Agent-Based Social Information Gathering on Internet

Eun-Seok LEE Roberto OKADAand Norio SHIRATORI Dept. InformationEngineering GraduateSchool of InformationSciences SKKU(Sung-Kyun-KwanUniversity) TohokuUniversity Seoul, Korea Sendal. ,Japan

Background difforent users such that whenone agent alone cannot satisfy user request it asks for advices or help to other IJA it The popularization of computers and the Internet have trusts. produced an explosion in the amountof inform~ttion, ntak- Wedivide.d the impleme.ntation task in two phases. In the ing difficult to find the.hi. The. problem with oxisting tirst phase, wefocus on distributed intelligence with the in- information-locating support systemns is that ahhoughthey troduction of several agents and the negotiation aspects in allow user to search througha large, nunfl)e.r of information such a way. that this cooperative agents showan intelligent sources, they provide very limited capabilitie.s for locating. bchaviour. combining and processing informations. The load of find- In the second phase, we focus on autonomousintelligence, ing informationis still on the user. i.e. makingeach of the agents more "’intelligent", where In order to support the user in finding information in we’ll investigate more powerful real-time planning and such environments, we propose, what we called a "com- le.arning algorithms. pletely agent-based framework for Information Gather- ing", named Cooperative Agent Society for In.formation GatheT~ngor CASte: in short. This CASIa is b;Lsed on The prototype the integration of collaborative, interface agents, domain- The multi-alAent system under construction is implemented specific mediators and cooperative informatiou ~mrces. in C and using the libwww. It uses Netscape as a web browser and user interface, mxdthe agents run ‘as sepa- Architecture: rate processes. This system incorporates one UAper user. We dividethe necessary functionalities between the. user several primitive Man(meta-index like) , and the existing andthe distributed information sources basically in three web(including search engines e.g Lycos, Inh)Seek and the typesof agents: documentservers distributed all over the Internet). When (1) User Agent(UA) u,s er-specific agents: one. per user. the. user invokes his personal UA, it opens a personal log It takes care. of user preferences, managepersonal infor- file. to keep meta-information about the documentswhich mation and acts ,as the user’s electronic personal ~ssis- the user flints interesting. This UAalso maintains a talfle tmtt. These age.nts are. sinfilar to the existing interface of "trusted peers", their location and their trust relation- agents, which perform tedious, repetitive and time con- ship - which is updated based on user feedback ~ffter his sumingtasks on behalf of the user or act ,as an abstract- evaluation of retrieved documents. Whenthis UAreceiw,s ing interface bet~.en the. user and the low-level details. the query as a set of keywordsfrom the. user, it first se- In our framework, UAwouhl dialogue, with the user in lect the agents with higher trust values and send them order to acquire user reque.sts, helping him fornmlating the query. The selected UAslook at their per~nal log file the proper query (2) Machine Agent (MA)inf ormation- for documentsretrieved in the past related with the query specific agents: attached to the information sources, con- in question, while the selected Manschoose somesearch trolling access to the information they provide. By do- engines to send this request. This engine selection is done. ing so, negotiation capability, security and consistency of base.d on a trust tahle with the list of knownsearch engines, data can be ensured (3) Manager (Man}tas k-specific their location and a weight indicating howsuccessful they agents: they stand in betweenthe. provide.rs and consumers were in the past answering about this query. Mansreceive of information or services. In this franmwork, they have pointers to candidate solutions from the selected search en- Domain-specific knowledge and plan howto satisfy user gine.s, re.-order themand return this orde.red list of pointers requirements in their domain of expertise. This type of to candidate, solutions as an HTMLdocument. This doeu- agents have been researched at different levels of complex- me.at is presentedto the user, he selects one pointer causing ity, from ve~" simple routers for supporting communica- the documentto be retrieved. The user evaluates this doc- tion to complexmediators with task division, ,allocation ument and provides a positive or negative feedback. This and planning capabilities. causes the update of the weights of the trust relationships of this UAto the agent which proposed this solution and In thispaper, we focuson twoaspects: (I) the UA wlfich the trust weights of the Manto the search engines as well (i)monitors user actions for finding information, (ii)learns as the user’s personal profile. userpreference and keeps it as a personaldata in user’s personalprofile, (iii) builds a trust relationship withother Acknowledgement: agentsbased on pastexperiences and (2)the. social aspect The authors would like to tllank Hideto Kihara for his i.e. the mechanismfor interaction between UAowned by contribution in the prototype implementation.

