Peter C. Lockemann(Universität Karlsruhe, D)
Heinz Schweppe (FU Berlin, D)

Computer science has long pretended that information systems are perfect mirror images of a perfect world. Database management systems, e.g., work under the assumption that the data stored represent a correct subset of the real world. Of course, this idealized assumption is rarely true. Information systems contain wrong information caused, e.g., by data entry errors: This is a common problem for instance in genomic databases

• wrong information caused, e.g., by data entry errors: This is a common problem for instance in genomic databases
• imprecise or falsely precise information, e.g., a measuring device will provide information with a certain precision only. Typically, information systems store the measured date, but do not store information about the conditions under which this data is true and the precision achieved.
• incomplete information. A certain piece of information may not be available to the information system.
• inconsistent information. Different information systems may contain contradictory information.

In the past information systems have worked around these flaws by extensive consistency checking, plausibility checks, or human discovery and correction. These solutions are bound to fail as systems become ever more distributed, the information more globalized, and the individual systems more autonomous. Hence, we need to find ways for our information systems to directly deal with the uncertainty induced by them.

Nor is imperfection necessarily a bad thing. Take inconsistent information. It may reflect information collected under different circumstances or in different contexts, i.e., it may represent different views on the same phenomenon, and the sum total may very well carry more information than any single one.

The challenge, then, is to make system operation resilient to imperfect data. Resilience is not simply a matter of correction but more so of reconciling what appears contradictory information.

Meeting the challenge becomes particularly pressing when we consider the modern development of the computing environment into large-scale, open, mobile, extremely widely distributed systems. Even if everything works correctly, it will no longer be possible to guarantee consistency across such systems. Consider as an example a large-scale peer-to-peer organization. Each peer observes part of the environment only. Even collectively the peers will never observe the environment at the same instance in time. Hence, there is neither individually nor collectively a consistent image of the environment. This reminds one of the uncertainty principle in physics which states that predictions can be made only within certain probabilities. Consequently, what such systems need to incorporate is what is referred to as uncertainty management. We will need mechanisms that allow the individual components to function despite the fact that they have incomplete and maybe incorrect knowledge, and that the system as a whole reaches its goals by limiting the collective uncertainty of collaborating subsystems to acceptable levels of uncertainty.

Uncertainty is an issue that appears in many disciplines. The aim of the seminar is to bring together researchers from all these communities. Some of these communities have a long history of dealing with this problem, for others, it is a new challenge. These are in particular:

• Database Management Systems
• Multi-Agent Systems
• Peer to Peer Systems
• Sensor Networks
• Data Stream Management Systems
• Reputation Systems
• Context-Aware Systems
• Artificial Intelligence
• Information Retrieval
• Self-organizing Systems
• Semantic Web
• Market Economics, Decision Science
• Fuzzy Systems

Each of these communities needs to deal with the issues described above, and many of them do in their own ways. Unfortunately, up to now, there has been little exchange between the communities in their approaches. The outcome of this seminar should be a classification of the different types of uncertainty, an overview of the state of the art on dealing with them in the different communities, the applicability of these solutions to other types of systems, and an identification of promising avenues of research and possibly a joint research program.

The seminar was roughly structured along the following three areas:

Fundamentals, e.g. models for representing uncertain data, impact of uncertainty on (database) operations, consistency models, error correction

Methods for uncertainty reduction and inconsistency tolerance

Applications, e.g., obtaining information from sensor networks or stream management systems, structuring unstructured data, object localization, belief revision in agent and reputation systems, personal information management, data integration.

The seminar confirmed our belief that uncertainty management is an extremely important area of computer science research that will need contributions from a number of disciplines to be successfully tackled. The seminar identified a number of potential killer applications and many advantages of incorporating uncertainty management in information systems. The seminar provided an excellent basis for the initiation of such interdisciplinary work, but also for the exchange of ideas and the organisation of future collaboration among groups working in the same area, as is evidenced for instance by the probabilistic databases benchmarking initiative. Here, the "database-heavy" nature of the group turned out to be very beneficial to achieving a concrete outcome of the seminar. For a more detailed description of the results, please refer to the workgroup reports included in the seminar proceedings.