Our NSF grant “Bayesian Analysis of Nuclear Dynamics” has been awarded! This large, multi-institutional award seeks to bring the statistical tools of model mixing, calibration and uncertainty quantification to important problems at the frontier of nuclear physics. You can read the official announcement of this incredibly exciting project here and here.
Our latest BART paper, “Assessing variable activity for Bayesian regression trees,” has been posted to arXiv! In this work, Akira Horiguchi investigates alternatives to the usual split-count variable activity metrics. Building on the idea of Sobol indices, Akira finds that the popular split-count can, in fact, be quite misleading. Subsequently, an efficient algorithm for computing such indices in BART models is developed and an interesting connection to the usual split-count metric is described. We will be adding this feature to the OpenBT code base over the next few months! https://arxiv.org/abs/2005.13622