Keywords
representing knowledge, uncertainty, intersubjective probabilities, imprecise probabilities, bias, environmental modelling
Start Date
1-7-2012 12:00 AM
Abstract
Environmental decision support intends to use the best available scienti-fic knowledge to predict the consequences of management alternatives. This raises 3 questions:(i) How to formally represent and quantify scientific knowledge?(ii) How to find adequate model structures and parameter values for predicting the behaviour of environmental systems under different driving conditions?(iii) How to implement efficient numerical procedures to actually calculate such predictions?Approaches to address all three of these questions will briefly be discussed. With respect to (i) an intersubjective interpretation of probabilities with an extension to imprecise probabilities is suggested as the most adequate representation of scienti-fic knowledge. Conceptual arguments in favour of this approach are discussed as well as problems of alternative approaches. To address (ii) the importance of con-sidering input errors, model structure deficiencies, and internal stochasticity of the modelled system is emphasized, as well as handling remaining systematic errors or bias in model output adequately. This typically requires Bayesian inference and the propagation of probability distributions through deterministic or stochastic models. Under (iii) numerical difficulties of Bayesian inference are briefly mentioned as well as approaches to overcome these. Of particular importance are adaptive Markov Chain Monte Carlo methods, Approximate Bayesian Computation (ABC) tech-niques that make it possible to simulate from the likelihood function instead of evaluating it, and recent approaches of statistically emulating computationally demanding dynamic models. Finally, an overview of research needs is provided.
Conceptual and Practical Aspects of Quantifying Uncertainty in Environmental Modelling and Decision Support
Environmental decision support intends to use the best available scienti-fic knowledge to predict the consequences of management alternatives. This raises 3 questions:(i) How to formally represent and quantify scientific knowledge?(ii) How to find adequate model structures and parameter values for predicting the behaviour of environmental systems under different driving conditions?(iii) How to implement efficient numerical procedures to actually calculate such predictions?Approaches to address all three of these questions will briefly be discussed. With respect to (i) an intersubjective interpretation of probabilities with an extension to imprecise probabilities is suggested as the most adequate representation of scienti-fic knowledge. Conceptual arguments in favour of this approach are discussed as well as problems of alternative approaches. To address (ii) the importance of con-sidering input errors, model structure deficiencies, and internal stochasticity of the modelled system is emphasized, as well as handling remaining systematic errors or bias in model output adequately. This typically requires Bayesian inference and the propagation of probability distributions through deterministic or stochastic models. Under (iii) numerical difficulties of Bayesian inference are briefly mentioned as well as approaches to overcome these. Of particular importance are adaptive Markov Chain Monte Carlo methods, Approximate Bayesian Computation (ABC) tech-niques that make it possible to simulate from the likelihood function instead of evaluating it, and recent approaches of statistically emulating computationally demanding dynamic models. Finally, an overview of research needs is provided.
