Sampling from multimodal distributions is a long-standing challenge in computational statistics. Annealed importance sampling (AIS) tackles it by moving a single sampler along a path from a tractable starting distribution to the target. The new method, proposed by Haoxuan Chen and Lexing Ying, replaces that single sampler with an interacting ensemble, combining AIS with population-based Monte Carlo techniques.
Interaction is introduced through either the snooker algorithm or the genetic algorithm used in evolutionary Monte Carlo. These moves let members of the ensemble share information, encouraging exploration of modes that might otherwise be missed. In the continuous-time and mean-field limit, the authors derive a partial differential equation that governs the ensemble's evolution, providing a theoretical foundation for the approach.
Experiments on both continuous and discrete distributions show efficiency gains over standard AIS. The paper, accepted by the SIAM Journal on Scientific Computing, is available on arXiv.