The fortnightly BRAG meeting will be held this Thursday 25/08/22 at 1 pm via Zoom/GP-Y801. This week we will have presentations by @Paula Sobenko Hatum and @Laurence Davies.
Zoom link: https://qut.zoom.us/j/98910888234?pwd=STZDUEt0R3d1YzRRd0RLczdpeTdodz09
Password: brag@QUT (if prompted)
Paula’s Talk
Title: Predicting seagrass ecosystem response to an extreme climate event using Dynamic Bayesian network model
Abstract: Although mathematical models to assess seagrass resilience to climate change have been developed, these models account for only a small number of processes that affect seagrass dynamics and often fail to incorporate an understanding of how multiple stressors affect resilience synergistically. Furthermore, the predictive capacity of deterministic models is often constrained by the limited spatial and temporal resolution of input data. As extreme climate events (ECEs) are predicted to become more frequent and intense, understanding the direction and magnitude of ecological responses to these events is necessary for predicting their effects on an ecosystem. Since seagrasses are particularly sensitive to heat stress caused by marine heat waves (MHWs), this leads us to a considered research problem: How can heat stress from marine heatwaves be incorporated into the DBN to model the resilience of seagrass meadows? To respond to this question, we aim to develop a model that will provide an approach for capturing whole-of-systems dynamics using expert knowledge and a more feasible, smaller range of observed data to predict seagrass responses to the impact caused by the marine heatwave.
Laurence’s Talk
Title: Transport-Based Reversible Jump Proposals
Abstract: Transdimensional sampling methods are not often an efficient means to conduct Bayesian model selection and averaging. The efficiency of reversible jump Markov chain Monte Carlo (RJMCMC) proposals, measured by the number of required likelihood evaluations, is problem-dependent and is usually difficult to improve. Inspired by recent advances in Deep Neural Network based Normalising Flows for particle transport and density estimation, we demonstrate an approach to improve the efficiency of RJMCMC sampling by performing transdimensional jumps between reference distributions corresponding to each model. We further propose the use of an SMC sampler, allowing the application of flow estimation to any transdimensional static inference or inversion problem on continuous parameter spaces. In contrast to other RJMCMC proposal design approaches, our method does not assume that the conditional target densities are differentiable or uni-modal. The performance of our approach is illustrated for a number of simulated examples of increasing complexity.
Thanks,
Katie & Jamie
BRAG Co-Chairs