Date & lieu : Lundi 26 mai 14h en Lagrange Gris (L101), Inria Sophia Antipolis
Titre : Distributed solutions for large Markov chains
Résumé : This talk focuses on computing the stationary distribution of large Markov chains, a key problem in applications like PageRank. For chains with billions of states, exact methods are infeasible, so we turn to iterative algorithms.I will introduce RLGL, an algorithm that empirically achieves exponential convergence. Despite its effectiveness, understanding its behavior analytically is challenging—convergence is proven only in a few simple cases. My PhD project aims to deepen the theoretical understanding of this process and improve its performance through smart heuristics.I will present some simple results under simplifying assumptions, and discuss open questions in the distributed setting.
À propos de Lorenzo : Lorenzo Gregoris, who is doctorant from Eindhoven University under the supervision of Nelly Litvak, visited us for the month of May. Lorenzo studied mathematics at the University La Sapienza in Roma, especially probability, statistical mechanics, algorithms and Deep Learning. His master thesis was entitled « The Random Path Representation of the Spin O(N) Model ». He has also a strong experience on teaching. Since October 2024, he started his PhD on the distributed algorithms for large Markov chains.
