I am Samuele Tosatto , a passionate researcher in reinforcement learning . I am currently an Assistant Professor at the University of Innsbruck . I was previously a postdoc in the RLAI Lab, specifically in the R-LAIR headed by Prof. Mahmood . I took my Ph.D. under the supervision of Prof. Jan Peters at the Intelligent Autonomous System lab (IAS) , defending my Ph.D. dissertation “Towards Off-Policy Reinforcement Learning for Robotics”.
My current research goal is to develop reinforcement learning capable of directly learning on a real system . That is a challenging problem! Robotic systems are slow (compared to simulated systems), and therefore the learning algorithm must be sample efficient . Furthermore, the learning agent must safely interact with the system.
My previous research focused on off-policy reinforcement learning, allowing higher sample efficiency. Its offline formulation allows us to learn from a fixed dataset (usually provided safely by a human demonstrator). I believe that off-policy reinforcement learning plays a crucial role in developing algorithms applicable to the real world. I developed a novel algorithm capable of extracting the optimal policy from a few suboptimal trajectories.