Agent-Based Perception of an Environment in an Emergency Situation

Agen t-Based P erception of an En viron men t in an Emergency Situatio n F ahem Kebair, F r ´ ed ´ eric Serin and Cyrille Bertelle ∗ Ab str act —W e are interested in the problem of mul- tiagent systems development for risk detecting and emergency response in an un certain and partially p er- ceived e n vironment. The e v aluation of the current situation pas ses by three stages insi de the multiagen t system. In a first time, the si tuation i s represented in a dynamic w ay . The second step, consists to c harac- terise the situation and finally , it is compared wi th other si milar known situations. In this pape r, we present an information mo delli ng of an observed envi- ronment , that we hav e applied on the Rob oCupRes- cue Simulation System. Information comi ng from the environmen t are formatted according t o a t axonomy and using semantic features. The latter are defined thanks to a fine on tology of the domain and are man- aged by factual age nts that aim to repres ent dynam- ically the current situation. Keywor ds: F actual agent, Multiagent system, Ont ol- o gy, Se mantic fe atur e , T axonomy 1 In t ro duction Recent catas tr ophic disasters hav e bro ug ht ur gent needs for diverse tec hnologies for disaster r elief. Currently , there is an o verwhelming need for b etter informatio n techn ology to help supp ort the efficient and the effective management of the disaster management (also known as emergency re sp o nse). In particula r, ac to rs and a gencies need an assis ta nce to help them to mak e a decision in a fashion time and to b e able to co ordina te their effor ts in a flexible wa y in or de r to preven t further problems or effectiv ely manage the a fter math of a disas ter . Our pro ject is situated in this context and co nsists to develop a gene r ic Decision Supp or t Sys tem (DSS), able to detect a risk in an uncertain and partially p er c eived en vir on- men t a nd to preven t its evolution. The DSS kernel is a m ultiagent system with thre e lay e r s, wher e ea ch one has a spe cific r ole. The r ole o f the lower lay er , that we call the repr esentation lay er , is to repres ent the environmen t state and its evolution ov er the time. The e n vironment is per ceived as a whole o f entities, directly or indirectly ob- serv able a nd of which states change p ermanently . These ent ities are mo deled accor ding to a taxo nomic o rganisa - ∗ Laboratoire d’Informatique, de T raitemen t de l’Information et des Syst ` emes, University o f Le Hav re, 25 rue Phili ppe Lebon, 76058, Le Havre Cedex, F rance. Email: { fahem.k ebair, fr ederic. s erin, cyrille.b ertelle } @univ-lehavre.fr tion and information t hat describ e them are for matted according to a mo del of “sema nt ic features ” , inspired b y the memento design pattern rules [Gamma and al. 1995]. Moreov er, the system appr ehends these information via softw ar e agents (called factual agents) and acco rding to an ontology of the studied domain. The collab ora tion of these agen ts and their comparis o ns with each other, form dynamic agents clusters. The latter a r e compa red by past known sce narios. The final ob ject of the s tudy is to p ermit to preven t the o ccur of a crisis situation a nd to provide an emerg e ncy manag ement pla nning. This mo delling was elab ora te s tarting from the game of Risk [Person 2 005] and tested on the Rob oCupRescue Sim ulation System (RCRSS) [Rob oCupResc ue]. In this pap er, we provide a mo delling o f informatio n extracted from an observed environment in a n emer gency context. Inside the system, informa tion ar e manag ed thanks to factual agents that interact by compar ing each o ther . The mo delling includes a definition of a tax onomy . The latter was applied to the RCRSS environmen t, for whic h we hav e defined an ontology of the do main. The struc- ture of the pap er is as follows: fir st we present the general architecture of the DSS and its internal kernel. Then, we define the taxo nomic o rganisa tio n o f the p erceived envi- ronment. After that, we present the RCRSS environmen t and the ontology o f the domain. Finally , we present fac- tual agents and some tests using gr aphic too ls. 2 Decision Supp ort System The r ole of the Decision Supp ort System is quite wide. In genera l, the pur po se is “to improve the de- cision ma k ing ability o f mana gers (and op era ting p er- sonnel) b y allowing