Integrating Information and Knowledge with Software Agents

UDC 519.681:681.3.068

VHironobu Kitajima VRyusuke Masuoka VFumihiro Maruyama (Manuscript received June 9, 2000)

This paper describes research and development activities for integrating distributed information and knowledge with software agent technology. First, we give an overview of our software agent architecture. Then, we describe the application of our software agent technology in a system that integrates distributed databases and in a document-oriented information integration system which integrates search engines. Our agent-based approach is suitable for dynamic information sources.

1. Introduction for users to retrieve information and knowledge The advent of the Internet has brought about simultaneously from distributed sources and to what is called an “information flood.” There is an make consecutive retrievals of associated infor- enormous amount of information and knowledge mation and knowledge. digitally available through networks. Although Distributed and disparate information sourc- such information and knowledge can be accessed es need to be integrated. This is where software on an individual basis, it is not an easy task for agent technology comes into play. In this paper, a users to acquire appropriate information and “software agent” is a computer system that is sit- knowledge. uated in an environment and is capable of The sources of information and knowledge autonomous action in this environment to meet have platform-level, system-level, and represen- its design objectives.1) A system of software agents tation-level differences. Platform-level differences hides platform-level, system-level, and represen- include the differences between hardware and tation-level differences and realizes a virtual operating systems, for example, differences be- integration. The areas where virtual integration tween Solaris, Windows NT, and mainframe can be applied include Supply Chain Management computers. System-level differences include the (SCM), Enterprise Application Integration (EAI), differences between database management sys- Enterprise Information Portals (EIPs), and Knowl- tems such as Oracle, SQL Server, DB2 and edge Management (KM). proprietary applications on mainframe comput- This paper describes our activities in this ers. Representation-level differences come from area under the Smart AGent Environment (SAGE) the different database structures and different project at Fujitsu Laboratories. Chapter 2 gives vocabularies used in databases, which include dif- a general overview of our software agent archi- ferent field names for corresponding fields and tecture and introduces its application to the different field values with the same meaning. integration of distributed databases for SCM. These three types of differences make it difficult Chapter 3 describes a system for document-

162 FUJITSU Sci. Tech. J.,36,2,pp.162-174(December 2000) H. Kitajima et al.: Integrating Information and Knowledge with Software Agents

oriented information integration which has been realize virtual catalogs. SAGE: Anthony, details in practical use by Fujitsu’s systems engineers of which are given in Chapter 3, is an application since October 1998. Chapter 4 describes the stan- of the agent system to a virtual integration of doc- dardization efforts and future directions of our ument search engines. In both applications, the project. Chapter 5 concludes the paper. agent system enables the system integrators to give end users a virtually integrated view of dis- 2. Integrating distributed databases tributed and disparate information sources In this chapter, we first introduce the Smart without changing the way in which information AGent Environment (SAGE) project and its intel- sources are operated. ligent agent systems, which enable users to This chapter mainly describes SAGE: Fran- virtually integrate distributed information sourc- cis. (SAGE: Francis lead to the development of an es. In the rest of this chapter, we mainly discuss agent system software product called AGENT- SAGE: Francis, which is an Electronic Commerce PRO.4),5) Fujitsu released AGENTPRO in August (EC) application of the SAGE project in which dis- 1999.) tributed and disparate Relational Databases (RDBs) are virtually integrated. Section 2.2 de- 2.2 Architecture scribes the architecture of SAGE: Francis, Section Figure 1 shows the architecture of SAGE: 2.3 describes its main features, and Section 2.4 Francis. It consists of user agents (UAs), facilita- describes our mediator agents called “facilitators.” tors (FAs), and database agents (DBAs).note 1) All Then, in Section 2.5, we explain how SAGE: Francis agents communicate with each other via Agent works by describing an example application in a Communication Language (ACL) messages, the Supply Chain Management (SCM) prototype sys- syntax of which is defined by the Knowledge Que- tem. Finally, in Section 2.6, we discuss ry and Manipulation Language (KQML)6) and centralized-type and decentralized-type solutions Knowledge Interchange Format (KIF).7), note 2) for the integration of distributed information sources. Browser Browser Browser Browser 2.1 SAGE HTML, etc. We are conducting research on intelligent User User User User agent agent agent agent agent systems under the SAGE project. Under Facilitator Facilitator this project, we have developed our agent system ACL and several agent system applications, which include SAGE: Francis2),3) and SAGE: Anthony. In Database Database Database Database agent agent agent agent both SAGE: Francis and SAGE: Anthony, the agent SQL, etc. system is used for virtual integration of distributed databases and other disparate information Database Database Database Database sources. SAGE: Francis is an application of our agent system to Electronic Commerce (EC) set- Figure 1 tings where RDBs are virtually integrated to Architecture of SAGE. note 1) There are also application agents (AAs) in AGENTPRO, which are created as wrappers for general applications. AAs are very close to DBAs. Essentially, they are DBAs in which database information sources are replaced with general applications. note 2) KQML and KIF are often called DARPA ACL since they are outcomes of Knowledge Sharing Efforts (KSE), which is a DARPA project. Future versions of AGENTPRO will be FIPA-compliant.8)

