Versatility through Planning, Curriculum Learning, and Meta-Reinforcement learning

Matteo Leonetti
King's College London Professor

DEIB - Seminar Room "N. Schiavoni" (Bldg. 20)
December 6th, 2022
10.30 am

Contacts:
Marcello Restelli

Research Line:
Artificial intelligence and robotics

Sommario

On December 6th, 2022 at 10.30 am Matteo Leonetti, Professor at King's College London, will give a seminar on "Versatility through Planning, Curriculum Learning, and Meta-Reinforcement learning" in DEIB Seminar Room.

Truly versatile agents that can carry out diverse tasks have somehow eluded AI so far. I will present work in three directions that I think will play a role, in addition to reinforcement learning, in achieving multi-task generality: reasoning and planning on inaccurate models, to compute /reasonable/ plans and decide what not to explore; curriculum learning to shape exploration and form knowledge representations as the agent acquires new skills; meta-reinforcement learning, to generalise across similar tasks.

Biografia

I am lecturer at King's College London. I have been here and there: before KCL I was a lecturer at the University of Leeds, and before that a post-doc at the University of Texas at Austin with Peter Stone. Before even that I was a post-doc at the Italian Institute of Technology with Petar Kormushev. I received a PhD from Sapienza University of Rome working with Luca Iocchi. During my PhD I was lucky enough to spend six months at the University of Massachusetts at Amherst and learn the fundamentals of RL from Andy Barto himself. I also spent a little over a year at the University of Edinburgh, with Subramanian Ramamoorthy. In addition to people, I worked with different kinds of robots: 4-legged, 2-legged, underwater vehicles, and more recently laptops on mobile bases with or without an arm.