Agent Technology for Distributed Organizational Memories

Comprehensive approaches to knowledge management in modern enterprises are confronted with scenarios which are heterogeneous, distributed, and dynamic by nature. Pro-active satisfaction of information needs across intra-organizational boundaries requires dynamic negotiation of shared understanding and adaptive handling of changing and ad-hoc task contexts. In this paper, we present the notion of a Distributed Organizational Memory (DOM) as a meta-information system with multiple ontology-based structures and a workflow-based context representation. We argue that agent technology offers the software basis which is necessary to realize DOM systems. We sketch a comprehensive Framework for Distributed Organizational Memories which enables the implementation of scalable DOM solutions and supports the principles of agent-mediated knowledge management. 1 DISTRIBUTED ORGANIZATIONAL MEMORIES Knowledge Management envisions the comprehensive use of an enterprise's knowledge, whoever acquired it, wherever it is stored and however it is formulated in particular. Organizational Memory Information Systems shortly Organizational Memories, OMs shall support the effective handling, conservation, and use of knowledge across time and space and as far as possible in person-independent ways. An OM comprises a variety of information sources where information elements of all kinds, structures, contents, and media types are available. The OM has to control and access these information sources in accordance with the users’ information needs, which are determined by a combination of personal, organizational and contextual circumstances: The useful interaction with the OM is influenced by the actual task at hand, but also by the individual’s role in the organization, his personal skills and interest profiles (and their overlap with the requirements of the current activity), as well as by prior knowledge and experience. The internal structure of an OM reflects this principle: By representing explicit interconnections between information elements and formalized models (particularly the domain, the enterprise, and the work context) the content of the information elements is partially made available to automatic processing and reasoning. As the various models form a basis for common reference across an enterprise, ranging from lists of shared vocabulary to more detailed ontological representations, common and shared understanding is supported by this approach. An explicit modelling of business processes as a means for context representation facilitates the situation-specific mark-up and retrieval of information elements; the integration with workflow systems which enact the process models enables pro-active information services. Consequently, an OM is best described as a metainformation system with tight integration into enterprise business processes, which relies on appropriate formal models and ontologies as a basis for common understanding and automatic processing capabilities (Figure 1; Abecker et al., 1998 & 2000). This description seems to motivate a central approach, and in fact a number of OM systems were realized as central repositories with globally valid ICEIS 2003 Artificial Intelligence and Decision Support Systems 4 ontologies and structures. However, centralized OM approaches have drawbacks with respect to two important aspects: a) Knowledge generation and use in an enterprise is distributed by nature. Departments, groups and individual experts develop individual, differing views on given subjects. These views are motivated and justified by the particularities of the actual work, goals, and situation. Obtaining a single, globally agreed-upon vocabulary on a level of detail which is sufficient for all participants is very expensive or even outright impossible. Consequently, an OM should benefit from balancing both local expertise – which might represent knowledge which is not easily shareable on a global level–and overall views on a more global level. A strictly centralized approach neglects this opportunity. b) Knowledge resides in changing environments. A centralized OM is ill-suited to deal with continuous modifications in the enterprise: The maintenance costs for its detailed models and ontologies simply get too high. Furthermore, centralized OMs assume a strict sequence of design, implementation, and use, while in reality a more evolutionary approach seems more promising: OM-like structures evolve in different groups and departments, using appropriate formalizations and conceptualizations. Integrating these elements under a common roof without disturbing their individual value should result in solutions which offer common benefit with reduced efforts while reaching better acceptance on the individual level. The reality of enterprises' environments thus asks for a distributed approach to OM realization: Distributed, heterogeneous OM cells let local expertise prevail while striving for maximal integrated benefit. Evolutionary growth and scalability on all levels is reached by allowing individual OM cells to grow and mature independently while interaction and communication brings enterprise-wide exchange and understanding. The natural approach for building complex software representations of distributed scenarios is agent technology. In the following, we will outline the characteristics of software agents which are helpful for Distributed Organizational Memories (DOM). We will argue that a comprehensive framework for DOMs requests the notion of agent societies. Further, an overview of typical instantiations of agents within the framework is given. 2 AGENT INFRASTRUCTURE FOR THE DOM 2.1 Agent-based Software Systems The DOM scenario is obviously characterized by a high degree of heterogeneity and distributedness, it can easily lead to a highly complex software system, and it is an open environment in the sense that we have to expect that frequently new components (even formerly unknown ones) may be plugged into the overall system, be replaced by other modules, or plugged out. Over the last years, the paradigm of agent-based computing turned out to be an appropriate means for dealing with such application scenarios. In this paper, we suppose the reader to have basic knowledge about agent-based software systems and engineering. We employ the “weak definition” of agents introduced in (Wooldridge & Jennings, 1995) with the following definitional features: (i) autonomy; (ii) social ability; (iii) reactive behaviour; and (iv) pro-active behaviour. Other possible characteristics of software agents like some level of intelligence, mobility, or techniques for learning and adaptation may also be relevant for parts of the overall solution we aim at. However, in this paper we will focus on multiagent systems’ capabilities for selforganization and social organization as a means for dealing with complex and dynamically changing situations which are mainly constituted by the characteristics mentioned above. In principle, human as well as software agents can be described with respect to the following dimensions corresponding to Newell’s (1982) knowledge level: Goals: Agents operate in a regularly changing environment. In doing so, they not only react to such changes, but also have their own goals and objectives which they try to achieve. Figure 1: OM as a Meta-Information System AGENT TECHNOLOGY FOR DISTRIBUTED ORGANIZATIONAL MEMORIES 5 Knowledge: Agents have knowledge with respect to the relevant realms of their environment, e.g. objects and other agents, as well as with respect to their own goals. Competencies: An agent’s abilities to perceive and manipulate its environment and its own internal state. In a multi-agent environment, the abilities to communicate with other actors are particularly important. Through communication, knowledge about facts, goals, competencies, etc. can be exchanged. This allows for negotiation and agreements which may lead to a distribution of tasks between agents or to changes of an agent's knowledge and goals. 