP , PS and P A re pr esent thresholds that define the conditions of the A TN transitions. The definition o f this c onditions are sp ecified to a given application. As shown in the previous fo rmulae, only PP is sp ecific. How ever, PS and P A are generic and are de duce d fr o m PP . SI and CI a re also indepe nden t of the studied domain and ar e computed according to the agent mov ement in its A TN. 5 Game of Risk Use Case The fir st lay er mo del has b een tested on the game of Risk. W e chose this game as applica tion not only b eca use it is well suited for crisis ma na gement but a lso we apprehend the elements and the actions on such an en vironment. Moreover we hav e an expert [8 ] (quo ted ab ov e) in our tea m who is able to ev a lua te and v alidate r e sults at any moment. As result, this test proved that this mo del allows the dyna mic information representation of the current situa tion thanks to factual agents o rganisa tion. Moreov er we could study the be haviour and the dynamic evolution of these agents. Risk is a stra tegic ga me which is comp ose d of a playing b oar d repres entin g a map o f forty-t wo territories that are distr ibuted on s ix con tinent s. A player wins by co nquering all territor ies or by completing his secr et miss ion. In turn, each play er receives and places ne w a rmies and may attack adjacent ter ritories. An attack is one or more battles fought with dice. Rules, tricks and stra tegies are de ta iled in [12]. The r epresentation layer of the s ystem has as r ole to simulate the game un- winding a nd to provide a semantic instantaneous des c ription of its curre n t state. T o ac hieve this task, w e be gan by identifying the different ob jects that define the g ame bo ard (figure 5) and which ar e: terr ito ry , play er, ar my and continen t. Continen ts and territories are r egarded as desc r iptions of a per sistent situation. Whereas, armies and play er s are activities resp ectively observed (o ccupying a territory) a nd dr iving the actions . Fig. 5. Class Diagram for the Game of Risk R epresentatio n F rom this mo del we distinguish tw o differen t t yp es of semantic features: a play er t yp e and a terr itory type. F or ex ample (Quebec, play er, gr een, n bAr mies, 4, time, 4) is a territor y sema nt ic feature that means Queb ec terr itory is owned by the green play er and has four armies. How ever, (blue, nbT errito ries, 4 , time, 1) is a player semantic featur e that s ignifies a blue play e r has four territories at step 1 . The first extracted semantic featur es of the initial sta te o f the game cause the creation of factual agents. F or example, a semantic feature as (red, nbT erritor ies, 0, time, 1) will cause the cr eation of red play er factua l agent. During the game prog ression, the ent ry of a new sema nt ic feature to the system may affect some agents s ta te. A factual agen t of type (Alask a, play er, red, nbArmies, 3, time, 10 ) b e c ome (Alask a , play er, red, nbArmies, -2, time, 4 9) with the e ntry of the semantic feature (Alask a, play er , red, n bArmies , 1, time, 49). Alask a agent sends messages containing its semantic feature to all the other factual a g ents to inform them ab out its change. The other agents compare their own information with the received one. If an agent is int erested by this message (the proximity measure b etw ee n the tw o semantic features is not n ull) it up dates its semantic feature acc o rdingly . If the red play er owned GB b efore the semantic feature (GB, player, blue, nbArmies, 5, time, 52), b oth red play er a nd blue player will r eceive messa ges b ecause of the change of the territory owner. If we take aga in the preceding example (Alask a territory), Alask a ag ent c o m- putes its new PP (v alPr oximit y). The computation of v a lP roximit y in our case is given by: num ber of armies (t) - n um b er of armies (t-1) e.g. here v alP r oximit y = 1-3 = -2. P S and P A are deduced thereafter fro m PP . The agent verify then the pr edicates o f its curr ent state outgoing transitions in order to change state. T o pass from Delib er ation state to De cision state for example the PS must be strictly po sitive. During this trans itio n, the a gent will send a Supp ortMessage to a friend and an A gr essionMessage to an enemy . 