Quantifying uncertainty in the predictions of complex process-based models

点击翻译按钮获取中文摘要

Making responsible decisions about landscapes is facilitated by the use of complex models able to represent multiple competing demands on land use. Decisions about land use require that trade-offs between competing demands be identified, and their consequences through time be characterised. Methods for representing consequences through time on maps generally take the form of complex models such as stochastic computer simulations. Such models are increasingly used to make realistic predictions about real world processes from socio-ecological systems involving land use to the effects of climate change. Because these models attempt to simulate all relevant aspects of a real physical system, they may involve many parameters, some of which will be difficult to set correctly. As the final objective of these models is to assess the possible consequences of management decisions, such as the placement of wind turbines, it is crucially important that the uncertainty introduced by calibrating parameters be understood.Approximate Bayesian Computation, or ABC, is a promising technique for estimating parameter values together with their credible intervals, and this allows calculation of the uncertainty deriving from parameter calibration. The overarching aim of this proposal is to improve ABC methods to make them sufficiently fast and accurate that they can be widely used for the evaluation and calibration of complex stochastic computer models, and to quantify the uncertainty attached to their predictions. The end goal of the project is to be able to fit and evaluate the accuracy of complex models for real, challenging applications, and for this approach to be more widely used in practice. We will work with investigators in the landscape decision-making programme, and others involved in landscape decision modelling, to apply the methods we develop to their models. Our proposal develops and brings to bear cutting-edge mathematical and statistical methodologies to calibrate complex models, and to quantify the uncertainty in their predictions that derives from parameter calibration.