OntoBayes Approach to Corporate Knowledge

OntoBayes Approach to Corporate Knowledge Yi Yang and Jacques Calmet Institute for Algorithms and Cognitive Systems (IAKS) University of Karlsruhe (TH) 76131 Karlsruhe, Germany {yiyang,calmet}@ira.uka.de

Abstract. In this paper, we investigate the integration of virtual knowledge communities (VKC) into an ontology-driven uncertainty model (OntoBayes). The selected overall framework for OntoBayes is the multiagent paradigm. Agents modeled with OntoBayes have two parts: knowledge and decision making parts. The former is the ontology knowledge while the latter is based upon Bayesian Networks (BN). OntoBayes is thus designed in agreement with the Agen Oriented Abstraction (AOA) paradigm. Agents modeled with OntoBayes possess a common community layer that enables to define, describe and implement corporate knowledge. This layer consists of virtual knowledge communities.

1

Introduction

Corporate knowledge was defined as the overall knowledge detained by agents within a system and their ability to cooperate with each other in order to meet their goal. Decision making based on corporate knowledge has become crucial for a society made of distributed agents each possessing its own knowledge, particularly when the knowledge is uncertain. Uncertainty is an inevitable feature in most environments. Agents do not act within a static and close environment but within a dynamic and open one. The available information is mostly incomplete and often imprecise because agents almost never have access to the whole truth implied by their environment. Agents must therefore act under uncertainty. During a process of decision making agents can profit from the corporate knowledge of their society better than only from their sole knowledge. Agents are required to know not only ”what it knows” but also ”what they know”, and are expected to make maximum use of the knowledge. The initial step to build uncertain knowledge-based decision support systems is to adopt a representation. It was solved in an ontology-driven uncertainty model (OntoBayes)[1]. This model proposed an approach to integrate Bayesian Networks (BN) into ontologies, in order to preserve the advantages of both. Bayesian Networks provide an intuitive and coherent probabilistic representation of uncertain knowledge, whereas ontologies can represent the organizational structure of large complex domains excellently. The next step consists of reasoning under probabilistic knowledge. The Bayesian probability theory is an acknowledged standard for probabilistic reasoning.

It provides not only exact but also approximate reasoning mechanisms. Exact methods exploit the independence structure contained in the network to efficiently propagate uncertainty [2, 3]. The approximate methods are based on stochastic simulation, which constitute an interesting alternative in highly connected networks, where exact algorithms may become inefficient [3–5]. OntoBayes should be easily combined with existing reasoning methods to meet its objectives. The last important step is to construct decision mechanisms. The decision making part of OntoBayes makes use of the classic decision-theoretic technique. Such techniques possess explicit management of uncertainty and tradeoffs, probability theory and the maximization of expected utility. This part is described in [6]. In this paper, we investigate the integration of virtual knowledge communities (VKC) into OntoBayes, in order to introduce uncertain corporate knowledge in a society of agents. The selected overall framework for OntoBayes is the multiagent paradigm. Agents modeled with OntoBayes have two parts: knowledge and decision making parts. The former is the ontology knowledge while the latter is based upon BN. OntoBayes is thus designed in agreement with the Agent Oriented Abstraction (AOA) paradigm[7]. Agents modeled with OntoBayes possess a common community layer that enables to define, describe and implement corporate knowledge. This layer represents virtual knowledge communities. The remaining sections of this paper are structured as follows. Section 2 discusses the previous works in this field in further detail. Section 3 provides an overview on the existing approach VKC. Section 4 summarizes the current work on OntoBayes. Section 5 is devoted to investigate the integration of VKC into OntoBayes and finally, Section 6 contains a brief overview on future work and concludes the paper.

2

Previous Works

The work of modeling corporate knowledge based on AOA was a major source of inspiration of this research. The grounding theory behind the design of OntoBayes is the agent oriented abstraction paradigm presented in Calmet’s work[7]. The AOA paradigm covers the concepts of agents, annotated knowledge, utility functions and society of agents. Indeed, AOA is based on Weber’s classical theory in Sociology [8]. AOA assumes that agents are entities consisting of both a knowledge component and a decision making mechanism. The former is partitioned into four components, also called annotations: ontology, communication, cognition and security. The latter is related to its tasks and goals. It generates utility functions and is based upon the knowledge component. Chiefly, agents can be defined in terms of knowledge and utility. The AOA model can be abstractly summarized by a number of basic definitions. A detailed instruction of that is to be found in [7]. In [9] the application of AOA model to the abstract modeling of corporate knowledge is investigated. Corporate knowledge was defined as the amount