6 Conclusion The pap er has presented a decision supp ort system which aims to help decis ion- makers to analyse and ev alua te a current situation. The core of the system r ests on an a gent-orien ted multila yer arc hitecture. W e hav e de s crib ed here the first lay er which aims to provide a dynamic informa tion representation of the current situation and its e volution in time. This par t is mo delled with an original in- formation repr esentation metho do logy which is ba sed o n the ha ndle of semantic features us ing a factual agents or ganisatio n. The mo del of the first lay er was applied on the g ame of Ris k. Results provided by this test corres po nd to our attempts, which consist on the dynamic repre- sentation o f informa tion. This applicatio n allow ed us to track the b ehaviour of factual agents a nd to understand their par ameters whic h are the most accu- rate to characterise information. Mor eov er , we consider that a great pa r t of the system is generic and may b e ca rried in to other fields. Curre n tly , we int end in a first time to connect the represent ation lay e r to the t wo other and to apply thereafter the whole s y stem on more sig nificant domains as RoboC up Res cue and e-lea rning. References 1. Bouk achour, H.: Syst ´ eme de v eille pr´ eventiv e p our la gestion de situations d’urgence: une mo d´ elisati on par organisation d’agen ts. Application aux risques industriels. PhD Thesis, U niversi ty of le Havre, F rance (2002) 2. Bouk achour, H., Galinho, T., Person, P ., Serin, F.: T o w ards an architecture for the representatio n of d ynamic situations. In Las V egas IC-A I 659-664 (2003)q 3. Cardon, A .: A multi-agen t mo del for coop erative communications in crisis man- agemen t system: the act of comm unication. I n 7th Europ ean-Japanese Confer- ence on In formation Mo d eling and Kn o wledge-Based Simulation 111-123 (1997) 4. Cardon, A .: Mo d´ eliser et concevoir une machine p ensante: approhce de la con- science artificielle, V uib ert, ( 2004) 5. Durand, S.: Repr ´ esen tation des p oints d e vue multiples dans une situation d’urgnece: u ne mod´ elisa tion par organisation d’agents. PhD Thesis, U niversit y of Le Havre, F rance ( 1999) 6. http://www .fipa.org, last retrieved March (2006) 7. Galinho, T., Serin, F.: Semantic F eatures And F actual Agents: A Mo del T o Represent Changeable Situations A Mo d el T o Rp resent Changeable Situ ations (2003) 8. Galinho, T., Coletta, M., P erson, P ., S erin, F.: Dynamic Representation of In- formation for a Decision Supp ort System, 8th I nternational Conference on En- treprise I n formation, I CEIS2006, Pa phos, Cyprus, (2006) 9. Kitano, H., T adokoro, S.: Rob oCup Rescue A Grand Challenge for Multiagen t and Intellige nt Systems, American Association for A rtificial Intelligence (2001) 10. Pers on, P ., Bouk achour, H ., Coletta, M., Galinho, T., S erin, F.: F rom Three MultiAgent Systems t o On e Decision Su pp ort System, 2 nd Indian International Confer enc e on Art ificial Intel ligenc e , Pune, India (2005) 11. http://sas emas.org/2 005/c hallen ge.html, last retrieved March (2006) http://w ww.rescuesystem.org /robo cuprescue, last retrieved March (2006) 12. http://ww w.thegamesjournal. com/articles/Risk.sh tm l, last retrieve d Marc h (2006) http://w ww.tuxic k.net/xfrisk/doc/riskfaq.html, last retriev ed March (2006) 13. W o oldridge, M., Jennings, N.R.: Pitfalls of agen t-oriented d evelopment 2 nd In- ternational Conference en Autonomous Agents, p p . 385-391, Minneapolis, (1998) 14. W o oldridge, M.: An Introdu ction to MultiAgent Systems, Wiley , (2002)