mor e or better decisions within the constraints of cog nitive, time, and economic limits” [Holspace C.W. and al. 199 6]. More sp ecifically , the pur- po ses of a DSS are: • Supplemen ting the decision maker, • Allowing b etter intelligence, desig n, or choice, • F acilitating pro blem solv ing, • Providing aid for non structured decisions, • Managing knowledge. Decision ma kers need quick re s po nses to even ts that take place at a contin ually incr easing ra te and they should incorp ora te an enormo us amount of knowledge such as data, choices and consequences . Also they must have fast access to consis ten t, hig h-quality knowledge to comp ete [Kim 200 5]. In our context, the DSS is used as an emergency man- agement system, able to ass ist actor s in urban disasters mitigation and to preven t them a b o ut p otential future critical consequence s. The system inc ludes a b o dy of knowledge which describ es some a sp ects of the decision- maker’s world a nd that compris es the ontology of the domain and past known scena rios. Figure 1: Ker ne l structure The kernel of the DSS is a m ultiagent system with thr ee lay er s. Agents o f each lay er hav e their own wa y of b ehav- ing and communicating. Representa tion la yer : This layer is comp osed by fac- tual agents and has as essential aim to repres e n t dyna mi- cally and in real time the informa tion of the cur r ent situ- ation. E ach new entering information is dealt by a factual agent that intends to reflect a par tial part of an obser ved situation. Ag ent s interactions and more pr ecisely , aggres - sions and mutual aids reinforce so me agents and w eaken some other. Characterisation lay er : This layer has as aim to gather factual agents, emerged from the precedent lay er, using clustering a lgorithms. W e consider a cluster o f agents, a group of which agents are close from dyna mic and evolution manner p oint o f view. The goa l here, is to form dynamic s tructures, where each o ne is manage d b y a characterisation agent. Prediction la yer : This layer is made up of predic- tion ag ent s. E ach one repres e n ts an obser ved scena rio originally from the current situation. The ta s k of the prediction age n ts is to compare their scena rios by past ones to provide a clos ed one o f which result may b e a po tent ial co ns equence. This mechanism is bas ed on the case base reaso ning , the latter differs from a class ic one by its ability to manage a dyna mic and a n incremental developmen t. 3 T axonom ic Or ganisation of the Studied En vironmen t Our p erception of the en vironment fo cuses on tw o as- pec ts: on the one ha nd, we obser ve the concrete ob jects of the w orld, the c hanges o f their states and their interac- tion. On the other hand, we obse r ve the even ts and the actions that may b e c reated natura lly or artificially . W e hav e defined therefore , three catego ries of ob jects (Figure 2): Concrete ob ject, Action ob ject and Messag e ob ject. Concrete ob ject : Three t yp es o f concrete ob jects are distinguished. The first t ype is the Person o b ject, which represents an actor of the environment. It is the o nly ob ject that has the ability to act a nd to interact and of which b ehaviour a nd state evolution a re usua lly pre- dictable. The se cond t y pe is the Passive ob ject. Tw o sub- categorie s ar e identified: immobile o b jects as buildings and roads netw or ks, and mobile ob jects like the means of tr a nsp ort. The obser v ation of these ob jects is the simplest one, bec ause they do not hav e any b ehaviour. Their o bserv ation is reduced only to the descr iption of their current sta te. The third type is the Mean o b ject. It is created at a given time and for a par ticular pur- po se. Its ex istence duration v aries in time, acco r ding to the ob jectiv e for whic h it is cr e ated. F or example, a car is considered a s a mea n s ince it is driven by a driver, otherwise, it is considere d a s an immo bile ob ject. Action ob ject : This type is divided into activities and phenomena o b jects. Both are crea ted at a given time a nd are limited temp orally without a pr io ry knowledge of the bo unds. P henomena are unpredictable even ts that start at a given time. Their observ ation is the most complex bec ause o f their uncertainties and their rapid evolutions. Activities are the actions s equences p er formed by actors. Generally , they are ordered and emitted for a par ticular purp ose. Message : Messag es represent the interactions be t ween per sons and more precisely the information flows ex- changed b etw ee n the actor s. The impact of a message is not easily mea surable and it dep e nds o n its sender, its receiver and its p erforma tive. If the messa g e is stor ed, its impact will b e deferred. The o bserv ation of a n o b ject in the environmen t may concern a p ersis ten t, tempo rary o r punctual state. A per sistent state can b ecome inv a lid following a rupture. F or example, a building with ten flo or s is a p ersis tent state a s lo ng a s is no t destroyed. A blo ck ed road is also a p ersistent state until it will be unblo cked. How ever, a fire is a temp orar y state, b ecause it is foreseea ble that it will cease , fa ult of combustible. Finally , a punctual state is immediate and instantaneous, like sending a message. Figure 2: T axonomy of the obs e rved e nvironment 4 F orma lisation of I nformation: Appli- cation on the RoboCupRescue Sim u- lation System 4.1 Rob oCupRescue Simulation Pro ject Rob oCupRescue (RCR) is an ann ua l in ter na tional comp etition within th e framework of the Robo Cup [Rob oCup]. This pro ject intends to promote resea rch and development in the disa ster rescue domain by cre- ating a sta nda rd sim ulator a nd forum for re searchers and participator s . RoboCupRes c ue pr o ject intends to s imu- late a urba n disaster caused by an earthqua ke. The sim- ulation disaster integrates v arious asp ects of disa s ters. These includes, fire, ho us ing a nd building da mages, dis- ruption of roads, elec tr icity , water supply , gas, and other infrastructures, movemen ts of r e fug e s, status o f victims, hospital op erations, etc. R CRSS is co mpo s ed by several distributed mo dules: a k ernel, a geogra phic informa tion system, simulators (fire, tr affic a nd collapse simulators), a v ie wer and a n RCR ag ents mo dule. W e are interested in o ur work in the RCR agents mo dule. The w ork con- sists in designing rescue teams that have as missio n to sav e civilians and mitigate disaster co nsequences. The final g oal is to set up a stra tegy planning that p ermits teams co ordination. The picture Figure 3 shows the hiera rch y classes of the R CR disa ster space . Each ob ject in the world has pro p- erties such as its p osition, its shap e ans its state. W e distinguish tw o main o b jects categ ories: moving ob jects and motionless ob jects. Fir st ones represent ac tors o f the disaster world and they are mo delle d by Person ob ject in our taxonomy . The second categor y cons ists of bo th buildings and ne tw orks r oads and they are mode lle d by Passiv e ob ject in the taxonomy . Seven RCR ag e n ts types exist in the RCR simulation world: three plato on agents which are fire brigade, p olice force and ambulance team, three ce nter a g ents whic h are fire station, po lice office and ambulance c e nter and civil- ian agents. W e will not develop the b ehaviours o f the la t- ter, b ecause they a re simulated indep endently with the other s imulators. Ea ch RCR a g ent has a pa rtial knowl- edge of the whole e n vironment state. This knowledge is upda ted thanks to tw o capacities : visual a nd auditory capacities. These capa cities p ermit agents to receive in- formation send by the kernel of the simulator each cy- cle (one second in the s imulation which repres ent s one min ute in the reality). Agent centers represent in reality per sons inside, so they can only s ee their surr o unding a rea of the world and exchange messa ges with other ag e n ts. The r ole of these centers is to co ordina te the communi- cation b etw een the three ag ents types. Platoo n agents hav e more c a pacities, they can act by p er forming seven different a ctions: rescue, load and unloa d actions for a m- bulance team ag ents, extinguish for fir e brigade ag ent s, clear for po lice force agents and move for a ll agents. Figure 3: Cla ss hiera rch y of the RCR ob jects in the dis- aster space 4.2 On tology of the Domain The definition of the ontology of the domain is the result of the taxonomy application on the R CR simulation en- vironment. T he determination of the concepts is based on the ob ject mo delling of the RCR environment and resp ects the ta xonomic or ganisation. The next picture (Figure 4 ) shows the ontology that we have implemented using prot´ eg´ e [Pr ot´ eg´ e ]. The abstract class Ob ject is situated o n the top level o f Figure 4: