The fortnightly BRAG meeting will be held this Thursday 06/04 at 1 pm via Zoom/GP-Y801. This week we will have presentations by @Matt and @Sarah.
Zoom Link: https://qut.zoom.us/j/86982060024?pwd=dHFHdm44R2NjYWNEbWhwTmFWQ09PZz09
Password: brag@QUT (if prompted)
Title: Coupling methods in MCMC
Abstract: A natural idea to improve computational challenges in Bayesian computation is to run multiple MCMC chains in parallel. Unfortunately, as each chain incurs a different (and unknown) amount of bias from initialisation the overall combined results cannot be trusted. Recently there has been an interest from the comp-stats community for a class of algorithms based on coupling MCMC chains. A pair of MCMC chains is said to be faithfully coupled if they meet exact and match at all subsequent iterations of the chain. I’ll introduce some key coupling ideas and how they can be used to remove bias and allow parallel computing for MCMC methods.
Title: Testing ecosystem models with experimental time-series data
Abstract: Calibrated ecosystem models are often used in conservation planning – to (a) forecast the populations of threatened species, (b) predict responses to conservation actions or (c) understand more about species interactions. These models are high-dimensional, nonlinear, and have limited data for calibration, yet they are rarely tested or validated. We test the capabilities of the generalised Lotka-Volterra ecosystem model when calibrated to experimental microcosm datasets and assess the accuracy and usefulness of modelling. We find that the models have poor predictive accuracy, give ambiguous predictions of how populations will respond to an action and yield confusing parameter inferences.
Jamie & Scott