Efficient modeling of temporally variable user properties with dynamic Bayesian networks

The overall goal of the project that I am working within is the automatic adaptation of the behavior of a mobile assistance system to a user’s resource limitations in order to realize a situationally appropriate presentation of instructions and information. Specifically, our assistance system models the temporally variable user properties of cognitive load, time pressure, and affective states. Because these properties are not directly observable, they have to be estimated on the basis of indirect evidence. Such evidence can be found, for example, in the user’s speech and motor behavior, in data from physiological sensors, and in knowledge about possible causes of resource limitations, such as the system’s own behavior or the user’s activities. The system needs to track the user’s state from moment to moment, taking into account previous states as well as new evidence. Dynamic Bayesian networks (DBNs) are a suitable computational framework for this problem, but they raise serious problems of computational complexity. Rollup methods must be applied that cut off older time slices but incorporate their impact on the remaining time slices of the DBN. There are two lines of research that I will first pursue in parallel and then bring together. One of these lines is to investigate relevant ways of increasing efficiency and the other one is to construct and test DBNs using relevant data.