Active Acquisition of Information for Behavioral AI Agents

To this day, behavioral AI agents largely explore their environments through random sampling, rather than through targeted strategies aimed at acquiring information that is either directly relevant to their goals or valuable in improving the accuracy of internal representations at critical points. The latter objective –which we can term as “active seeking and acquisition of information”–is essential for building competent and recursively improving AI agents. Yet, it remains an open problem in classical architectures such as deep neural networks, in part due to their entangled and non-decomposable internal representations.
 
A recently proposed alternative paradigm, varsel networks [1], generates structured internal representations and shows promise in overcoming these limitations of conventional machine learning. Importantly, this paradigm may also lend itself to integrating targeted, intelligent mechanisms for information acquisition in behavioral agents.
 
The aim of this project is to develop such an extension of [1]. Your tasks will include studying varsel networks, implementing an extension of their base code that enables active information acquisition, designing simple test environments that demand active information acquisition for their solution, and evaluating your method against existing benchmark approaches on these environments.
 
Prerequisites: Fluency in Python. Prior familiarity with machine learning or planning is a plus, but not required.
 
[1] Erden, Z. D. (2025). Foundations of a New Learning Paradigm in AI Grounded in the Principles of Evolutionary Developmental Biology, Part III: Varsel Networks (Doctoral dissertation, EPFL).