A Multiagent Meeting Organizer that satisfies Soft Constraints Joo-Hwee Lira t, Jiankang Wut f~~ Siet-Leng Lag

tRWCPNovel ISS Laboratory ~Center for ComputerStudies Institute of Systems Science Ngee Ann Polytechnic IIeng Mui Keng Terrace, Singapore 119597 535, Clementi Road, Singapore 599489 Email: (joohwee, jiankang}@iss.nus.sg Email: [email protected]

As a distrihuted resour(~ allocation problem, meet- be conveniently expressed as preferrences (soft con- ing sdmduling is a tedious and time-consuming pro- straints) since satisfaction-seeking is more realistic cess. This paper proposes a multiAgent MEETing than optimality-seeking. Indeed, in fuzzy decision organiZER (AMEETZER)that respresents and rea- making, the decision set is the intersection (denoted sons with soft constraints related to the meeting at- as A) (or more general, confluence) of the goals and t~mdees and resources. An AMEETZERaccepts call- the constraints, for-meetingrequest fron| its user(host) and conmmni- cateswith otherAMEETZERs of the proposedatten- ItDCX’) = m~xb,G(X) ^ t,c(X)} 0) dees and withthe agentmanaging the meetingrooms where D, G, C are decision set, goal fuzzy set, con- to arriveat somecommonly free time slot, taking into straint filzzy set respectively and I~D,PC,, PCare ttmir ~u:countthe pre-specified and dynmnicallycreated hard membership fimctions respectively, X" is the optimal and softconstraints of all attendeesand meetingre- solution in tim domain over whic~ X ranges. sources.Whcn necessar);the AMEETZERcould per- In AMEETZER,conventional crisp constraints (n- formnegotiation of mc~.ting time or relaxationof con- ary relations) are extended to fuzzy constraints straintswith or withoutthe intervention of its user. founded upon filzzy relations. Constraints imposed on Last but not least,the AMEETZERreminds its user meeting time by the attendees or meeting rooms can of forthcondng meetings aC appropriate times. be expressed as filzzy relations defined on the Carte- With distrihuted AMEETZERs,meeting requests sian product space of Day (in a week) x Time (in that involve disjoint se.ts of attendees and resources day), denoted as T. In a simple case, a time preference can be scheduled concurrently, as contrast to a cen- is represented as tralized approadl which faces conmnmication and pro- crossing bottlenecks as well as fanlt tolerance mid com- ci = {l,(t)lt E (2) plex scheduh., maintemmceissues. To filrther increase whereI~(t) ¯ [0, 1] represents the preference for a meet- the concurrency of the system, eacJl AMEETZERis ing to be scheduled at time t as seen by the system, dc~composedinto four subagents, Receptionist (inter- an attendee., or a meeting room. While l~(t) = 0 de- ~:ts with user to manage his calendar and prefer- notes ml impossible time for meeting possibly due to cnces and can’ies out meeting proposal and negotiation other commitment, l~(t) = 1 indicates maximal pref- ctc), Scheduler (infers the optimal time (and room) erence. Priorities among attendees (or meeting re- for a meeting under hard and .soft constraints etc), sources) and meetings can easily be incorporated by Messenger (connmmic:~tes with other AMEETZERs), introducing weighting coefiieients for ci or sieving ci and Learner (learns scheduling preferences). Together through modifiers (or linguistic hedges). the calendm’s and preferences are used to g~merate in- Using the framework of (1), AMEETZERinfers the stmme.sof soft constraint set that specify the preferv.nce most preferred time t" by all N attendees as ~tlues- (E [0, l]) on the meeting times. These prefi~.r once values are conmnmicatedto and inferenced by the I,(t’) = max/trainci(t)} (3) host AMEETZERto generate utility-optimal time for t -i_~N the meeting. whic]| provides a we.U-defined utility measure (optimal Another challenging asl)eCt in real life scheduling degree of joint constraint satisfaction). l~rohlem is the representation mid reasoning of hard Ameetzer is implemented in Java to take advantage and soft constraints. In real life, goal.s can also of its object-orientation anti platform independence.

Lira 449