By Ana Busic (INRIA, Paris)
Talk Abstract: Maximizing energy production in wind farms requires addressing wake effects, a phenomenon in which wind turbines create sub-optimal conditions for downstream turbines. Developing optimal control strategies is challenging, as high-fidelity models of complex aerodynamics are not tractable for optimization. However, promising results have been achieved using cooperative multi-agent reinforcement learning. In particular, the independent learning approach significantly increases power production in simulated farms.
Despite this success, the independent learning approach lacks a convergence guarantee due to non-stationarity. We frame the wind farm control problem as a transition-independent decentralized partially observable Markov decision process, where the interdependencies between agent dynamics are represented by a directed acyclic graph. We show that non-stationarity can be mitigated with a multi-scale approach and provide convergence guarantees for multi-scale Q-learning.
Joint work with Claire Bizon Monroc, Jiamin Zhu, and Donatien Dubuc.
Speaker Bio: Ana Bušić is a researcher at Inria (Paris Center) and the Computer Science Department at École Normale Supérieure (DI ENS), as well as an associate professor at PSL University. She leads the joint ARGO project team (Learning, Graphs, Distributed Optimization) between Inria and DI ENS. She obtained her PhD in 2007 from the University of Versailles. Her research activities lie at the intersection of stochastic operations research and machine learning. Her contributions include stochastic modeling, performance evaluation, Markov decision processes, reinforcement learning, and distributed control, with applications to communication and electrical networks. She is a member of LINCS (Laboratory of Information, Networking, and Communication Sciences), a research and innovation laboratory for academia-industry collaboration between Inria, the Institut Mines-Télécom, UPMC Sorbonne Universités, Nokia Bell Labs, and SystemX. She received the Google Faculty Research Award (2014).