Tommaso Marzi

Lugano, Switzerland (CH)

About me

Ciao! I’m a Ph.D. student at USI (Lugano, Switzerland ) in the Graph Machine Learning Group, which is a research team of the Dalle Molle Institute for Artificial Intelligence IDSIA USI-SUPSI, under the supervision of Professor Cesare Alippi.

I’m currently working at the intersection of Graph Deep Learning and Reinforcement Learning.

Background

I obtained the Bachelor’s degree in Physics at the University of Bologna (Bologna, Italy ) in 2020, working on dynamical models for pedestrian dynamics in real-case scenarios. Subsequently, in 2022 I obtained the Master’s degree in Applied Physics from the same University, working on theoretical aspects of predictability and entropy for non-reversible Markov systems. In december 2022 I started my PhD at USI.

News

Mar 24, 2025	I’m co-organizing the Inductive Biases in Reinforcement Learning workshop, accepted at RLC 2025 (University of Alberta )! Check the website for more information.
Nov 24, 2024	Our paper Feudal Graph Reinforcement Learning has been published in TMLR!
Aug 01, 2024	Our paper Feudal Graph Reinforcement Learning has been accepted to EWRL 2024. See you in Toulouse

Selected Publications

TMLR

Feudal Graph Reinforcement Learning

Tommaso Marzi, Arshjot Singh Khehra, Andrea Cini, and Cesare Alippi

Transactions on Machine Learning Research, 2024

Abs Website

Graph-based representations and message-passing modular policies constitute prominent approaches to tackling composable control problems in reinforcement learning (RL). However, as shown by recent graph deep learning literature, such local message-passing operators can create information bottlenecks and hinder global coordination. The issue becomes more serious in tasks requiring high-level planning. In this work, we propose a novel methodology, named Feudal Graph Reinforcement Learning (FGRL), that addresses such challenges by relying on hierarchical RL and a pyramidal message-passing architecture. In particular, FGRL defines a hierarchy of policies where high-level commands are propagated from the top of the hierarchy down through a layered graph structure. The bottom layers mimic the morphology of the physical system, while the upper layers correspond to higher-order sub-modules. The resulting agents are then characterized by a committee of policies where actions at a certain level set goals for the level below, thus implementing a hierarchical decision-making structure that can naturally implement task decomposition. We evaluate the proposed framework on a graph clustering problem and MuJoCo locomotion tasks; simulation results show that FGRL compares favorably against relevant baselines. Furthermore, an in-depth analysis of the command propagation mechanism provides evidence that the introduced message-passing scheme favors learning hierarchical decision-making policies.