FUJITSU Sci. Tech. J.,36, 2,(December 2000) 163 H. Kitajima et al.: Integrating Information and Knowledge with Software Agents

Every ACL message in SAGE is sent and re- and which ontology they use. There are also “un- ceived asynchronously. Agents can automatically advertise” messages, which enable DBAs to deny match ACL messages such as a query message partially or entirely the knowledge sent by previ- and its resultant message. This is important be- ous advertise messages.note 3) cause agents can send the messages when Based on such knowledge, FAs can route and appropriate and operate freely. translate ACL messages appropriately. Hence, the The two main functions of UAs are to pro- system configuration is changed according to the cess conversions between ACL messages and advertise messages from DBAs. These advertise expressions on user interfaces such as Web pages messages enable a decentralized method of sys- and to offer a support service for users. One of tem configuration, which allows DBAs to have the main functions of DBAs is to process conver- autonomy over their operations. This advertise- sions between ACL messages and SQL queries: ment mechanism insulates UAs, and therefore the SQL is the query language of database manage- users, from the locations and other details of in- ment systems (DBMSs) such as Oracle and formation sources. The transparency provided by Microsoft Access. Two other key functions include this mechanism is essential for virtual informa- converting DB data into virtual knowledge to be tion integration. used in ACL messages from DBAs and informing 3) Ontology and ontology translation FAs of the capabilities of DBAs (a function called SAGE is based on ontologies. An ontology is “advertising”). The main functions of FAs, which a set of definitions of terms and the relationships act as mediators between UAs and DBAs, are for- between those terms. An ontology can be roughly warding messages to suitable DBAs, merging and defined as a vocabulary used by communicating sorting (if specified) results from multiple DBAs, agents. replying accordingly if there are no suitable Since distributed and disparate information agents, and translating terms in ACL messages sources are developed independently, they usual- as necessary. ly use different sets of terms. Even though agents use the same syntax for ACL messages, for exam- 2.3 Main features ple, KQML and KIF, ontologies used in ACL As a complete agent system, SAGE: Francis messages can be different. Therefore, FAs pro- has the following main features: vide an ontology translation service. With this 1) Agentification of legacy information sources service, the agent system hides vocabulary-level SAGE: Francis provides DBAs to agentify leg- difference from users. This is another essential acy information sources such as databases. After element for virtual information integration. the agentification, these information sources are FAs also provide flexible routing of ACL mes- made into agents which communicate via ACL sages based on the messages’ ontology, which uses messages using CORBA Object Request Broker its own category information. This is one of the (ORB). This hides platform-level and system- merits in having explicit ontologies. level differences between information sources. 4) Ontology Alignment Tool (OAT) 2) Dynamic system configuration by advertise This support tool provides users with a visu- messages al interface for manipulating ontology and DBAs can communicate knowledge about ontology translation information (Figure 2). It themselves to FAs by sending advertise messag- was created because only humans can provide es. This knowledge includes their addresses as note 3) DBAs and AAs can edit an FA’s knowledge agents, their information processing capability, in- about themselves with advertise and unad- formation on what kind of information they have, vertise messages.

164 FUJITSU Sci. Tech. J.,36, 2,(December 2000) H. Kitajima et al.: Integrating Information and Knowledge with Software Agents

Menu bar Ontology editor Correlation editor

Protocol Inference processor engine

Inference

Agent Knowledge base communication Advertise module Ontology Translation Parse and create KQML message

