- HAL publications
SIMSMART Research team
SIMulating Stochastic Models with pARTicles
- Leader : Mathias Rousset
- Type : team
- Research center(s) : Rennes
- Field : Applied Mathematics, Computation and Simulation
- Theme : Stochastic approaches
- Inria teams are typically groups of researchers working on the definition of a common project, and objectives, with the goal to arrive at the creation of a project-team. Such project-teams may include other partners (universities or research institutions)
SIMSMART is a computational probability and statistics research team, dedicated to the study of Monte Carlo algorithms within a mathematical perspective.
The main applications of interest are related to the simulation and statistical inference of stochastic complex dynamical physical systems; in particular systems arising in meteorology and computational physics at the molecular scale.
Using an appropriate level of mathematical abstraction and generalization, SIMSMART aims at providing non-superficial answers to methodological challenges related computational complexity reduction, statistical variance reduction, and uncertainty quantification.
Those challenges arise in a context where computational power is nurturing scientists into simulating the most detailed features of physical reality, relying on more and more cumbersome models, data sets, and (stochastic) algorithms.
SIMSMART’s main research topics are the following:
– Particle methods and rare event simulation.
– Advanced particle filtering and data assimilation (non-parametric inference, high dimension).
– Model reduction and sparsity.
Rare event simulation is ubiquitous in simulation, either to accelerate the occurrence of
physically relevant slow or rare phenomena or to estimate the risk associated with uncertain variables.
The increasing size of recorded observational data and the increasing complexity of models also suggest
to devote more effort into advanced filtering/data assimilation Monte Carlo methods, where managing high dimension, non-linearities and non-parametric models are current open challenges.
The need to simulate and compare complex and observed dynamical systems depending on uncertain parameters motivates the construction of relevant reduced-order models.
One of the main classes of algorithms that encompass the three latter topics consist of 'particle Monte Carlo methods'. In such methods,
several copies – a.k.a. particles – of the random system at stake are simulated, while being
split or killed according to some importance rules, for instance based on some real observations and its
associated likelihood (particle filtering), or on some score function (rare event simulation).
International and industrial relations
CEA/IRSN - Ifremer - French/Argentinian Climate Institute
Naval Group - Scalian Alyotech - EUMETSAT