About this blog

Hello! I’m Ahmed Akakzia, PhD candidate in AI at the ISIR Lab under the supervision of Olivier Sigaud and Mohamed Chetouani.

I will be defending my PhD thesis in few days. The title is: Teaching Predicate-based Autotelic Agents: Learning Goal Representations with a Social Partner for Intrinsically Motivated Agents. I am interested in the study and design of embodied artificial machines that learn in open-ended environments. Such environments are distinctively non-stationary, where sudden drift might occur due to unknown dynamics. When interacting within these environments, artificial agents need to efficiently explore and discover interesting skills without any external rewarding signal. They need to be intrinsically motivated to represent, generate and pursue their own goals.

The key insights of my work are twofold. First, I believe that learning abstract and task-agnostic goal representations promotes the transfer and the adaptation of previously learned skills to unknown domains. Such representations could be based on some arbitrary spatial semantic relations that perfectly describe a scene regardless of the nature of the involved objects. Second, I think that developing open-ended repertoires of skills ought to involve not only the physical sensorimotor experience, but also external socio-cultural signals. Such signals, typically delivered by humans, would in principle guide the agents’ exploration and influence the learned goal representations. To make this possible, artificial agents need to be able not only to learn in autonomy, but also to benefit from external teaching signals. We call such agents Teachable Autotelic Agents.

This blog is a portfolio of my publications, including scientific papers and posts.

Thank you for reading!