 |
Center for Nonlinear Studies |
Importance sampling is essentially the only Monte Carlo method that can successfully estimate rare event probabilities. It is often necessary to adapt importance sampling to a specific problem. The cross-entropy method is one famous example of this. That method can fail in some easy circumstances, by finding the wrong importance sampler. This talk introduces some simple problems in which one can be assured of finding a good importance sampler via convex optimization. Elaborations of the algorithm can handle harder problems. For hard enough problems one loses the gaurantees but might still find an improved method. The main goal of this talk is to present some ideas for adaptive importance sampling methods to LANL people and then learn more about the specific importance sampling problems arising in grid science.
Host: Michael Chertkov