An agile model for situation assessment: How to make Evidence Theory able to change idea about classifications

A relevant issue felt in the domain of Situation Awareness is related to the definition of models describing situations and threats of interest. Actually, the widely adopted approaches are based on two phases: employ training data as input of learning algorithms, and then validate the built model through other sets of data, gathered from the field. Model construction is therefore considered as an off-line process, and model correction is contemplated in terms of little adjustments in real-time applications. Great advantages could be derived by the employment of agile models, able to revise themselves evaluating model inconsistencies, contraddictions and errors, or taking into account user informations. In this paper, the analysis of model agility is conducted with regard to the Evidence Theory approach. The technique contemplates automated reasoning on time-independent models and it is therefore addressed to static pattern recognition, in the domain of Situation Awareness. The mathematical formalism adopted is the Transferable Belief Model, defined by Smets, able to simply identify modelling disagreements or deviations among several information sources. In this work, we investigate about possible metrics to adopt in the correction process. Firstly, we shown the classical model inability to identify two consequent situations; then we propose an algorithm using contradiction information, to allow the model to be aware of different, consequent situations and to change opinion respect to previous situation classified. Some simulation results related to a simple case study in the Critical Infrastructure domain are then reported.