The fortnightly BRAG meeting will be held this Thursday, December 12th, at 1 pm via Zoom/GP-Y801. This week we will have presentations by Sarah and Dilishiya.
Come along and enjoy some light snacks at the final BRAG session of the year! 🍿🍩
Zoom link: https://qut.zoom.us/j/83437677358?pwd=jjkloCpL11QnzyCme0g02LS2JliXR1.1
Meeting ID: 834 3767 7358 / Passcode: 374799
Dilishiya’s Talk
Title: Bayesian Designs for Detecting Temporal and Spatio-Temporal Change via Mean Field Variational Bayes
Abstract: Bayesian spatio-temporal design provides a framework to guide efficient data collection across both space and time by incorporating prior knowledge and dynamically updating beliefs as new data become available, ensuring targeted and adaptive data collection in evolving environments. However, the practical application of Bayesian spatio-temporal designs is often hindered by challenges related to high dimensionality and computational complexity, limiting their usability in real-world scenarios. This study addresses these challenges by presenting a model-based Bayesian design framework that employs Mean Field Variational Bayes (MFVB) to optimally capture temporal and spatio-temporal changes in complex ecological systems. MFVB offers significant computational advantages over the commonly used Laplace approximation, providing more efficient approximations of posterior distributions for utility evaluations in design optimisation and typically capturing posterior distributions more accurately than normal-based approximations derived from the Laplace method. The framework is applied to ecological monitoring of shoals in Western Australia, focusing on detecting changes in hard coral cover across space and time. Two models are employed: one to capture overall temporal change and another to detect spatio-temporal variations. Both models are constructed within the generalised additive mixed modelling framework, recognised for its robustness in addressing model uncertainty, and Bayesian model selection, guided by the evidence lower bound, ensures the selection of the best-fitting model to inform Bayesian designs. Two types of Bayesian designs are developed: one tailored to detect consistent temporal changes and another aimed at identifying spatial variations over time. Utility functions are proposed to quantify the effectiveness of these designs in detecting changes accurately and precisely across space and time, and the approach is demonstrated for monitoring changes in hard coral cover at the Barracouta East shoal in Western Australia and compared to alternative designs.
Sarah’s Talk
Title: Calibrating models using non-empirical features
Abstract: Applied mathematical models – such as differential equation models or agent-based models – are often limited by a lack of data available for calibration. Many systems are data limited, due to the costs, time-constraints, or invasiveness of data collection. Yet, experts studying these systems often have rich knowledge of its behaviour and function that are often not considered as “data”. However, in highly uncertain applications, this knowledge is powerful for model calibration. In this work, we place non-empirical observations within an approximate Bayesian framework and combine this with any available data to constrain models, even when data is limited. We demonstrate these ideas using three examples: fitting coral growth models to expert knowledge, incorporating theory within ecosystem population models, and identifying biochemical networks that can adapt to changes in stimulus.