WG4: Knowledge Extraction and Learning

Support for decision making is increasingly provided by computing systems of various kinds. Such systems increasingly consist of numerous computing devices that must learn about themselves and their environment in situations of dynamic change. Learning has traditionally played an important role in decision support and in particular machine learning by a single agent in a relatively static environment has been well studied. However, today's complex decision making problems involve numerous agents that must continually adapt to their environment and each other. There are very few theoretical results to guide us in understanding learning by multi-agent
systems. Specific subjects which are going to be considered include:

- Efficient Implementation of Learning and Classification Algorithms. The literature provides several different approaches for learning from examples (decision trees, boolean functions, rough sets) and has studied their complexity aspects. Less is known when the problem concerns making decisions from previous examples and from revealed preferences from past decision situations. The project will explore the problem of efficiently implementing learning algorithms for decision problems.

- Interactive Preference Learning. The combinatorial structure of the space of alternatives can be a serious problem for interactive preference learning. The acquired preference information may be incomplete, inconsistent. Furthermore, the user's preferences may even evolve during the elicitation process as the user discovers the space of possibilities. Interactive preference elicitation is able to repair the current preference model by acquiring new information from the user. The risk is that questions posed to the user become very complex due to the combinatorial problem structure. Explanation techniques can allow an identification of critical parts and a reduction of user interaction to relevant information.

- Critical Management. The management of complex decision situations arising from actual or potentially critical situations (homeland security, humanitarian disasters, conflicts etc.) often requires the use of past knowledge (best practices) and the possibility to synthesise the existing information (partial and uncertain) in useful indices to use for prevention purposes or immediate action. Data mining and data analysis techniques combined to decision theory turn out to be extremely powerful tools enabling a more efficient handling of critical situations where the time to response is a key issue.

- Decision compilation. Whereas decision theory targets problems in which difficult individual decisions need to made, many business decisions are made on the basis of decision tables or decision rules in a quite deterministic way. The project provides an excellent opportunity to study the relationship between elaborated decision making models as promoted by decision theory and fast decision deployment as promoted by the business rules community. The idea is to use decision-theoretic approaches to determine scenario-dependent decisions off-line and to compile them into rules, which can easily be deployed on-line. This approach can thus provide decision-theoretic foundation for business rules and thus improve the quality of everyday's decision making.