Meta-Reinforcement Learning
Andrea Tirinzoni
DEIB Ph.D. student
DEIB - Conference Room "Emilio Gatti" (building 20)
February 27th, 2018
11.00 am
Research Line:
Artificial Intelligence and robotics
Sommario
Meta-learning, the process of learning how to learn, has been widely studied in the supervised learning community during the last two decades. The need for intelligent agents able to solve complex control problems with only little experience, just as human beings are able to do, made this field emerge as a fundamental paradigm for reinforcement learning as well. The idea is to build algorithms that can themselves generate reinforcement learning algorithms, where the latter are highly customized for a specific domain. This way, the generated algorithms can achieve dramatically better performance than any general-purpose, hand-designed learning procedure. In this seminar, we provide an overview of state-of-the-art approaches to meta-learning. Although focus is given to reinforcement learning, we present algorithms that, based on a common formulation, can be independently applied to regression, classification, and control problems. We demonstrate the power of meta-learners by showing a series of recent successful applications. Finally, we propose some open questions in this field and provide ideas for future research.
Biografia
Andrea Tirinzoni was born in Morbegno, Italy, on July 18th, 1993. In July 2015, he received a Bachelor's Degree cum laude in Computer Engineering from Politecnico di Milano. In 2017, he received a Master of Science in Computer Science from University of Illinois at Chicago and a Master's Degree cum laude in Computer Science and Engineering from Politecnico di Milano. From November 2017, he is a Ph.D. student at Dipartimento di Elettronica, Informazione e Bioingegneria (DEIB) at Politecnico di Milano. His main research interest is machine learning, with a focus on reinforcement learning and transfer learning.