of knowledge provided by individual agents. To avoid the separation between agents and knowledge, it was considered that agents have explicitly represented knowledge and communication ability. A knowledge company was modeled as a scenario to demonstrate the corporate knowledge modeling within the AOA. The concrete implementation for corporate knowledge within AOA was introduced in [10]. The realized concept was introduced as virtual knowledge communities [11, 12]. VKC is the applied method to enhance the OntoBayes model for corporate knowledge retrieval, so it is reviewed in the next section separately and in more details.

3

Virtual Knowledge Communities

Traditionally, information is mostly centralized within a uniform information structure. This viewpoint is not truly compliant with the nature of knowledge that is subjective, distributed and contextual [13]. From the perspective of the knowledge information society, modern knowledge management often focuses on the constitution of communities of practice and communities of interest [14]. The concept of a community of practice or a community of interest can be supported in a virtual community in order to bring the concerned agents together to share their knowledge with each other. A community is a place where agents can meet and share knowledge with other agents which share a similar domain of interest. The concept of VKC was introduced as a means for agents to share knowledge about a topic [10]. It aims to increase the efficiency with which information is made available throughout the society of agents. From the point of view of corporate knowledge management, agents can be individuals, software assistants or automata. Agents possess knowledge and processes within the society tend to make agents produce and exchange knowledge with each other. These processes are distributed throughout the society and contribute through their own intrinsic goals to solve a unique highlevel challenge. This provides the link between corporate knowledge and virtual knowledge community [9]. In the implemented model [11] there are two main modeling for VKC: agent modeling and community modeling. The former has four key notions: personal ontology, knowledge instances, knowledge cluster and mapper. Personal ontology represents the knowledge of an agent. It describes the taxonomy of the relationships between the concepts and predicates that an agent understands. The knowledge instances are instances of objects defined into the personal ontology. It was assumed that an agent’s knowledge consists of both its personal ontology and knowledge instances according to its personal ontology. The knowledge clusters as sub-part of an ontology can be shared among agents. They are defined by their head concept, a pointer to the different parts of knowledge existing in a cluster. The mapper chiefly contains a set of mapping from personal terms to mapped terms, and allows an agent to add such mappings, and use the mapper to normalize or personalize a given knowledge cluster or instance using

these mappings. It facilitates knowledge sharing among agents with regards to the heterogeneity of knowledge. The community modeling has also some key notions: domain of interest, community pack, community buffer. A domain of interest exists in each virtual knowledge community and is similar to the concept of ontology for an agent. It is given by the community leader which created the community. The community pack is what defines the community. It consists of a community knowledge cluster, a normalized ontology which contains at least the head of the community cluster, and the identification of the leaders of the community. The community buffer can record messages which are used by the member of a community to share their knowledge. This approach is compatible with blackboard systems, but still has its difference, because agents cooperate to solve their respective problems, not for a unique goal. The VKC approach has been designed and partially implemented as a prototype system. The implementation is based on Java Agent Development Framework (JADE) and Java Runtime Environment (JRE) platform. It was tested and evaluated [11]. A component of the system enables to simulate VKC.

4

Knowledge Representations in OntoBayes

The typical characteristic of domains such as insurance or natural disaster management is the risk factor. The normal ontological model is not sufficient to express the domain specific uncertainty or risks. In order to overcome this shortcoming the OntoBayes model was proposed [1]. This model consists of knowledge and decision making parts. Its decision making part will not be discussed in this paper, because it is not fully relevant to define corporate knowledge within this model. We remind in subsections 4.1 and 4.2 well-known concepts that are basic for ours approach. 4.1

Web Ontology Language

OntoBayes makes use of OWL as its formal language to facilitate its knowledge representation. OWL is a semantic markup language for publishing and sharing ontologies on the World Wide Web. It is intended to provide a language that can be used to – conceptualize domains by defining classes and properties of those classes, – construct individuals and assert properties about them, and – reason about these classes and individuals [15]. OWL is developed as a vocabulary extension of the Resource Description Framework (RDF) [16]. The underlying structure of any expression in RDF is a collection of triples, each consisting of a subject, a predicate (also called as property) and an object. OWL and RDF have much in common, but OWL is a language with better descriptive power than RDF.