ACC ACC: Agent Communication Channel

Message bar INTERSTAGE (CORBA) Other agents

Figure 2 Figure 3 Ontology Alignment Tool (OAT). Architecture of a facilitator (FA). correct meanings, relations, and correspondences ing message. A protocol object is a Finite State of terms. We are currently considering adding Machine (FSM) that keeps the state of the ses- higher-level support functions to alleviate human sion. For each protocol object, the protocol tasks related to ontology manipulation. processor checks, at fixed intervals, whether a tim- 5) Merging and sorting of ACL messages eout occurs and whether the conditions required FAs merge the reply messages from multi- by the protocol are satisfied. If a timeout occurs ple DBAs for an original query message and sort or the conditions are satisfied, the actions speci- them if so specified. Because of this function of fied in the protocol object are executed. This FAs, a user sees the result of a query as a single mechanism enables FAs to be involved in multi- list, even though the results may have come from ple, complex sessions at the same time. several DBAs. Other important modules are the inference engine and the knowledge base. The knowledge 2.4 Facilitators (FAs) base keeps ontologies and ontology translations In this section, we give some details about loaded at the startup of the FA. It also keeps ad- facilitators (FAs), which constitute the core of our vertise information, which is dynamically inserted agent system. Other agents, for example, UAs and by advertise messages from DBAs and is dynam- DBAs, differ from FAs in that they have connec- ically deleted by unadvertise messages. The tions to outer resources such as users and inference engine is used to match the query mes- databases. Another major difference between FAs sages from UAs and advertise information. and other agents is in their implementation Ontologies and ontology translations are utilized languages. Common Lisp is used for the imple- during this matching process. mentation of FAs, while Java is used for the FAs communicate with other agents through implementation of UAs and DBAs. the agent communication module. The module Figure 3 shows the architecture of an FA. stores all received messages and allows other One of the most important modules is the proto- modules to read them by function calls. It pro- col processor. FAs need to manage multiple and vides the functions to parse and create ACL complex sessions of message exchanges. To real- messages. The module, in turn, uses the Agent ize this management, a protocol object is created Communication Channel (ACC) to send and for each session at the receipt of a session-initiat- receive the messages. The ACC supports asyn-

FUJITSU Sci. Tech. J.,36, 2,(December 2000) 165 H. Kitajima et al.: Integrating Information and Knowledge with Software Agents

parts suppliers (Figure 4), goes as follows. The TV manufacturer has a factory, a shipment center, a sales center, and parts suppliers distributed throughout a country. Each department or supplier has its own databases. Since they are developed more or less independently, they use different Database Management Systems (DBMSs), for example, Oracle and Microsoft Ac- cess, and they have different applications on their mainframe computers. They also use different ontologies for their field names and data in their databases. This prototype was built for an imaginary TV manufacturer and The manufacturer introduced the agent sys- associated parts suppliers. By agentifying databases and applications at a factory, shipment center, sales center, and parts tem to create an SCM system that will enable the suppliers distributed in a country, mediation by the facilitator enables the prototype to answer a query by a user at the headquarters sales persons to provide estimated delivery dates. of the TV manufacturer regarding the delivery date for a specified First, they agentify the distributed databases into amount of product. the DBAs. Then, they use a visual tool, OAT, to Figure 4 SCM prototype by agent system. create ontology files and ontology translation files for translation between different ontologies. These chronous messaging between the agents. This files are provided for the FA, a mediator agent, agent communication mechanism is built on which provides ontology translation and other INTERSTAGE,9) which is Fujitsu’s CORBA mid- ontology-related services. dleware. When the DBAs and AAs are started, they send out advertise messages to the FA. The ad- 2.5 How it works: an SCM system vertise messages describe the DBAs’ and AAs’ prototype capabilities and what kind of information those In this section, by using an SCM system pro- agents have. They also create UAs, which inter- totype, we describe how SAGE: Francis is applied pret the users’ intentions into agent messages and to virtually integrate various information sourc- which communicate received agent messages to es and how it works. This SCM system prototype users. The users’ side of the agents is realized by was developed for a demonstration of AGENTPRO. Web interfaces. Through the Web interface of the It uses application agents (AAs) along with DBAs UA, a sales person at the headquarters makes a which agentify general applications. The agent query about the estimated delivery date of a spe- system used in this prototype is somewhat more cific product. Then, the UA sends out the query advanced than the currently released version of in an ACL message to the FA. AGENTPRO. The advanced features will be Based on the information provided by the ad- provided as a part of the future versions of vertise messages from the DBAs and AAs, the FA AGENTPRO. queries the DBAs and AAs about the stocks at the One of the purposes of this SCM prototype production line, the shipment center, and sales system is to provide an estimate of the delivery center. The FA sends out to each DBA or AA a time of products by virtually integrating distrib- query message in the ontology used by the agent. uted information sources with the agent system. Then, the FA waits for the asynchronous replies The scenario of the prototype, which was built for from the DBAs and AAs. an imaginary TV manufacturer and associated If there are not enough stocks, the FA que-

