People

Research activity in 1997

Since MBR is known to have high computational costs, the interest for combining MBR with other problem solving methods (potentially less expensive) has grown in the recent years. In particular we have investigated how to combine Case-based reasoning approach with MBR in order to speed-up diagnostic problem solving, and making it more practical also for large applications. During 1997 we have developed an entirely new Retrieval module which is able to retrieve the most promising cases by taking into account easiness of adaptation and without exploring the entire case memory.

The development of this efficient Retrieval algorithm (whose performances are much better than the ones obtained with different case memory organizations) is a significant step in improving the overall performance of ADAPTER, the diagnostic system which is based on the strict integration of case-based and model-based and we have developed in the past years.

In co-operation with University of Amsterdam, we have also investigated how a multimodal system like ADAPTER suffers from utility problem (the increase of performance in one component does not implies an increase of performance of the whole system). We have carried on an extensive set of experiments in order to evaluate the benefits and the costs of the different modules and we have developed a complex cost model which is able to estimate the impact of modifications of the case memory on the performance of the whole system. We have also started the development of learning strategies in order to allow the system to automatically decide when new cases have to be learnt and stored in the case memory.

We have also investigated the relation of Model-based reasoning approach, in particular the one based on a logical representation of the behaviour of the system to model, with other representation formalisms. In particular, interesting results have been obtained in comparing logical approach to MBD with probabilistic models (in particular bayesian networks), despite the apparent differences our investigation has pointed out that both formalisms are strongly based on the assumption of completeness. The use of Petri Nets formalism and of the analysis mechanisms developed for such a formalism is investigated in order to model diagnostic reasoning in terms of concepts and methods typical of Petri Nets.

As regards compilation of model-based knowledge, an approach has been developed for compiling, from libraries of component models and a model of the structure of a device, functional dependency graphs of the type used in the IDEA system developed by the FIAT research center for vehicle diagnosis. The approach would allow an easier generation of such models and then a speed-up in the development of diagnostics for brand-new automotive systems.

A lot of attention has been given to the temporal dimension in model-based diagnosis. A general characterization of diagnosis with time has been provided, dealing with the different ways time can be taken into account in diagnosis (reasoning in different contexts, reasoning on temporal and dynamic behaviour, reasoning about time-varying behaviour). A characterization on when some of the temporal dimensions can be ignored has also been given. Specific results have been obtained on the diagnosis of dynamic systems, modelled by means of equations relating qualitative deviations of variables in the system: different definitions of diagnosis in such a context have been compared and a way of reasoning on the injection of abrupt faults into the system has been adopted to provide more accurate diagnostic solutions with respect to simpler definitions of diagnosis across time.

The research work in temporal reasoning has found application in the field of patient management and in temporal queries to to database in industrial applications.

Related to temporal reasoning is the issue of reasoning about actions and change. In cooperation with the Catholic University of Leuven, a general solution to the ramification problem, that is, reasoning on side effects of events, has been provided and embedded in a narrative-based language modelled after the Event Calculus.

Vittorio Brusoni: