Active Inference one step closer to practical, real-world application
Magnus Koudahl defended his PhD thesis at the Department of Electrical Engineering on January 23rd.
![](https://assets.w3.tue.nl/w/fileadmin/_processed_/b/e/csm_Koudahl_Magnus_EE_VH_PROM_bbe7d94aba.jpg)
For his PhD research Magnus Koudahl focused on making the framework of Active Inference (AIF) practical. He has developed a version of Active Inference that works on freeform graphical models and brought Active Inference closer to being a practical engineering tool.
AIF has a rich history in the field of computational neuroscience where it has been used to design intelligent agents that solve a host of decision-making and control tasks. A core draw of AIF is that it imbues agents with a natural, epistemic drive to seek out novel information. This sets it apart from other approaches and makes it especially suited to tasks where information is scarce or expensive but can be acquired in situ. Yet despite its myriad promises AIF has not made the leap to engineering and practical application. Laying the groundwork for making this leap is the fundamental task tackled in this PhD project.
AIF becomes composable
Koudahl made several contributions of which the main ones are listed. Firstly, he derived the equations for AIF in linear Gaussian dynamical systems - an extremely common model for practical control tasks. Secondly, he developed a novel preference learning system by applying AIF to Gaussian process classifiers. This formed part of a larger system called AIDA which handled personalized audio processing for hearing aids. And finally, he has formally integrated AIF with the framework of factor graphs. To do so, he developed a novel graphical syntax for writing constrained free energy minimization problems. He then comprehensively translated AIF to the new syntax and derived the necessary results for applying AIF to practical applications. A key feature of this new description of AIF is that it now meshes with an established body of work on message passing algorithms for Bayesian inference. This means AIF becomes composable with off-the-shelf tools that are already used in production, bringing it one step closer to practical, real-world application.
Title of PhD thesis: A Factor Graph Approach to Active Inference. Supervisors: Prof. Bert de Vries and Dr. Thijs van de Laar.