A Petri-Net Based Approach to Measure the Learnability of Interactive Systems

We propose an approach to measure the learnability of an interactive system. Our approach relies on recording in a user log all the user actions that take place during a run of the system and on replaying them over one or more interaction models of the system. Each interaction model describes the expected way of executing a relevant task provided by the system. The proposed approach is able to identify deviations between the interaction models and the user log and to assess the weight of such deviations through a fitness value, which estimates how much a log adheres to the models. Our thesis is that by measuring the rate of such a fitness value for subsequent executions of the system we can not only understand if the system is learnable with respect to its relevant tasks, but also to identify potential learning issues.