Séminaires, soutenances

Acceleration of MCMC methods by non-reversibility and factorization

par Dr Manon Michel (UCA)

Europe/Paris
Amphi Recherche (LPC)

Amphi Recherche

LPC

Description

During this talk, I will present non-reversible Markov-chain
      Monte Carlo methods, based on piecewise deterministic Markov
      processes (PDMP). First developed for multiparticle systems, the
      goal was to emulate the successes of cluster algorithms for spin
      systems and was achieved through the replacement of the time
      reversibility by symmetries of the sampled probability
      distribution itself. These methods have shown to bring clear
      accelerations and are now competing with molecular dynamics
      methods in chemical physics or state-of-the-art sampling schemes,
      e.g. Hamiltonian Monte Carlo, in statistical inference. Finally, I
      will explain how the factorization of interaction terms can lead
      to computational complexity reduction, for instance in presence of
      long-range interactions.