Cole Monnahan1
1SAFS
February 07, 2017 3:30pm (PST): FSH 203
Introducing the no-U-turn MCMC sampler in ADMB and TMB: faster run times for large, complex fisheries models.
Bayesian analysis is a powerful tool for statistical inference in fisheries science. However, some analyses remain impracticable due to extremely long runtimes, despite increases in processor power. This is the case for data-rich age structured assessments written in AD Model Builder (e.g., stock synthesis) and large, complex spatio-temporal models for species distributions, typically built with Template Model Builder (TMB). Recently, the software package Stan, powered by the no-U-turn (NUTS) MCMC sampler, has emerged as state of the art for Bayesian analyses and it promises faster runtimes. However, for many fisheries scientists, it is not immediately feasible (nor desirable) to rewrite ADMB/TMB models in Stan. In these cases, it makes more sense to bring the algorithms to ADMB and TMB, rather than bring the models to Stan. In this workshop I will demonstrate Stan and how to interpret the diagnostic output from NUTS. I will then present and discuss my work on integrating NUTS into both TMB and ADMB. The workshop will contain a hands-on element in TMB, so bring your laptop with it installed, and will be very informal. Interested ADMB users can test their own models but need to compile an ADMB fork (contact me about this). I will expect that attendees are familiar with NUTS (read Monnahan et al. 2016) and the brave amongst you should try Betancourt 2017.
Monnahan, C. C., Thorson, J. T. and Branch, T. A. (2016), Faster estimation of Bayesian models in ecology using Hamiltonian Monte Carlo. Methods Ecol Evol. doi:10.1111/2041-210X.12681
Betancourt, Michael. A Conceptual Introduction to Hamiltonian Monte Carlo. arXiv preprint arXiv:1701.02434 (2017).