166 FUJITSU Sci. Tech. J.,36, 2,(December 2000) H. Kitajima et al.: Integrating Information and Knowledge with Software Agents

ries the DBA at the factory about the products on With centralized-type solutions, the perfor- the production line. If there are not enough prod- mance is independent of the performance of the ucts on the production line to make up for the software and hardware used for distributed infor- shortfall, the FA further queries the DBA at the mation sources. It is also independent of the factory for the parts needed for the necessary ad- communication environment between the central ditional production. Then, the FA asks the DBAs server and distributed information sources. Those and AAs of the parts suppliers for the delivery factors can work against decentralized-type dates of those parts. With the obtained informa- solutions. On the other hand, decentralized-type tion, the FA comes back to the DBA at the factory solutions work better when changes and variet- and queries about the estimated date for comple- ies are abundant. Changes include addition and tion of the additional production. Then, the FA deletion of distributed information sources. Vari- summarizes all the information and sends a mes- eties include those of platforms, systems, and sage containing the result to the UA, which in turn vocabularies. displays the result to the user. Sometimes an open environment prohibits A new parts supplier can join the system by the use of centralized-type solutions, because the agentifying its database and by sending an ad- participating distributed information sources are vertise message to the FA when it is ready. There operated independently and autonomously. These are also unadvertise messages that remove the sources will not or cannot submit all their infor- information of the senders from the knowledge mation to the central server, but only accept base of the receiver. This advertisement mecha- one-time queries. nism enables dynamic configuration of the total Now, there is a new situation which makes system, allowing departments and suppliers to centralized-type solutions difficult. That is, more join and leave at their convenience and to have and more information is being created or composed autonomy over their operations. dynamically, especially for Web pages. This situation prevents centralized-type solutions from 2.6 Centralized-type versus collecting all the data from distributed data sourc- decentralized-type solutions es efficiently since often an indefinite number of In this section, we compare centralized-type pages can be created by the combination of the and decentralized-type solutions for integrating original data held there. In order to deal with distributed information sources. such cases, all of the distributed information sourc- Centralized-type solutions retrieve data from es have to be queried each time a user makes a distributed information sources and send it to a query and, therefore, decentralized-type solutions central server using robots or FTP. The central are the only options. server usually consists of one or more very high performance computers. Then, indices are given 3. SAGE: Anthony to the collected data. The central server provides 3.1 Overview services such as a query service for the collected We have already developed an intelligent data. Our agent system is an example of a decen- agent environment named SAGE. It uses a kind tralized-type solution. of mediation agent, called a facilitator (FA), as an Both types of solutions have their own mer- essential part of the technology. SAGE has proved its and demerits. Our argument is that recent its usefulness in the SAGE: Francis project, the rapid changes in the corporate environment and target of which is the virtual integration of multi- the need to interoperate with other enterprises ple RDBs over a network. As a next step, we favor decentralized-type solutions. started another project to build a new informa-

FUJITSU Sci. Tech. J.,36, 2,(December 2000) 167 H. Kitajima et al.: Integrating Information and Knowledge with Software Agents