2.2 Socially-Enabled Software Agents In Section 3 we will show that for a fully agent-based realization of the DOM scenario a huge amount of agents with possibly diverging goals and maybe highly complex communication and negotiation threads is required. As discussed in much detail by (Schillo et al., 2002), optimal work distribution and collaborative performance in such a group of agents benefits not only from task delegation and knowledge exchange, but also from social delegation as the basis for dynamic self-organization of agent societies, in order to achieve optimal group performance, yet staying flexible enough to cope with changing requirements. Via social delegation, groups of agents constitute Agent Societies with less communication effort because of clear responsibilities, with better task distribution because of specialization, etc. The phenomenon of society creation and self-organization can be observed in sociology (Bourdieu & Wacquant, 1992) and is a major topic of organizational theory. (Castelfranchi, 2000) considers it a crucial point for the introduction of agents into Enterprise Information Systems to complement the mechanisms for bottom-up control (system behaviour emerges from goals and negotiation at the micro level), which are inherent to the agent paradigm, by new mechanisms which appropriately reflect the global directives to be propagated top-down in a stable organization. In order to achieve this goal, we propose to build a DOM as a set of collaborating societies of sociallyenabled agents. These notions are being further refined in (Vicinus, 2002) and are exemplarily illustrated in (Elst & Abecker, 2002). In this paper we sketch the conceptual foundations and sketch their application in the DOM. Hence we define an Agent Society as a set of agents (an agent can be member of several societies at the same time) with at least one manager agent (which administers membership, role assignments, etc.) which enact for a certain time one or more Agent Roles with respect to this society. The role concept is not new in agent-oriented analysis and design methods like GAIA (Wooldridge et al., 2000), because analysis and modelling of an application domain is the easier the more similar the modelling paradigm is to the phenomena occuring in the real world. And, obviously, business situations and complex organizations are typically characterized by roles. Further we define Socially-Enabled Agents as software agents equipped with the required mechanisms to process appropriately rights and obligations, which together constitute a role in a society: Rights: Rights describe a subset of an agent's competencies. They describe under which conditions an agent is allowed to do something, like send a message to another agent, change his own goals, or grant rights to other agents. Obligations: Obligations also describe also a subset of an agent's competencies. They describe under which conditions (i.e. if a certain event occurs or another agent – maybe in a specified role – send a specific message) an agent is obliged to perform some action. Figure 2 above gives a rough idea of the software agent implementation we did for socially-enabled agents on top of the JADE (Bellifemine et al., 2001) platform. The major design decisions illustrated here are the fact that an incoming message must be first be sorted into the appropriate society module, because an agent may belong simultaneously to several societies. The respective society behaviour implements a Reactive Rule system which encompasses the obligation processing. This leads to a list of candidate actions which is then filtered by the right processing unit before being executed by the agent. Figure 2: Socially-Enabled Agent ICEIS 2003 Artificial Intelligence and Decision Support Systems 6 2.3 Competencies as Speech Acts In order to make the idea of rights and obligations a bit more concrete and to show how their semantics could be defined, we sketch how their introduction leads to speech acts in the agent society. We describe these speech acts similarly to FIPA: The sender, receiver and content of a speech act are specified; feasibility preconditions contain the qualifications; the rational effect shows the reasons for which a speech act might be selected. Table 1 shows two examples of FRODO speech acts for forming agent societies. With ApplyForRole an agent expresses the intention to take a specific role in a society. In the table two alternative specifications are given: a) In the simple specification the sender just wants the receiver to know that it wants to take the role and therefore the semantics of inform is used. Here, the receiver itself must infer that an appropriate reaction might be a GrantRole or a Deny. b) The second alternative is much more specific. Here, a request for a GrantRole action is used. This action should be applicable as soon as the receiver believes the desired role is possible for the sender. The precondition for ApplyForRole is that the sender really wants that role in the respective society and that it not already believes to have the role. Accordingly, the precondition for a GrantRole is that the sender i) has the right to do so (hasRole(sender, society, Manager)), ii) has a belief that the receiver wants the role, and iii) the specific role is appropriate for the receiver. So the manager of an agent society is responsible for forming the society by granting roles to other agents. The operationalization of a role’s rights and obligations for a concrete agent is done by a social layer in FRODOs agent platform sketched in Figure 2. 3 AGENT SOCIETIES FOR THE DOM In this section we briefly sketch the agent (sub-) societies required for building a DOM which arise directly from going through the several layers of the architecture in Figure 1. 3.1 Ontology Management As indicated in Fig. 1 and explained in more detail, e.g., by (Abecker et al., 2000; Davies et al., 2002), the future’s corporate-internal and external information systems will rely to much more extent than today on ontologies as shared, formalized accounts of domain knowledge structures. Table 1 Two Examples of FRODO Speech Acts for Agents Societies. FRODO speech act ApplyForRole Description An agents wants to take a specific role in a society and therefore sends an application to the manager. Sender S Receiver R Content role, society Feasibility Precondition NOT(Believes(S, hasRole(S, society, role))) AND Wants(S, role, society) Rational Effect Believes(R, Wants(S, role, society))) FIPA_action (inform :sender S