Ontology of the Rob oC upRes cue e nvironment the classe s hierarch y . Each ob ject of the en vironment has a type and is lo calised in time and spa c e . W e have assigned therefor e to O b ject cla ss a type, a time and a lo calisation attributes. In the seco nd level, three clas ses inherit the Ob ject class. Two abstract classes: ActionOb- ject and Concre teO b ject, and a concrete class Messag e. ActionOb ject clas s is the sup er c lass o f P henomenon a nd Activit y classes . The first o ne is the sup erclass of Fire, Break, Injury and Blo ck ade classes and has an a dditional attribute intensit y . The la tter re presents the intensit y and the pro gressio n degr ee of the phenomeno n. F or ex- ample, a fir e may have the following intensities: sta rting, strongly and extremely strongly . Activity class is the s u- per class of Load, Rescue, Unloa d, Extinguish, Mov e and Clear which are the RCR agents a ctions defined ab ov e . This class has tw o additional attributes: a ctor and tar- get. Actor attribute takes a s v a lue an RCR agent name and target a ttribute has as v alue a Concrete o b ject name that may b e: a building, a ro a d, a civ ilian, etc. ConcreteOb ject class is the sup ercla s s of the concrete classes: Person, PassiveOb ject a nd Mean class es. Person class has three a dditional attributes: buriednes s, dama ge and hitPoint. The fir st one shows how muc h a p erso n is buried in the colla pse buildings. The seco nd o ne s hows the necessity of medica l treatment. The last one shows the health level, a p ers on in go o d health has a hitPoin t = 10000 , and 0 when his is dead. PassiveOb ject and Mea n classes has only the inherited attributes. Finally , Message cla ss is a concr ete cla s s and ha s t wo ad- ditional attributes: re ceiver and se nder . In Robo CupRes- cue, a message conten t has the following forma t: “ac- tion name ob ject name” . F or example, “clear roa d#ID”, “extinguish building#ID”, or “ rescue civilia n#ID”, etc. The lo calisation attribute means therefore more precisely , the lo calisa tion o f the target ob ject in the mess age con- ten t a s the road#ID, building#ID and civilian#ID in these examples. 4.3 Seman tic F eatures Information coming from the environment ar e written in the form o f semantic features. The latter will b e man- aged therea fter by factual a gents in the r epresentation lay er . The idea to use a sema ntic feature is inspired from the memento design pattern and consis ts to store infor - mation, that des crib e the internal sta te of an obser ved ob ject orig ina lly from the taxonomy . The structure of a semantic featur e is gener ic a nd comp osed by a key and a set o f co uples < q ua lifier,v alue > . The key is defined from the taxonomy and the qualifiers ar e defined fro m the ontology . In Rob oCupRescue, infor mation are sent by RCR age n ts each cycle and may b e visua l or a uditive informa tion. The system trea ts these data in o r der to extract the im- po rtant o nes. F or example, an RCR agent who sends an information descr ibing an intact building will not be taken into a c count. Howev er , an infor mation a b o ut a burning building is int eresting, the system interprets it and cr eates therea fter new semantic features, r elated to ob jects defined by the taxonomy . As exa mple, a Building#14 has a pr op erty “ fieryness = 25” , this means that a fire has just star ted in this building. The s ystem creates therefor e , a se ma nt ic feature: (Phenomenon#14 , t yp e, fire, intensit y , s tarting, lo calis ation, 2 0 | 25, time, 7). This sema n tic feature is rela ted to a pheno menon ob ject, that means a fire is lo ca ted in 20 | 25 co ordina tes a t the seven th cycle of the simulation. In the ca se of a n au- ditive infor ma tion, the system cr eates semantic features according to messag es co nt ents. F or example, an RCR agent se nds a message ” clear roa d#15”. F r om this mes - sage, a semantic feature (Phenomenon#22 , type, blo ck- ade, in tensity , unknown, lo ca lis ation, 30 | 4 0 , time, 11 ) is created. This semantic feature is r elated to a blo ck ade phenomenon, a priory we do not know the intensit y o f the blo ck a de, but we can determine the co or dinates of the blo ck ed road from the world map, using its identifier (15). Thu s, by tr eating the messa ges a nd the visual informa- tion sent by R CR ag ent s, the s ystem gathers the par tial knowledges o f these agents to build a global k nowledge