tion system, called SAGE: Anthony,10) for systems documents. engineers at Fujitsu. 4) Dynamic integration of information sources The systems engineers of our company had The DBAs are dynamically integrated by been using a lot of isolated information systems sending advertise messages to the FA. Such mes- which are mainly based on WWW servers and are sages can keep the FA informed of knowledge administrated independently by different depart- about DBAs and ready for providing facilitation ments. Since most of the systems have no relation services. to each other and their authentication processes 5) Unified authentication were not unified, it was a very laborious task The user authentication processes of multi- to search the network for information. SAGE: ple information sources are unified into the Anthony is an integrated document search engine one-time login step of SAGE: Anthony’s entry system which can address this issue, even over a point. This service is brought by an authentica- very large scale network. tion agent, which is a special kind of DBA that In this chapter, we present a brief descrip- has information about all users and groups. tion of the system from the architectural and 6) Scalability functional points of view. Compared with conven- SAGE: Anthony’s natural dispersiveness pro- tional document search engine systems, SAGE: vides a certain degree of system scalability. Since Anthony has the following features. we allotted a portion of the document-searching 1) Dispersiveness task to each DBA and kept its index-file size small- Documents to be searched and index-files to er than that of the conventional centralized search be used by search engines in their keyword-match- engine system, the time efficiency is higher and ing processes are not centralized. Document updating of indexes and files can be done more indexes are stored in multiple local sites equipped frequently. with search engines. Each site is managed inde- Efforts to integrate document-searching ser- pendently and is only responsible for searching vices for enterprise information systems are its locally owned documents. already underway. One of these systems is Dom- 2) Standardization of information sources ino Extended Search (DES).12) DES provides We defined a document database equipped distributed heterogeneous searches across the with a search engine as a basic unit for informa- network using the user’s explicit selection of tion sources and transformed it into a DBA of information sources. Compared with this, SAGE: SAGE: Anthony. If there were differences in the Anthony can automatically determine the recom- designs of the search engine programs, we stan- mendable DBAs by analyzing the previously dardized them through their agentifying acquired knowledge about DBAs. processes. The DBA usually holds index-data SAGE: Anthony is a kind of three-tier about documents only; the documents themselves system whose middle tier is a facilitator agent. may be accumulated in other sites such as WWW The middle tier stores information to be used for servers. the virtual integration of information sources. 3) Efficient searching with facilitation There has been some research into heterogeneous SAGE: Anthony has a federated architec- searches with three-tier systems;13)-15) however, the ture11) so that the facilitation services of its FA middle tiers of these systems are rather static. can perform efficient searches. These services are SAGE: Anthony’s FA is so dynamic that the sys- based on the FA’s internally stored knowledge tem can realize on-the-fly modification of about other agents and play a key role in SAGE: searching policies by real-time updating of stored Anthony’s economical distributed searching of knowledge in the FA.

168 FUJITSU Sci. Tech. J.,36, 2,(December 2000) H. Kitajima et al.: Integrating Information and Knowledge with Software Agents

HTTP and an automatic-query agent. The former is a GUI WWW User agent User agent server special kind of DBA that maintains information

ACL about all users and groups for the unified authen- ACL SMTP ACL tication process. The latter is a substitutional Authentication Inference Facilitator engine Automatic-query agent that executes scheduled queries on behalf agent agent ACL of human users. The automatic-query agent uses ACL the SMTP protocol for returning answers of auto- mated queries to the users. Search Search Search DB engine DB engine DB engine agent agent agent 3.3 Facilitation services A typical document search proceeds through Figure 5 message-passing routines which are very similar SAGE: Anthony. to those of SAGE: Francis. For example, when a UA of SAGE: Anthony issues a query by sending 3.2 Architecture a KQML’s ask-all message to the FA, it redirects SAGE: Anthony inherited its federated the message to the recommendable DBAs based architecture from the design concept of on its knowledge about DBAs. After receiving the SAGE: Francis. Figure 5 shows a schematic of ask-all message from the FA, each DBA executes the agent system. As can be seen, the architec- an exhaustive document search and returns the ture is very similar to that of SAGE: Francis. answer to the FA by sending reply messages. The Actually, there is little difference between these message includes information about searched doc- two systems from the architectural point of view. uments such as URLs, titles, and file-sizes. The The principal difference is in the content of the FA collects such answers and returns a merged facilitation services. reply message to the UA that originated the que- A WWW browser is used for the GUI. The ry task. browser can access SAGE: Anthony via a WWW In other aspects, however, there are differ- server using a UA and can also display the re- ences between SAGE: Francis and SAGE: Anthony. sults of document searching for human users. The Whereas SAGE: Francis’ target of virtual integra- UA, which is an entrance of SAGE’s agent world, tion is an RDB, SAGE: Anthony integrates search can translate the user’s operational actions into engines. The contents of the facilitation service, appropriate ACL messages. therefore, needed to be designed uniquely. Each document database has a keyword- The FA of SAGE: Anthony provides intelligent matching-based search engine for its own use. The facilitation services. It determines recommend- database is agentified to a DBA that can only be able DBAs based on various kinds of internal accessed from an FA with ACL messages. knowledge and redirects received query messag- The FA provides facilitation services based es to them. The knowledge can be classified into on the content of the message from the UA, utiliz- the following three domains. ing previously acquired knowledge concerning 1) Database category other agents. Since there is only one FA in the DBAs are classified by a pre-defined taxono- system, it may become a bottleneck that reduces my concerning their contents and owners. Some the searching performance. We therefore enabled example classifications are “Technical Informa- mirroring of the FA in order to further enhance tion,” “Product Information,” “Industrial System the scalability of this system. Department,” and “Medical System Division.” This system features an authentication agent Such classifications, which are usually organized

FUJITSU Sci. Tech. J.,36, 2,(December 2000) 169 H. Kitajima et al.: Integrating Information and Knowledge with Software Agents

