 |
 |
19
T.W. Alexander Drive P.O.Box 14006 Research Triangle Park, NC 27709-4006 Tel: 919.685.9350 Fax: 919.685.9360 info@samsi.info |
|
|
19
T.W. Alexander Drive P.O.Box 14006 Research Triangle Park, NC 27709-4006 Tel: 919.685.9350 Fax: 919.685.9360 info@samsi.info |
October
November
December
February
| 2/12/09 |
| 11/12/08 |
Mark huber suggested that we break into subgroups to focus on some specific questions raised in previous meetings. One such question is that of how small we can make the constant used by Dyer and Freize (1991) in the Product Estimator. Another idea is to explore the relation beween Parallel and Simulated Annealing and the Eberle and Marinelli papers. It was also decided that Wednesday November 19 will be the last meeting for this semester. We will reconvene in the middle of January.
| 11/05/08 |
Mark Huber discussed the papers of Eberle and Marinelli (see the Resources under the Password-Protected portion of the site) on convergence of Sequential Monte Carlo methods. He concluded that the these don't really improve on the simulated and parallel annealing of last week. One open question: can we relate the persistence of parallel annealing with the G_i defined by Eberle and Marinelli?
| 10/29/08 |
Mark Huber introduced parallel and simulated tempering, citing his work on both Conditions for Rapid Mixing of Parallel and Simulated Tempering on Multimodal Distributions and Sufficient Conditions for Torpid Mixing of Simulated and Tempering Annealing. The main interest is in whether mixing times are fast or slow in parallel and simulated tempering. Then if r is defined as a measure of the mixing time, we can use conductance and the second largest eigenvalue to bound r. The hope is that we can use the works of Eberle and Marinelli to improve estimates where tempering is slow.
| 10/22/08 |
Mark Huber gave an overview of the ideas presented in Adapted Simulated Annealing: A Near-optimal Connection Between Sampling and Counting by Stefankovic, Vempala and Vigoda. After introducing the notation of the paper, a cooling schedule and a B-cooling schedule were defined with examples in the Ising and Hard-Core Gas models. We were left with the open problem of finding cooling schedules for continuous problems.
| 10/15/08 |
If we want a uniform draw from a region A, where A is a subset of a region B from which we can take uniform draws, the basic Acceptance/Rejection algorithm involves analyzing the measure of A divided by the measure of B. But this quantity usually decreases exponentially in the size of the problem. Mark Huber introduced the Product Estimator, a method to avoid such problems. Although the product estimator proves to be flexible, this comes with a computational cost.
| 10/08/08 |
Because the first meeting was so sparsely populated, Mark Huber gave an extensive refresher on what was covered last week. With the new participants, many ideas were discussed for better analysis of the Randomness Recycler. Several relevant papers were mentioned, and will hopefully be added to the website resources for future meetings.
| 10/01/08 |
Mark Huber gave a presentation on the Randomness Recycler. He introduced perfect simulation and specifically, the Acceptance/Rejection method. The Randomness Recycler algorithm "recycles" randomness while undoing bias. So, unlike the A/R method, we don't have to throw everything away after the rejection step. James Lynch has provided a paper on ergodicity and Markov Chains, as a topic for a future working group meeting.
Entire site © 2002-2008, Statistical and Applied Mathematical Sciences Institute. All Rights Reserved.
This page updated on September 17, 2008 10:50 AM