This thesis deals with interaction between human and informed virtual environment (IVE). An IVE is a virtual environment including knowledge, which is, in our case, linked with virtual objects (Smart Objects). The purpose of this thesis is to enhance the coupling between human and IVE by allowing it to relevantly react to user’s behavior. To do so, we add to the IVE a decision-making module which is able to choose, in real time, a multimodal feedback (a set of aids) according to user’s activity which is observed by the system by using various sensors. We choose the belief function theory for decision-making in order to deal with the uncertainty and the incompleteness of collected data describing user’s behavior. In this way, our system brings an informed interaction: the reaction of the system to user’s actions is based on a priori knowledge combined with other knowledge acquired in real time. Thus, we bring a personalized interaction, adapted to each user, to increase its efficiency. The decision-making module includes an evidential network with conditional belief functions allowing graphically representing influences, a priori known, between different variables of the system. Input variables correspond to data collected in real time about the user: his/her actions and physiological state. These data can be uncertain (ex: data coming from sensors with some reliability) or incomplete (ex: failure of one of the sensors). The belief function theory allows us to represent these data with the uncertainty and the ignorance which are associated. These beliefs are then propagated in the evidential network to obtain, in output, a belief on the utility of the triggering for each available multimodal aid. The choice concerning the display of a set of multimodal aids, among those having an important utility, is done by solving a constraint satisfaction problem. Indeed, the set of multimodal aids to choose must respect a set of constraints which are a priori or dynamically set up. This allows ensuring the adequacy of the chosen solution with the system (ex: some aids can be incompatible with each other) and the adequacy with the user (ex: user’s sensory canals should not be overloaded with too many aids). Depending on the complexity of the problem, a good solution, without guarantee about its optimality, is computed by a genetic algorithm so as to be able to propose a feedback in a short due time. An information file for each user is created a priori and then updated by the system according to an estimation of the user’s preferences about the aids. These preferences are respected at best for the choice of the aids. This work has been applied to a fluvial navigation simulator in order to bring a training module allowing giving a personalized interaction to the learners with the simulator. The learner’s behavior is interpreted by the system (actions on ship controls, navigation errors determined from an estimation of the future position of the boat, stress level, etc.) and with these data the decision-making system determines the most appropriate multimodal feedback according to the current situation. The aids proposed can be, for example, visual assistances, audio messages and simplifications of the navigation conditions in order to help the learner to anticipate the maneuvers to do. On the contrary, in the case of experienced learners, the decision-making system will rather choose to remove assistances and to increase the navigation complexity. This personalization of the feedback for each learner brings autonomy in the training allowing a trainer to follow several learners at the same time. Other applications of our work could be considered such as, for example, driving assistance in augmented reality. An experiment has been realized to evaluate the contribution of our system for learners of different levels. Our system has been compared to a system without aids (control system) and to a system with non-adaptive aids. The results show that, between the beginning and the end of the training, novice learners obtained a two-time greater score with our system in comparison with the control system. The system providing always the same aids (training without adaptation) did not allow the learners to improve themselves. The learners’ questionnaire answers and the trainers’ comments show a real interest of our approach
