Skip to main content
Department of Information Technology

Probabilistic models gaining situational awareness

The smart systems needed by industry and society are becoming significantly more complex. The system design challenges include increasing levels of data and environmental heterogeneity, and this requires systems that are situationally aware and capable of dynamically adapting to a changing environment in order to function optimally. Our goal is to develop new models, inference algorithms, and automatic learning of controllers for complex tasks, in some cases directly from measured sensor data. Hence, within this theme we will develop solutions capable of blending modeling, inference and decision making.


In particular, we will consider a number of modeling approaches to situational awareness for dynamical systems, including: non-linear state space models, Bayesian nonparametric models (BNP), and flexible data driven discriminative models. The nonlinear state space model offers a general model of a dynamical system and when suitably combined with BNP models can obtain a solid and unified representation of the situational awareness capability. We see a very exciting and promising future here: our ideas on the combination of dynamical systems models and BNP models will provide the next generation of situational awareness models. Data driven discriminative models such as neural networks, on the other hand, offer a flexible approach to learn relations and appropriate actions directly from data in situations where explicit models are unavailable. Here our main challenge lies in how to incorporate such models into a lager complex system that also utilizes model based components. It is particularly challenging to combine these methods with probabilistic methods, as the probabilistic properties of data driven methods such as neural networks itself is a topic of current research. We will further study how to formalize and create appropriate abstractions for the the aforementioned methods so that they can be successfully incorporated into our probabilistic programming framework.

Specific objectives of this research theme:

  1. Develop new BNP models suitable for situational awareness.
  2. Develop methods for merging learnt discriminative models and probabilistic and model-based frameworks such the non-linear state space models.
  3. Identify suitable abstractions of current machine learning approaches so that the can be successfully combined in the probabilistic programming language.
Updated  2017-09-06 11:49:39 by Thomas Schön.