Automatic aggregation of qualitative reasoning networks

understanding and reasoning about the physical world is a task that is implied in large portions of both practical and theoretical education and training . In this paper we present techniques for the automated generation of domain knowledge models that support ITS functions for coaching these tasks . On the basis of a qualitative simulator (GARP), a complete reasoning network is generated for each exercise in a domain . However, a serious problem with such a network is that it easily becomes huge : it contains all necessary and grounded reasoning steps . In particular, complex coaching tasks like cognitive diagnosis (CD) soon become intractable . A solution can be found in 'summarising' the network into a hierarchical one by applying aggregation methods . The results these methods produce show similarities to the outcomes of learning mechanisms such as compiling out and chunking in human skill acquisition and in machine learning . The aggregated models present a more abstract view on the domain, in a way comparable to the result of a simulation using more abstract models . The major advantage of our approach, however, is that all different abstraction levels, plus their interconnections (i.e ., the abstraction operations) are available to the. tutor . The hierarchical reasoning networks are thus not only relevant for making CD more tractable, but also, and in particular, for enabling an ITS to communicate with the student about the reasoning at different grain size levels . A worked-out example is presented .