Using planning to predict and influence autonomous agents behaviour in a virtual environment for training

Virtual environments for training use technical systems simulation and virtual characters to put learners in training situations that emulate genuine work situations. In these environments, maintaining coherence is essential for the learning, whether in the perceived motivations of the characters or the reactions of the technical systems. However, with the complexification of simulated situations, it becomes difficult to maintain this coherence while exerting some control over the scenario, without having to define it explicitly a priori. We present in this paper the SELDON approach, which aims at dynamically adapting the scenario of a virtual environment for training to fit the learner's needs, and focuses on maintaining its coherence. We propose to generate this scenario by using a planning system with two different types of operators - prediction operators, and adjustment operators -, to influence the scenario unfolding in an indirect manner, while respecting the individual agent behaviours.