[Robot Learning Team Seminar] Talks by Prof. Matteo Papini (University of Milan) and Dr. Gianmarco Genalti (Politecnico di Milano)

This RIKEN AIP Robot Learning Team seminar pairs two reinforcement learning and online learning researchers for back-to-back technical talks on July 23. It runs over Zoom for all registered participants, with a members-only viewing room at the RIKEN Nihonbashi Office. Prof. Matteo Papini (University of Milan) opens with "Reusing Data in Policy Gradients to Improve Sample Efficiency." Policy gradient methods are typically sample-inefficient because they rely on on-policy data. Papini presents an actor-only policy gradient algorithm built on a novel multiple importance sampling estimator that re-uses trajectory data from the k most recent policies, with theoretical guarantees of improved sample efficiency and preliminary results for data reuse in PPO. Dr. Gianmarco Genalti (Politecnico di Milano) follows with "Bridging Rested and Restless Bandits with Graph-Triggering." He introduces the Graph-Triggered Bandits framework, in which a graph over the arms governs how each arm's expected reward evolves, unifying rested and restless bandits as special cases. The talk covers rising and rotting monotonic bandits, characterizes optimal policies, and discusses algorithms and their instance-dependent theoretical guarantees.