that can provide a clea r er idea a bo ut the s ituation. Semantic featur e s are related with each other, that mea ns they have a se ma nt ic dep endencies. W e defined therefore proximit y meas ur es in order to co mpare b etw een them. The proximit y v a lue is co mprised be tw een [-1,1]. Two se- mantic fea tures a re opp osite in their sub jects if the prox- imit y measur e is nega tive, they ar e closed if it is p ositive and indep endent if it eq uals zero. More the pr oximit y is near to 1 (-1), mo re the tw o semantic features a re clo sed (opp o site). W e distinguish three t yp es of proximities: a semantic pr oximit y which is determined thanks to the on- tology , a spatial and a time proximities tha t are r elated to sp ecific sca le s. As exa mple, a break and a blo ck are closed semantically , b ecause if a building is broken, the nearest road w ill b e cer tainly blo cked. Mor eov er , to give more prec is ion to this co nfrontation, w e compare the lo- calisations a nd the times o f observ a tion of the tw o even ts. If they are distant, we consider the tw o even ts ar e inde- pendent, and inv er sely . 5 F a ctual Agents of the Represen tation La yer 5.1 Structure and Role The r epresentation lay er is co mpo sed by factual agents. Each agent aims to re pr esent a partial part of the ob- served situation, thanks to the semantic feature that it carries . Figure 5: Internal str ucture of a fac tua l agent F actual agent is a reactiv e and proa ctive agent [W o oldridge, 20 02]. Its reactivity is ensur e d by a gener ic int ernal b ehavioural automaton of Augmented T ransition Net work (A TN) type [W o o ds 1970]. This automaton is comp osed by fo ur states [Ca rdon 2004]: initialisa tion, delibe r ation, decision and a ction. A TN T ra nsitions a re stamp ed by a set of co nditions and a seq uence of ac- tions. Conditions r epresent thre s holds, defined a ccording to three internal indicato rs o f the agent, which ar e: Pseu- doPosition (PP ), P seudoSp eed (PS) a nd PseudoAcceler a - tion (P A). The agent has tw o other indica to rs: a satisfac- tion indicator a nd a constancy indicato r, which r e present resp ectively the sa tisfaction degree of the ag ent ab out its progre s sion and its s ta bilit y in its A TN. The definition of these indicators a llow the factua l age nt to prog ress in its A TN, this characteristic ensures the pr oactivity of the agent, of which purp ose is to a chiev e the most imp ortant state, that is the actio n state. In addition, the factual agent is a so cial ag ent . It in teracts with the other agents in order to form a coalitio n with other ones, this p ermits it to acquire mor e force a nd p ow e r . The agent can also b e attack ed by other ones, with which it is opp osite seman- tically . The list of o ppo sites a g ents a nd closes agents is stored in an acquaintances net work, which is constr ucted and updated dynamically . 5.2 T ests and Graphic T o ols W e started to make tests o n a part of the ontology . W e lo calised our tests esp ecially on the detection of the dif- ferent ev en ts signalled by the RCR agents and the actions that they per form. Figure 6: View of the RCR disaster spac e W e hav e des igned some graphic to ols in or de r to follow and study the evolution of the factual agents. The graphic to ol is comp osed by a gr id that shows in real time po ints flow repr esenting factual a gents. Age n ts a r e pro jected on three axis: P P , P S and P A. F a ctual a g ents progre s s extremely quickly , so it is too har d to follow their evolution. W e hav e crea ted therefore, a n interactive in- terface (agent interface). This int erface has tw o essential functionalities. The fir st one p ermits to select a g iven factual agent and to show a ll its information: its seman- tic feature, its curr e n t state and its curr ent indicato r s v alues. The s econd one p er mits to fr e eze all the factual agents at a given time a nd to r eanimate them thereafter. This allows us to obtain an instantaneous view of all the agents during their evolution a nd to study consequently , information ab out any agent. Picture Figure 6 sho ws an instan ta neous image o f the cur- rent situation of the R CRSS disaster space in the eig hth cycle o f the simulation. Infor ma tion shown in the table, in the right, ar e rela ted to the blue building, that is burn- ing. A new factual agent, c arrying the s e ma nt ic feature (Phenomenon#670 68017 , type, fir