4.2

Basic Concepts in Probability

The basis element of Bayesian probability is the random variable which is considered as referring to a state (or an event) of the initially unknown environment of agents. Depending on the type of domain values, random variables are typically divided into three kinds: boolean, discrete and continuous random variables [3]. In the OntoBayes model only discrete random variables are used. The basic probabilistic concepts used in OntoBayes are the prior and conditional probability. In order to describe what is known about a variable A in the absence of any other evidence Bayesian probability uses the prior (or unconditional) probability. It is often called simply the prior and written as P (A). Once agents have observed some evidence which has influence over the previously random variables, prior probabilities are no longer appropriate. Instead, we use conditional (or posterior) probabilities P (A|B). It indicates the dependency or causal relation between variables A and B: A depends on B (or B causes A). 4.3

Probability-annotated OWL

To express uncertain information or the risk of actions we need a probabilistic extension of OWL which enables to specify probability-annotated classes or properties. For this purposes we define three OWL classes [1]: ”PriorProb”, ”CondProb” and ”FullProbDist”. The first two classes are defined to identify the prior probability and conditional probability respectively. They have a same datatype property ”ProbValue”, which can express the probabilistic value between 0 and 1. The last one is used to specify the full disjoint probability distribution. It has two disjoint object properties: ”hasPrior” or ”hasCond”. The property ”hasPrior” specifies the relation between classes ”FullProbDist” and ”PriorProb”. It indicates that the instances of ”PriorPro” are elements of one instance of ”FullProbDist”. The property ”hasCond” describes the relation between classes ”FullProbDist” and ”CondProb” and it has a similar semantic meaning as ”hasPrior”. These two properties are disjoint, because any instance of ”FullProbDist” can only have one probability type: either prior or conditional. 4.4

Dependency-annotated OWL

The probabilistic extension of OWL alone is not enough for modeling our ontologydriven BN. It is required to specify the dependency relations between the random variables explicitly. In order to solve this problem an additional property element was introduced in order to markup dependency information in an OWL ontology. Before the relation of dependency is formally defined, it is necessary to introduce some notations which are influenced by the Object Oriented Programming (OOP) approach. We denote an object property s between domain class A and range class B as s(A, B). It is considered as an available operation of the subject class A to object class B. A datatype property d of class A will be denoted as A.d, where d is considered as an attribute of A. Now the dependency between properties can be defined as follows.

Definition 1. A dependency is a pair X ← Y , where each of X and Y is either a datatype property A.d or an object property s(A, B) . It is read as ”Y depends on X”. This definition clearly points out that each random variable in OntoBayes modeled BN is either an object property or a datatype property. The dependency relation was not specified between classes, but between properties, because this avoids possible errors when extracting a Bayesian structure from ontologies [1]. Based upon this definition we can now explain the term of a dependency chain as follows: Definition 2. A dependency chain is a list X1 ← X2 ← ... ← Xi ← ... such that, for each i, Xi ← Xi+1 holds. Then X1 is called root of this dependency chain, and for each i, Xi is an element in this chain. Now we can build, for each variable X in a dependency chain, a dependency graph for X from a given BN. The graph contains all dependency chains with the condition that every chain contains the element X. There are no dependency circles in any dependency graph, because it is preconditioned that each BN is a directed acyclic graph (DAG).

5

VKC within OntoBayes

Nowadays most decision support systems put much greater emphasis on the knowledge management. This is clearly justified. Making a decision once one has the right knowledge is often the easy part. Under an open, dynamic and uncertain environment, an agent can make decision more easily, precisely and rationally based upon corporate knowledge than based on the sole knowledge of an agent. OntoBayes facilitates itself as an integral part of a decision support system [1], but does not address social ability of the relevant agents. They can not cooperate and communicate with each other. In order to overcome this limitation the approach of VKC is investigated for enabling corporate knowledge within OntoBayes. The work in [10–12] showed the link between corporate knowledge and virtual knowledge communities. Under this proposal we integrate VKC into OntoBayes. This integration aims at equipping agents with a layer through which they have the ability to act as members of knowledge communities. Basically we can use VKC to integrate corporate knowledge in OntoBayes because both of them are ontology-driven models. However there are still some differences due to the fact that the knowledge base of OntoBayes contains Bayesian information besides the ontological one. We illustrate these differences in the sequel and adapt them to the OntoBayes knowledge. 5.1

