Probabilistic Machine Learning and Decision Making

The next wave of automation in machines and services will consist of autonomous systems that are able to make their own decisions by leveraging artificial intelligence (AI) technologies. The ability of such AI systems to learn from data enables them to cope with situations which were not explicitly predicted at the design stage. The decision making should be a process internal to the machine learning system, as future data collection and potentials are governed partially by current decisions. In this project we will study inference in models of complex non-linear dynamical systems and rational decision making using modern probabilistic methods like Bayesian probability theory. We will study decision support or autonomous systems that are scalable, flexible and learning and explore their application in the domains of robotics, smart cities and computational chemistry.

Learning probabilistic models for sequential decision-making tasks poses several challenges. Firstly, exact inference under non-linear dynamics is usually intractable. This problem is exacerbated when working with high-dimensional and noisy observations such as images. Secondly, the model needs to be able to learn in the face of limited data availability, as collecting data can be laborious and expensive. Lastly, efficient learning of sequential decision-making tasks requires to trade off between exploiting knowledge about the task and exploring potentially better alternatives. Since information from the learned model can be used to guide exploration and increase data efficiency in an active learning fashion, well-calibrated models are crucial for such tasks.

The key objectives of this project are to improve the accuracy and computational efficiency of model inference, enhance the flexibility and quality of the learned model and policy, and guide exploration using model information. This work combines insights from probability theory, control theory and related fields to derive new machine learning models and inference procedures. We further develop a novel probabilistic data acquisition approach that is suitable for modern large-scale datasets and models. These approaches are empirically evaluated on both synthetic and real-world datasets. By comparing them to related work on carefully designed, repeated and reproducible experiments, we draw conclusions about their performance and limitations across decision-making tasks from different domains. By making contributions in the aforementioned areas we hope to advance the state-of-the-art in model-based learning for AI, particularly when data is scarce and noisy.