Multidisciplinary Institute in Artificial intelligence

Nowadays the volume and speed of data flows are constantly increasing. Many applications need to move from offline methods to sequential methods, such as online or reinforcement learning, that can make decision, acquire data, adapt to it and process it on the fly. Following this trend, Reinforcement learning (RL) algorithms have become increasingly popular for solving challenging optimization problems. Despite their success, they often require substantial computational power and lack optimal performance guarantees. If prior knowledge on the environment is available, it can be used to enhance learning efficiency, known as structured learning. This leads us to the research question that we want to tackle within the FunRL project : 

How to design reinforcement learning algorithms with optimal theoretical guarantees that exploit a (known or unknown) structure of the problem to solve.

This question will be developed in three directions: online control of queueing networks, which raises the important issue of stability and rarely visited states; Markovian bandits; and parametric learning in MDPs.

The FunRL project will have a positive impact on the MIAI ecosystem by:

– Contributing in core-AI research, focusing in RL, which is a fundamental part of AI;
– Developing new courses and formations on the subject in UCA and UGA;
– Enhancing new collaborations between members of the project that work on RL but, come from different laboratories and did not find the opportunity to work together before;
– Collaborating with industrial partners such as EDF and Criteo on key challenges such as the support of ecological transition and the fight against climate change