Differences in Agent Modeling

As mentioned in section 3 by the agent modeling of VKC there are four notions: personal ontology, knowledge instance, knowledge cluster and mapper. In OntoBayes model, the personal ontology is not only the metalevel to the ontology

concept level, but also to the Bayesian concept level. Due to this difference we rename this notion as personal OntoBayes. Correspondingly the knowledge instances are the instances of objects defined in the personal OntoBayes, and the concrete probability distribution on the Bayesian level. The knowledge cluster contains a sub-part of an ontology and a dependency graph with probabilistic information, which can be shared among agents. The mapper is then not only an ontological mapper, but also a Bayesian mapper. 5.2

Differences in Community Modeling

It was pointed out in section 3 that the community modeling has three key notions: a domain of interest, community pack and community buffer. Each VKC has a domain of interest, which is represented with the adapted knowledge cluster as mentioned above. The community pack consists of knowledge cluster, normalized ontology and identification of leaders of this community. Here, except the knowledge cluster, the normalized ontology must also be adapted. Similar to personal ontology, we rename normalized ontology as normalized OntoBayes. It contains at least the head of the knowledge cluster of the community and additionally the dependency graph for all predicates about the head of the knowledge cluster in this community. The community buffer works like a blackboard system, so that it can be smoothly adopted here. It consists of messages following the syntax of the FIPA-ACL standard. 5.3

Knowledge Exchange with an Example

The main objective to use VKC in OntoBayes is to enable the knowledge exchange between individual agents. Table 1 shows what are the exchanges in VKC, particularly in the OntoBayes model. While only the ontological knowledge can be exchanged in normal VKC, Bayesian knowledge must be also exchanged additionally. Table 1. The exchanged knowledge in OntoBayes model Knowledge instance Instances of BN and their Probability distribution Ontological knowledge Ontological structure Instances of ontology Bayesian knowledge

Knowledge cluster Bayesian Networks

We give an example in Fig. 1 to show how knowledge exchange with VKC works in OntoBayes. The dashed-arrowed line shows the operation in an exchange process and the solid arrowed line shows the dependency relation between properties. The notation mentioned in section 4.4 is used in this example.

Community (insurance product) Community buffer Product.Premium

dependsOn write

Agent Insurance

Product.Price read

dependsOn

Product.RiskCoverage

read Agent Customer

write

Product.Price dependsOn Buy(Customer,Product)

Fig. 1. A simple knowledge exchange example

There is a community with a domain of interest ”insurance product” in the community of communities. An agent ”insurance” joined this community and wrote a message in the community buffer. This message indicates that the price of an insurance product depends on the premium and risk coverage of the product. Besides this agent there is another agent ”customer” who has also interest in this community and writes a message. This message indicates that an action for buying an insurance product depends on its price. After the message input both of them read the messages from other agents in this community, to complete a simple knowledge exchange process. After the exchange each agent can adapt its old knowledge base respectively to its knowledge cluster and personal OntoBayes with the new information. The feature introduced in the above example is illustrated in Fig. 2. It must be pointed out that not only the structure of Bayesian Networks , but also the probability distribution according to the structure will evolve after the knowledge exchange. To assess the evolution of the probability distribution,

Product.Premium

Product.RiskCoverage

dependsOn

dependsOn Product.Price dependsOn Buy(Customer,Product)

Fig. 2. Knowledge evolvement following the knowledge exchange

methods such as conditional independency, Markov blanket etc. were introduced in [17]. 5.4

Knowledge Community Processes

Agents’ actions related to knowledge communities are the following ones: initiate, reorient, leave, terminate and join a community as well as exchange knowledge. Every agent can initiate a community by creating a topic and a community buffer and advertising about this community. Advertising is done through a specific agent called ”community of communities”, which has a central directory of all communities. All agents of the system are members of this community. At the same time of initiating a community, a community park is also created. It contains a knowledge cluster of the initiator, a normalized OntoBayes specified above, and the identification of the initiators. The information will be posted to the community buffer. Community reorientation is needed because the knowledge can not be uniquely considered at design time. It should evolve over time. Reorientation consists of sending a new community cluster to the community of communities. Agents can leave a community voluntarily, but it could be also forced out by the leaders. When a leader leaves a community, a new leader is required, if it is the unique leader in this community. Community termination consists of erasing the community buffer and its reference posted to the community of communities during the community’s lifetime. The community can only be terminated by one of the community leaders.

