SKM 2023 – wissenschaftliches Programm
Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe
DY: Fachverband Dynamik und Statistische Physik
DY 6: Statistical Physics: General I
DY 6.11: Vortrag
Montag, 27. März 2023, 12:45–13:00, ZEU 160
Population Annealing and the Role of Resampling in Population Annealing — •Denis Gessert1,2, Martin Weigel3, and Wolfhard Janke1 — 1Institut für Theoretische Physik, Leipzig University, Postfach 100920, D-04009 Leipzig, Germany — 2Centre for Fluid and Complex Systems, Coventry University, Coventry CV1 5FB, United Kingdom — 3Institut für Physik, Technische Universität Chemnitz, D-09107 Chemnitz, Germany
Studying equilibrium properties of thermodynamic systems with rough free-energy landscapes particularly challenges standard Markov chain Monte Carlo techniques such as the Metropolis algorithm. Sampling can be improved by using generalized ensemble methods, one of which is Population Annealing (PA). Although PA is not expected to outperform its contenders in terms of time complexity, it is particularly well suited for parallel execution with no theoretical limit on the level of parallelism, which makes it a viable option on modern HPC.
In PA a population of replicas is collectively cooled down. At each temperature a population control step is carried out before applying some replica-independent update moves. This population control is realized by means of resampling. Here, we compare various different resampling methods and their performance in PA applications. Using the d=2 Ising model as a benchmark system, we identify two resampling methods that appear preferable over the widely used multinomial resampling. Further, we point out when different resampling choices affect the statistical quality of the simulation outcome and obtain some model-independent guiding principles for the choice of PA parameters.