(ask-all (reply :content :content ( (and ("http://www.fujitsu.co.jp" (category ?database "finance") "Fujitsu's Homepage-1" 1345 (has ?database ?doc) "19971205T000000000" (and (keyword ?doc "keyword1") "Fujitsu's pen computer is adopted by (or (keyword ?doc "keyword2") NASA's Space Shuttle project..." (keyword ?doc "keyword3")) (not (keyword ?doc "keyword4"))) 12.5) (url ?doc ?url) ("http://abc.www.fujitsu.co.jp" (title ?doc ?title) "Fujitsu's Homepage-2" 2012 (byte-size ?doc ?byte-size) "19960303T000000000" (registered-date ?doc ?registered-date) "Here's very useful information about..." (summary ?doc ?summary) 31.3) (ranking ?doc ?ranking)) ... :aspect (?url ?title ?byte-size ... ?registered-date ?summary ?ranking) ) :id "fj900557" :language KIF :language KIF :language-encoding x-euc-jp :language-encoding x-euc-jp :ontology database.fujitsu.kif :ontology database.fujitsu.kif :in-reply-to abcdefg-6-0 :reply-with abcdefg-3-0 :reply-with abcdefg-6-1 :sender [email protected] :sender [email protected] :receiver [email protected]) :receiver [email protected])

Figure 6 Figure 7 Ask-all message. Reply message. in a family-tree style, can be regarded as a spe- be acquired in advance of the searching task by cial ontology to classify information sources. The receiving KQML advertise messages from other FA knows as basic premises which category of doc- agents. The knowledge can always be updated ument is indexed in which DBA. Users can specify dynamically by additional advertisement or by such categories when they query the system. direct updating of databases. The FA has an in- 2) Keyword information ference engine to store and handle such The FA has a keyword-database to store knowledge, which can be applied for intelligent records of keyword-related information consisting facilitation services. of keywords, names of DBAs, and frequencies of documents. The frequency indicates the number 3.4 Message examples of documents indexed in the DBA that match the In this section, we show some examples of keyword. If we expect the information to be com- the ACL messages used in the conversation pro- plete, the number of such records may be huge. cess among agents in order to present a clear Therefore, the keywords to be stored are automat- image of ACL message passing. ically selected according to the frequency at which Figure 6 shows an ask-all message sent to they appear in users’ query conditions. the FA from a UA. The UA is making a query 3) User profile about documents that are held by the DBAs of The user profile consists of attributive infor- the finance category and are matched to four spec- mation about users such as the user’s user name, ified keywords by boolean operations. The part of password, work, and post. The authentication the message tagged with “:aspect” prescribes that agent maintains a user-profile database and com- the answer message must be composed of docu- municates its knowledge to the FA on demand. ment properties such as a URL, title, byte size, The FA possesses DBA profiles which contain sim- and so on. After receiving the ask-all message, ilar knowledge to the user profiles. The FA utilizes the FA redirects it to the recommendable DBAs. these two kinds of profiles to check the user’s ac- In the example shown in Figure 7, a DBA is re- cessibility to each DBA. plying to the FA with the searched documents’ Most of the knowledge mentioned above must properties that are specified in the ask-all mes-

170 FUJITSU Sci. Tech. J.,36, 2,(December 2000) H. Kitajima et al.: Integrating Information and Knowledge with Software Agents

(advertise sage shown in Figure 6. :content ((database documentDB-1) Figure 8 shows an advertise message sent (has-database [email protected] documentDB-1) to the FA from a DBA. The DBA is advertising its (know-category documentDB-1 "finance") name, database category, ontology name, DBA (know-category documentDB-1 "medicine") (ontology [email protected] profiles, and other properties. These advertise- standard.database.fujitsu.kif) (access_level documentDB-1 "0") ments will be applied to the intelligent facilitation (department_flag documentDB-1 "c10000") services. (post_flag documentDB-1 "ffffff") (place_flag documentDB-1 "ffffff") (allows-relational-db-query [email protected]) 3.5 Current status (table-definition documentDB-1 "document database" After the successful prototyping stage, SAGE: '(description "database of documents")) Anthony evolved into a practical information sys- (field-definition documentDB-1 "document database" tem with industry-level robustness and with "keyword" '(type text description functionality for easy administration. The sys- "keyword to be searched")) tem, named FIND Future, has been running for ... ) 10s of thousands of Fujitsu’s systems engineers :sender [email protected] :receiver [email protected] since October 1998. Figure 9 shows a screen- :reply-with abcdefg-4-1 :language-encoding x-euc-jp captured user’s view of the system. :language KIF Currently, FIND Future integrates about 50 :ontology database.fujitsu.kif ) DBAs installed throughout Fujitsu’s intranet. It

Figure 8 boasts a huge number of accesses from every part Advertise message. of our company in Japan. In the very near future, FIND Future will be expanded to a worldwide enterprise information system by adding federated facilitation services among multiple independent FAs.