6

Conclusion and Future Work

This paper aims to give an overview of an ongoing, but long term, research project. There is apparently no similar work in the research domain of decision support system under uncertain knowledge. The paper builds upon the first version of the OntoBayes model [1] and the work of virtual knowledge communities [10–12], to model corporate knowledge under an open, dynamic and uncertain environment. OntoBayes provides an efficient way for representing uncertain knowledge in an ontological and Bayesian approach. The grounding theory for OntoBayes and VKC is the Agent Oriented Abstraction, which can describe in a fully generic way the concept of an agent and the society of agents. It is showed that such an abstraction mechanism leads to very practical applications for corporate knowledge [9, 10]. It was investigated in this paper that virtual knowledge communities can be applied to OntoBayes with some adaptations, because of the similarity: both of them are ontology-driven model, even though they differ themselves from the agent’s personal knowledge, knowledge instance and knowledge cluster. Although the VKC enhanced OntoBayes model in this paper is a considerable step towards corporate knowledge based decision support system, further investigations are still required. A system is being implemented for applications

in insurance and natural disaster management. Some of the future research directions deal with: – A trust mechanism for agents involved in uncertain knowledge sharing. – Methods to assess achieved decisions based upon corporate knowledge with possible conflicting information.

References 1. Yang, Y., Calmet, J.: Ontobayes: An ontology-driven uncertainty model. In: Proceeding of the International Conference on Intelligent Agents, Web Technologies and Internet Commerce (IAWTIC’05). Volume I., Vienna, Austria, IEEE Computer Society (2005) 457–464 2. Pearl, J.: Probabilistic reasoning in intelligent systems. 2 edn. Morgan Kaufmann (1997) 3. Russell, S.J., Norvig, P.: Artificial intelligence: A Modern Approach. 2. ed. edn. Prentice Hall (2003) 4. Shachter, R., Peot, M.: Simulation approaches to general probabilistic inference on belief networks. In: Proceedings of the 5th Annual Conference on Uncertainty in Artificial Intelligence (UAI’90), New York, NY, Elsevier Science Publishing Company, Inc. (1990) 221–234 5. Bouckaert, R.R., Castillo, E., Guti´errez, J.M.: A modified simulation scheme for inference in bayesian networks. International Journal of Approximate Reasoning 14 (1996) 55–80 6. Yang, Y., Calmet, J.: Decision making in ontology-based uncertainty model. 21st European Conference on Operational Research (2006) 7. Calmet, J., Maret, P., Endsuleit, R.: Agent-oriented abstraction. Revista (Real Academia de Ciencias, Serie A de Matematicas) 98 (2004) 77–84 Special Issue on Symbolic Computation and Artificial Intelligence. 8. Weber, M.: Economy and society. University of California Press (1986) 9. Maret, P., Calmet, J.: Modeling corporate knowledge within the agent oriented abstraction. In: Proc. International Conference on Cyberworlds (CW’04), IEEE Computer Society (2004) 224–231 10. Maret, P., Hammond, M., Calmet, J.: Virtual knowledge communities for corporate knowledge issues. In: Proceedings 5th International Workshop on Engineering Societies in the Agents World (ESAW’04). Volume 3451 of Lecture Notes in Computer Science., Toulouse, France, Springer (2004) 33–44 11. Hammond, M.: Virtual knowledge communities for distributed knowledge management: A multi-agent-based approach using jade. Master’s thesis, University of Karlsruhe and Imperial College London (2004) 12. Maret, P., Calmet, J.: Corporate knowledge in cyberworlds. IEICE Transaction, Information and Systems Society E88-D (2005) 880–887 13. Bonifacio, M., Bouquet, P., Cuel, R.: Knowledge nodes: the building blocks of a distributed approach to knowledge management. Journal of Universal Computer Science 8 (2002) 652–661 14. Fischer, G., Ostwald, J.: Knowledge management: Problems, promises, realities, and challenges. IEEE Intelligent Systems 16 (2001) 60–72 15. W3C: (OWL Web Ontology Language Guide) http://www.w3.org/TR/owl-guide. 16. W3C: (Resource Description Framework (RDF): Concepts and Abstract Syntax) http://www.w3.org/TR/rdf-concepts. 17. Jensen, F.V.: An introduction to Bayesian networks. UCL Press (1996)