Possibilistic Assumption based Truth Maintenance System, Validation in a Data Fusion Application

Data fusion allows the elaboration and the evaluation of a situation synthesized from low level informations provided by different kinds of sensors. The fusion of the collected data will result in fewer and higher level informations more easily assessed by a human operator and that will assist him effectively in his decision process. In this paper we present the suitability and the advantages of using a Possibilistic Assumption based Truth Maintenance System (0-ATMS) in a data fusion military application. We first describe the problem, the needed knowledge representation formalisms and problem solving paradigms. Then we remind the reader of the basic concepts of ATMSs, Possibilistic Logic and Il-ATMSs. Finally we detail the solution to the given data fusion problem and conclude with the results and comparison with a non-possibilistic solution. 1 A DATA FUSION APPLICATION SEFIR (Systeme Expert de Fusion Interactive du Renseignement) is a prototype of a data fusion expert system which receives messages about observations provided by intelligence officers describing the nature, the number and the disposition of enemy units on the battle field and tries to derive the enemy formation [Lastic 89]. Units can be of several types (tank, motorised rifle, etc.) and are organised in hierarchical levels ranging from the higher division level down to the regiment, battalion, company and section levels. A data fusion process consists normally of three phases [Waltz 90]: • the correlation phase, that is the association and combination of different informations concerning the same unit, obtained from different sources; • the aggregation phase, that is the identification of a unit of a certain hierarchical level given partial evidence of its component units of the lower hierarchical level; • the fusion phase, that is the elaboration of a(some) consistent situation(s) given partial informations provided by several sources. The correlation phase is not implemented in SEFIR since we assume that each intelligence officer sends the informations concerning the units moving on the axis that has been allocated to him solely. Each unit is thus observed only once. The messages received by SEFIR contain symbolic and numeric informations such as: • the type and level of the observed unit (a tank section, for example); • the time of the observation; • the axis of enemy progress on which the observer is located; SEFIR aggregates the observations described in the messages into higher hierarchical level units relying on its (incomplete) knowledge of the enemy's organisation and doctrine. This knowledge is encoded in rules of the form: "if three tank sections are observed on the same axis within an interval of time of one hour, then they may belong to the same tank company". The uncertainty resulting from such incomplete domain knowledge induces the generation of a relevant number of somewhat contradictory hypotheses of possible aggregations of units for a given situation. For example, if four tank sections are observed on the same axis within one hour, it is possible to create two hypotheses of aggregation at the superior hierarchical level of companies as shown in figure l . 83 84 Fulvio Monai and Chehire Intelligence officer