We are happy to announce a seminar by Prof Lucy Marshall and Ms Yiyi Ma. We hope to see you all there.
Any questions, please feel free to contact: secretary.nswbranch@statsoc.org.au.
Date: Tuesday, 27th June 2023
Time:
12.30 - 13.00: Lunch
13.00 - 13.45: Junior Speaker: Ms Yiyi Ma
13.45 - 14.00: Break
14.00 - 15.00: Prof Lucy Marshall
Venue: 14SCO T2 Theatre, Macquarie University and via Zoom.
RSVP: Register at https://statsoc.org.au/event-5298441 or https://uni-sydney.zoom.us/meeting/register/tZ0sf-CvrTopGdUX4dkTWL3gcGK49mu6gYec
Speaker: Prof. Lucy Marshall
Title: Quantifying the known unknowns: Data science and machine learning in environmental analysis
Abstract:
Environmental models are essential tools for understanding complex natural systems and predicting the impacts of human activities on the environment. By simulating the behavior of physical, chemical, and biological processes in natural systems, environmental models can help make informed decisions about environmental management, policy, and planning. Environmental modeling has undergone significant advancements in recent years, transitioning from simplistic representations of ecological systems to sophisticated and integrated modeling platforms. With this advancement comes increasing recognition of the value of modern data science methods for model building, uncertainty quantification, and improved predictions about future environmental states and hazards. These methods generate huge opportunities to leverage the increase in computational power, big data sets, and novel sensing technologies for unprecedented model performance.
However, these opportunities heighten the debate about the use of methods like machine learning in disciplines like hydrology, ecology, or biogeochemistry, that traditionally has relied on the knowledge of natural and engineered systems for model development (also known as ‘physically-based’ models). While recent data science methods suggest great promise, will their advancement be limited by the divorce of machine learning models from the knowledge-base of physical systems?
This presentation will provide insight into these issues across a range of modeling scenarios with a focus on Bayesian inference and its application to environmental systems. We track the evolution of Bayesian methods from conceptual rainfall model inference, through to multi-objective analysis of integrated ecohydrologic and water quality models, through to hybrid machine-learning/physically-based models. Through a series of case studies, we demonstrate recent advances in data science in two areas: (1) novel approaches to uncertainty quantification by characterizing errors in different environmental data sets; (2) the use of hybrid machine learning methods for improved uncertainty quantification of high-dimensional environmental models. Overall, our studies demonstrate the power of data science for future and current modelling applications, but recognizes the need for improved training of future engineers and scientists to think probabilistically, and be expert users of data and models.
Biography:
Lucy Marshall is a Professor of Engineering and Executive Dean of the Faculty of Science and Engineering. She is a water resources engineer, with expertise in hydrologic modeling, environmental model optimization, and quantification of uncertainty in water resources analysis. She has a special interest in understanding how environmental observations can be used to quantify uncertainty in systems undergoing change. Her research has spanned the development of new models in the most heavily instrumented watershed in the United States to making flood predictions in ungauged catchments across Australia.
Lucy completed her BEng (Hons), MEngSc, and PhD at the University of New South Wales (UNSW) in Sydney before moving to Montana State University as an Assistant Professor of Watershed Analysis, where she worked at the interface of engineering and environmental science. She returned to Australia as an Australian Research Council Future Fellow at UNSW, and went on to become the director of the UNSW Water Research Centre. She held multiple leadership positions at UNSW, as the inaugural Associate Dean (Equity and Diversity), Associate Dean (Research), and the academic lead for Athena SWAN. She joined Macquarie as Executive Dean in 2022.
Junior Speaker: Yiyi Ma (UNSW)
Title: A simulation study of the feedback particle filter with diffusion-map-based gain
Abstract: A general starting approach for investigating a complicated environmental system involving various parameters is summarising it using process-based models. On the other side, statespace models can be regarded as a further simplified version of process-based models, a set of differential equations consisting of observations and hidden states to describe a system's dynamic. A critical task for understanding such models is approximating the posterior distribution of the model at each time state, given initial conditions and historical information.
Ensemble methods (such as ensemble Kalman filters and particle filters) have achieved promising results for such tasks. A common feature of these methods is to rely on Monte Carlo samples to empirically approximate prior and posterior distributions. Although ensemble Kalman methods have shown great success in many real-world problems, they are inaccurate for filtering problems where the state space model or observation operators are nonlinear. Importance sampling-based particle filters are theoretically an elegant solution, but suffer from the curse of dimensionality. Feedback particle filter (Yang, P. G. Mehta, S. P. Meyn, 2013) provides an alternative solution for the non-linear filtering estimation task by introducing a control term to move the particles in state space, much like ensemble Kalman methods. This control term contains two main components: an innovation term and a gain function. The latter element is the solution to the Euler-Lagrange boundary value problem (EL BVP). (Taghvaei, P. Mehta, S. Meyn, 2020) provides a diffusion-map-based algorithm to estimate this solution. It also shows that a feedback particle filter with this diffusion-mapbased gain approximation achieves better results for one-dimensional stochastic processes than adopting a constant gain or using an SIR particle filter. We will explore the method's performance for higher dimensional state-space models, starting with the toy model Lorenz 63 to more complex ones related to environmental systems.
Yang, T., Mehta, P. G., Meyn, S. P. (2013). Feedback Particle Filter. IEEE Transactions on Automatic Control 58.10, pp. 24652480. doi: 10.1109/TAC.2013.2258825.
Taghvaei, A., Mehta, P., Meyn, S. (2020). Diffusion Map-based Algorithm for Gain Function Approximation in the Feedback Particle Filter. SIAM/ASA Journal on Uncertainty Quantification 8.3, pp. 10901117. doi: 10.1137/19M124513X.
Biography:
Yiyi is a second-year mathematics and statistics PhD student at UNSW. She graduated from Imperial College London with a BSc in Mathematics in 2020 and obtained an MSc in Statistics (Data Science) in 2021. Currently, she is interested in simulation-based likelihood-free inferences for spatial models and data assimilation, especially applying the techniques for nonlinear, non-Gaussian and high-dimensional state-space models.
@StatSocAus