e, intensit y , starting, lo calisation 2 29891 00 | 37 55100, time, 8), is crea ted a nd upda ted according to infor ma tion sen t by the fire briga de agent, situated just near to the building. This factual agent is re pr esented b y the gr e en ellipse in the grid and has as co ordinates (PP=2 07,PS=3 ,P A=1). In the ag ent int erface, we can see all informa tion ab out this ag ent, Figure 7: Gr aphic to ols to visualise factual ag ents notably , its indicato r s and its state which is the decision state. W e note, that all indica tors ar e str ictly p ositive and the agent is in adv anced state in its A TN. This means the ag ent ha s acquire d imp or tance and the even t that it represents is more and more sig nificant. This evolution is the r esult of informa tion sent by the fire br ig ade a gent and the interaction o f the factual a gent with other fac- tual agents. The latter carry o ther re la ted info r mation, that can b e messa ges announcing the fire, o r a ctions p er - formed to extinguish it. 6 Conclusion This pap er ha s presented an information mo delling of a per ceived environment in a n emerg ency co ntext. This mo delling is used to represe nt the evolution of the cur - rent situa tion thanks to factual age nts. Our fina l goal is to build a generic mutliagen t system that intends to detect a r isk a n to dea l with it. Information entering to the s ystem are structured in the form of semantic fea- tures. The latter are defined thanks to a taxonomic or- ganisation and to an ontology related to the domain. W e choose the RCRSS as a pplication to test this mo delling . W e hav e implemented therefor e , the ontology o f the stud- ied domain and s ta rted the representation of the R CRSS disaster s pa ce state, using a part of the ontology . This test allowed us to study the b ehaviour of factual agents and esp ecially A TN thresho lds and pr oximit y measur e s that are very de p enda nt on the applica tion and that re- quire more co n trol of the environmen t in order to v alidate them. Our future work consists in finis hing b oth, the im- plement ation o f the o n tology and the repres ent ation lay er that ar e still under r ealisation. W e have the inten tion thereafter, to co nnect this layer with the characteris ation lay er in order to test some observed scenarios of which definition constitutes also a future sub ject of study . References [Cardon 2004] Cardon A.: Mod´ eliser et concevoir une machine p ensante: appro hce de la conscience ar- tificielle, V uib ert, (2004). [Gamma and al. 1995 ] Gamma E., Helm R., Johnson R. and Vlissides J.: D esign Patterns : Elements of Reusa ble Ob ject Oriented So ft w are. Addison- W esley , 395 pages, (1995 ). [Holspace C.W. and al. 1996 ] Holspace C.W. and Whin- ston A.B.: Decision suppo rt systems. New Y ork: W est publishing co mpany (1996 ). [Kim 2005] Kim J.H.: K nowledge-bas ed decision suppor t systems and their future in Knowledge Manag e - men t Systems, December 14 , (2 0 05). [Person 200 5] Person P ., Bouk achour H., Coletta M., Galinho T., Serin F.: F rom Three MultiA- gent Systems to One Decisio n Supp ort System, 2 nd Indian In ternational Confer enc e on Artificial Intel lig enc e , Pune, India (2005 ) [Prot´ eg´ e] http://protege.sta nford.edu/ , last retreived March (200 7). [Rob oCup] ht tp://www.rob o cup.o rg/, last retrieved March (200 7). [Rob oCupRescue] http://www.robo cuprescue.org/cue, last retrie ved Mar ch (2007). [W o o ds 1970 ] W o o ds A.: T ra nsition Netw or k Gra mmars for Natur al Lnguage Analysis, W. A. WOODS, Harv ard Universit y , Ca m bridge, Massach use tts (1970). [W o oldridge, 20 02] W o o ldr idge, M.: An Introduction to MultiAgent Systems, Wiley (2 002). F AHEM KE BAIR is a PhD s tuden t in c omputer science since 2 006. His application domain co ncerns Agent-Based Soft ware Engineering . FREDERIC SE RIN r eceived his P h.D. degree in Ob ject- Oriented Simulation in 1996 from Universit y of Ro uen. Dr. Ser in is c urrently Asso cia te Professo r at Universit y of Le Havre. His curr ent applicatio n domain now concerns Agent-Based Softw ar e Enginee ring. CYRILLE BE R TELLE is professor in computer science in Le Havre University and develops research activities in complex systems mo deling. He is one of the co-dir e c tors of a r esearch lab ora tories amalg amation whic h pr omotes the Sciences and T echnologies in Infor mation and Com- m unication ov er the Haute-Nor mandie in F rance.