**DIFAI ERC Advanced Grant Project**

# **Model-based shared autonomy via active inference** Supervised by Sebastian Stein and Roderick Murray-Smith. Application deadline, 30th April 2025 This studentship is linked to the [DIFAI project](https://difai-project.org/index.html) Applicants are invited for a fully funded PhD studentship (international fees + stipend at [research council rates](https://www.ukri.org/what-we-do/developing-people-and-skills/find-studentships-and-doctoral-training/changes-to-the-minimum-stipend-from-1-october-2023/). For 2024-25 the stipend is £19237 per year. The stipend is usually non-taxable and does not need to be paid back.) in a collaborative project between the University of Glasgow and [Aegean Airlines](https://en.aegeanair.com/), Athens. Aegean Airlines is interested in decision support systems that derive policies from learned forward models of interactions between Aegean, their customers, and their competition, combined with human input. Probabilistic generative models (diffusion models, Gaussian Processes, Probabilistic Programming Languages, ...) and associated learning algorithms are powerful tools for Simulation Intelligence with different trade-offs between data efficiency, sample fidelity, inference delay, and interpretability. Active Inference is a promising framework for sequential decision-making in partially observable Markov Decision Processes that casts action selection as inference in an internal generative model of the world, which in combination with different generative models and inference amortization provides different trade-offs for inference delay, context adaptivity, and decision quality. Active Inference appears particularly suitable for shared autonomy tasks, where aspects of both the environment and the user's intent are unobserved and models of both have to be learned, but where there is a strong imbalance in data quantity pertaining to the environment dynamics vs. the user input. How to design practically useful active inference algorithms for these regimes is an open research question. This PhD will focus on investigating how probabilistic generative models and (partially amortized, potentially hierarchical) active inference algorithms can be combined effectively for interactive decision support, which will be evaluated on sequential business decision-making tasks relevant to Aegean Airlines using their vast internal data and their business teams as users. Depending on the candidate's interests and abilities, the focus can range from core machine learning algorithm research to application-specific approximation and emulation techniques. Outcomes of this project may contribute to the literature on explainable AI. This PhD will be split between the University of Glasgow campus and the Aegean Airlines offices in Athens, Greece. Students will have access to the training and academic opportunities offered at Glasgow as well as the opportunity to interact closely with Aegean Airlines' business and data teams - to learn about their business decision-making processes and machine learning production pipelines; for feedback on prototypes and participation in user studies; to work with large-scale real-world datasets and potentially run experiments in production. The successful candidate will have demonstrable experience with at least one modern numerical python library (jax, pytorch, ...) and a strong background in generative machine learning models, Bayesian statistics, or signal processing and control theory. The PhD will be supervised by Dr. [**Sebastian Stein**](http://www.ssteinresearch.com/) and Prof. [**Roderick Murray-Smith**](https://www.dcs.gla.ac.uk/~rod/) at the University of Glasgow. How to Apply: Please refer to the following website for details on how to apply: [https://www.gla.ac.uk/postgraduate/research/computing/)