The goal of this project, supported by the Austrian Research Fund (FWF), is to research methods for advanced information processing in an agent-based environment for distributed information access. Methods for knowledge representation and reasoning are explored for the construction of rational information agents, in order to equip them with reasoning capabilities. Suitable methods and principles are researched, and their applicability will be examined on a particular application, to be built in the context of the IMPACT agent environment.
The importance of accessing data and information scattered over a number of different sites, which are connected through wide area networks such as the Internet, has been rapidly increasing. The World Wide Web (WWW) for example provides a huge wealth of vastly unstructured heterogeneous data which is difficult to search. Different approaches to tackle this problem have been proposed:
A common feature of many multi-agent systems is that the agents have limited (if any) reasoning capabilities. This capability is in particular needed during the decision-making process, when the agent has to determine which actions are appropriate in response to a given stimulus. However, advanced reasoning capability requires that knowledge -which serves as the input of the reasoning process- is represented in the agent's knowledge base in an appropriate way. In particular, aspects of advanced knowledge representation such as incomplete information, inconsistent information, or preferences need to be dealt with in an explicit, declarative way in the agent's knowledge base.
Incomplete information plays a role in different issues: e.g. in the specification of user queries, where optional input values to a query are left out; in the profile of a user requesting some information; in the information actually provided by some agent; on the cost of executing a query, etc.
Inconsistency may arise by merging different sources of information (e.g. two different yellow pages description), or when agents report different results.
Preferences naturally arise in various situations, e.g. if one out of several provider agents should be selected for contact, or if different options exist for receiving the same service.
Appropriate formalisms and methods are needed for incorporating above discussed aspects of incomplete information, inconsistency, or preferences in a declarative knowledge base. The proposed methods must be supplied with a formal semantics, in order to predict and understand the behavior of agents. Without a clear formal underpinning, verification of correct agent behavior is almost impossible.
The IMPACT system is being developed as a platform for creating and deploying software agents in an open agent environment. In the IMPACT system, a society of software agents is mantained by agents which are registered at designed IMPACT servers. Each software agent which wants to be a member of the society may join by registering at some IMPACT server, by declaring the services it provides in a uniform service description language (SDL). The registered agent must be able to understand and process requests that arrive through this interface, following a specified protocol.
As an important feature of this platform, it allows for the integration and reuse of legacy code: an IMPACT agent can be built on the top of existing code software by "agentifying" it through a kind of IMPACT "wrapper" which is put around this existing code. In this way, large information repositories can be made accessible for effective search without redesign from scratch.
IMPACT agents are event-triggered and act on the occurrence of particular events. Agent communication is done via a reliable communication system using message boxes. Possible reactions of the agents on an event are described in an agent program, which is a declarative languag for agent decision making.
While the IMPACT system provides a general framework for creating and deploying software agents, it does not provide special support for particular kinds of agents such as information agents. The design of particular applications involving information agents requests for additional methods and techniques, which have to be researched and developed in order to make the IMPACT approach broadely useful for such applications.
In this project, we tackle this shortcoming and want to research and analyze methods for the construction of (systems of) rational information agents equipped with reasoning capabilities. These methods should then be applied in the context of the IMPACT system. However, to warrant general applicability of these methods beyond the IMPACT approach, they shall be developed at a conceptual level, and different possible architectures for systems of information agents will be taken into account. In particular, the methods should address the aspects which are missed by current systems of cooperative information agents.
We plan to research the following aspects of knowledge-based information agents:
We shall investigate issues in the context of information agents for which these aspects are relevant, and provide methods for an adequate treatment. They will be used for constructing specialized reasoning components for information agents, which can be added to enhance their capabilities. A reasoning component can be viewed as a module which comprises a knowledge base and provides access to it through a well-defined interface. A component may be accessed by the information agent for solving particular tasks.
The methods we plan to facilitate reasoning components are backed on concepts from the area of knowledge representation, and involve formalizations of common-sense reasoning. These formalizations are relevant for the issues mentioned above, because the kind of problems human agents may encounter in dealing with daily-life situations are very much the same as the corresponding problems of artificial agents in a virtual multi-agent environment. For instance, humans are constantly forced to make decisions in order to carry on, but they seldom have all the relevant information which would actually be needed to obtain a valid conclusion. Hence, the human common-sense "fills in" these gaps and allows the drawing of rational conclusions which seems plausible given the current circumstances, but may become unwarranted once new, more accurate information is known. For this reason, formalizations of human common-sense reasoning are termed nonmonotonic, as the set of potential consequences does not necessarily grow with an increase of information, as prior "reasonable guesses" may have to be retracted.
We illustrate on a couple of examples how the problems addressed by nonmonotonic formalisms that are intended for modelling human agents naturally apply to software agents as well:
The feasibility of the methods and techniques that we shall research will be tested on a suitable application, to be implemented as a prototype in the framework of the IMPACT system. Examples of potential targets for the application are:
We believe that enhanced reasoning capabilities of information agents in applications like these will profitably increase the quality of the information service, and lead to better results. Moreover, having a firm theoretical basis of the applied methods is essential for properly understanding the behavior of large and complex environments like multi-agent systems. Conversely, being able to show that formalisms invented for problems of a similar nature, like systems describing human common-sense reasoning, is a healthy stimulus for that area as well.