Figure 9 FIND Future.

FUJITSU Sci. Tech. J.,36, 2,(December 2000) 171 H. Kitajima et al.: Integrating Information and Knowledge with Software Agents

4. Standardization efforts and future Enterprise Application Integration (EAI) and direction Enterprise Information Portals (EIPs). For EAI In this chapter, we describe our standardiza- and EIPs, both within and between enterprises, tion efforts and the future direction of our agent customized integration of disparate and distrib- research. uted information sources is essential. Regarding our standardization efforts, we are For enterprises to stay competitive in these committed to the Foundation for Intelligent application areas, it will be more and more im- Physical Agents (FIPA), which is an agent stan- portant to adapt to changes. Enterprises need to dardization organization. We have been active in change their information systems to introduce new the area of ontology and content language stan- services for customers and to distinguish them- dardizations. Also, we have recently submitted a selves from their competitors. There is also an workplan called the “Agent Description Ontolo- ever changing environment, and new laws and gy.” The Technical Committee (TC) of the rules will continue to be introduced. Restructur- Agreement Management of the FIPA has decided ing of enterprises, mergers, acquisitions, etc., will to include this workplan in its mission statement necessitate the accommodation of new informa- and to produce the specification under the name tion sources. Enterprises should be ready to deal of the “Service Description Ontology.” One of the with new users and new requests from users. authors (Hironobu Kitajima) has been appointed With such an enormous pressure from chang- as the redactor of the specification. This work- es, enterprises will need to personalize and plan aims to specify ontologies to explicitly and optimize their services for each user. To deal with formally describe agent services, agent capabili- this situation, we firmly believe that automatic ties, and agent needs. Realizing these aims will service integration of disparate information enable more accurate and efficient interoperation sources and information systems is quintessen- between agents. Furthermore, their realizations tial. Inside and outside changes force us to have are the absolute prerequisites for content-based externalized rules instead of knowledge pro- routing. grammed into code and to constantly reconfigure Almost all application domains may benefit the services based on those rules to satisfy each from these specifications. In particular, those that user’s request. involve loosely coupled processes, for example, da- We consider that Artificial Intelligence (AI) tabase integration, electronic commerce, and technology will play a vital role in automatic ser- CALS tools, will benefit. vice integration. We also consider that ontology We are also conducting interoperability tests will be the key to gluing multiple services togeth- with other agent platforms. We have succeeded er automatically and that the Service Description in an interoperability test between a modified Ontology mentioned above provides the basic version of AGENTPRO and Comtec Agent Plat- mechanism for sharing knowledge needed to re- form based on FIPA97 specifications.16) We have alize automatic service integration. also succeeded in a private interoperability test between a modified version of AGENTPRO and 5. Conclusion JADE Agent Platform.17) We plan to provide a new We have described our research and devel- agent platform in the future versions of AGENT- opment activities for integrating distributed PRO which will conform to a series of new FIPA information and knowledge with software agent specifications released in 2000. technology. Applications of our integration tech- Regarding the future direction of our agent nology include the integration of distributed research, we expect that it will be connected with databases such as SCM systems and document-

172 FUJITSU Sci. Tech. J.,36, 2,(December 2000) H. Kitajima et al.: Integrating Information and Knowledge with Software Agents

oriented information integration systems such as cal Agents). information sharing systems. In August 1999, http://www.fipa.org/ Fujitsu released a software product called AGENT- 9) INTERSTAGE. PRO based on our integration technology. http://www.interstage.com/ Currently, we are working on extending our tech- 10) H. Kitajima, A. Kawamura, T. Yoshino, and nology to advanced EAI and EIPs. F. Maruyama: SAGE: Anthony. (in Japanese), Computer Software, 17, 1, pp.32-44 (2000). References 11) M. R. Cutkosky et al.: PACT: An Experiment 1) Gerhard Weiss, ed.: Multiagent Systems – A in Integrating Concurrent Engineering Modern Approach to Distributed Artificial In- Systems. IEEE Computer, January 1993, telligence. The MIT Press: Cambridge, MA, pp.28-38. 1999. 12) Lotus Development Corporation, Domino 2) R. Masuoka, T. Sugasaka, A. Sato, H. Kitajima, Extended Search. and F. Maruyama: SAGE and Its Application http://www.lotus.com/home.nsf/welcome/ to Inter-company EC. Proceedings of PAAM98, domsearch pp.123-135, 1998. 13) Y. Arens, Chee, Y. Chin, Hsu, Chun-Nan and 3) T. Sugasaka, K. Tanaka, R. Masuoka, A. Sato, Knoblock, and A. Craig : Retrieving and In- H. Kitajima, and F. Maruyama: A conversa- tegrating Data from Multiple Information tional agent system and its application to Sources. International Journal of Intelligent electronic commerce. Proceedings of the and Cooperative Information Systems, 2, 2, fourth & fifth world conference on integrat- pp.127-158 (1993). ed design and process technology (IDPT 14) H. Garcia-Molina, J. Hammer, K. Ireland, 2000), June, 2000. Y. Papakonstantinou, J. Ullman, and 4) INTERSTAGE AGENTPRO DataSheet. Jennifer Widom: Integrating and Accessing http://www.interstage.com/images/ Heterogeneous Information Sources in AgentProDataSheet.pdf TSIMMIS. Proceedings of the AAAI Sympo- 5) Dynamic Advertising – The Key to Intelligent sium on Information Gathering, Stanford, Data Retrieval. California, March 1995, pp.61-64. http://www.interstage.com/images/ 15) A. Y. Levy, A. Rajaraman, and J. J. Ordille: AGENTPRO%20Whitepaper1.pdf Querying Heterogeneous Information Sourc- 6) The DARPA Knowledge Sharing Initiative es Using Source Descriptions. Proceedings External Interfaces Working Group: Specifi- of the 22nd International Conference on Very cation of the KQML Agent Communication Large Databases. VLDB-96, Bombay, India, Language, 1994/2/9. September 1996. http://logic.stanford.edu/papers/kqml.ps 16) FIPA-related software developed by mem- 7) Draft proposed American National Standard bers and non-members. (dpANS): Knowledge Interchange Format http://www.fipa.org/fipa9706.pdf (KIF), NCITS. T2/98-004. 17) JADE (Java Agent DEvelopment Frame- http://logic.stanford.edu/kif/dpans.html work). 8) FIPA (Foundation for Intelligent and Physi- http://sharon.cselt.it/projects/jade/

FUJITSU Sci. Tech. J.,36, 2,(December 2000) 173 H. Kitajima et al.: Integrating Information and Knowledge with Software Agents

Hironobu Kitajima received the B.A.S. Fumihiro Maruyama received the B.S. degree from the University of Tokyo, degree in Mathematical Engineering Tokyo, Japan in 1986. He joined from the University of Tokyo, Japan in Fujitsu Laboratories Ltd., Kawasaki, 1978. He joined Fujitsu Laboratories Japan in 1993 and has been engaged Ltd., Kawasaki, Japan in 1978 and has in research and development of simu- been engaged in research and devel- lated annealing and agent systems. He opment of computer-aided design and was a visiting scholar of Stanford Uni- artificial intelligence. He received his versity from 1996 to 1997. He is a Dr. of Engineering degree in Informa- member of the Information Processing tion Engineering from the University of Society of Japan (IPSJ). Tokyo in 1991. He received the IPSJ 20th Anniversary Best Paper Award and Prof. Motooka Com- memorative Award in 1980 and 1988, respectively. He is a member of the Institute of Electrical and Electronics Engineers (IEEE), the Information Processing Society of Japan (IPSJ), the Japa- nese Society for Artificial Intelligence (JSAI), and the Institute of Ryusuke Masuoka received the B.S. Electronics, Information and Communication Engineers (IEICE) and M.S. degrees in Mathematics from of Japan. the University of Tokyo, Tokyo, Japan in 1985 and 1987, respectively. He joined Fujitsu Laboratories Ltd., Kawasaki, Japan in 1988 and has been engaged in research and development of neural networks, simulated annealing, and agent systems. He is a member of the Institute of Electrical and Electronics Engineers (IEEE), the Institute of Elec- tronics, Information and Communication Engineers (IEICE) of Japan, and the Information Processing Society of Japan (IPSJ). He received the Best Author Award from the IPSJ in 1995 and his Ph.D. degree in Mathematical Sciences from the University of Tokyo in 2000.

174 FUJITSU Sci. Tech. J.,36, 2,(December 2000)