model - ampere
Transcription
model - ampere
Validation/Evaluation in Environmental Modelling AMPERE & PIAMDDI Workshop Seville May 28-29 2013 Tony Jakeman, Joseph Guillaume, Barry Croke, Sondoss El Sawah, Baihua Fu and John Norton iCAM, Fenner School of Environment and Society & National Centre for Groundwater Research and Training [email protected] Socioeconomic & environmental impacts of climate change, technology and water policy drivers in the Namoi catchment – adaptation opportunities Tony Jakeman, Jenifer Ticehurst, Rachel Blakers, Barry Croke, Baihua Fu, Wendy Merritt, Darren Sinclair, Neil Gunningham, Joseph Guillaume, Andrew Ross (ANU) Allan Curtis and Emily Sharp (CSU) David Pannell, Alex Gardner, Alison Wilson and Madeleine Hartley (UWA) Cameron Holley (UNSW) Rebecca Kelly (iSNRM and ANU) Steering Committee: State and local agencies, Namoi Water (irrigators) Integrated Model • Integrates the work of each of the disciplinary sub-teams • Three components – Social Bayesian Network using results of the social survey – Core integrated deterministic model • Simulates hydrogeological system, constraints on extraction, farmer decision making, crop yields and ecological impacts • With inputs of the possible practice changes , climate change scenarios and water allocation policies – An integrated trade-off analysis Spatial Scale Hydrological model zones Hydrological Model Development • A key challenge was the choice of hydrological model structure, including: – Surface-groundwater, groundwater level and routing sub-modules needed – Which hydrological processes should be simulated? – The spatial resolution – The level of process detail – conceptual or physics-based? • The driving consideration was the needs of the Integrated Assessment Project Fraction CMD Module NONLINEAR MODULE Effective Slowflow Rainfall Fraction Surface Storage Shallow Subsurface Storage Shallow Groundwater Storage Surface Water Extractions Surface Runoff Shallow Streamflow Sub-surface Runoff Discharge Recharge Infiltration Deep Groundwater Storage Groundwater Extractions SHALLOW GROUNDWATER Quickflow DEEP GROUNDWATER Temperature Rainfall SURFACE WATER Model Structure Natural Losses and Lateral Flow Overview questions • • • • What is the evaluation practice (and why)? What sorts of technical tools do we use? How useful are they? The need for avoiding poor predictions is obvious but what else is required: - To address uncertainty in problem definition - To ensure modelling is fit-for purpose - To constantly question assumptions • So what philosophies, processes and methods can we build on? 9 The underwhelming modelling practice • a lot of development of models of environmental processes and integrated models with some history matching/calibration – and modellers stubbornly preferring their familiar paradigm and methods • widespread acceptance of key models – incremental modifications the norm versus rigorous creation • much talk but less analysis of uncertainty, and even sensitivity, in those models; stress testing infrequent • scant discussion of model assumptions, strengths and weaknesses; very little frank reporting of uncertainties • model purpose and objective functions weakly argued and matched • stakeholders only beginning to be involved in the whole ‘process’ - for saliency, legitimacy and new knowledge 10 Reasons for poor practice • Too easy to publish models with modest but inadequate evaluation, and obvious overparameterisation masquerading as knowledge • Modeller s ignorance and/or complacency • Lack of resources • Too much science and technology push, not enough customer pull • Sheer volume and complexity of uncertainties 11 Some high-level solutions • Enhance publication standards and modelling guidelines eg Position Papers of Env Mod & Software • Improve education and training of modellers – including fora like MSSANZ and iEMSs • Set aside resources for model evaluation • Link the application context/purpose(s) to the model development and testing • Devise a workplan capturing a set of issues that should be addressed and how to identify and prioritise uncertainties including: • Begin/continue serious analysis, revision and associated documentation of common models 12 Quantitative techniques and model types: variously used • Error analyses of models, time variation of parameters – eg bias, correlation with inputs and conditions (rainfall-runoff example) • Global and regional sensitivity analyses and model emulation to improve model identifiability, discriminate between alternative hypotheses, and assess what uncertainties are crucial • Psuedo MC (inc. Bayesian) methods to quantify model uncertainty • Bounding methods to propagate through models and for solving inverse problems – non-probabilistic • Fuzzy logic approaches • Modelling families – various alternates and hybrids • Modelling intercomparisons • BUT we often forget part of the qualitative – e.g. sensibility/sanity and explanatory testing, use of stylized facts 13 And in IAM we have some bigger problems • Limited data – and only a handful of parameters can be identified in models, even when data is informative • Sensitive parameters can change according to driving conditions • The usually unquantified uncertainty may dominate – the neglected sources (see next slide) Ramifications • Catalogue, rank and appreciate all relevant types and sources of uncertainty • Analyse model components in context of their linkages • Evaluation will unlikely get us all the confidence we d prefer in our uncertainty bounds – but it can help reduce and define the margins of uncertainty 14 Unquantified uncertainty – the neglected sources beyond errors in inputs, outputs and parameters • Neglected sources in the decision pathway may dominate the usual quantified uncertainty eg - the conceptual model error (ie model structure assumptions) may dominate the numerical model - transfer from mental models to collective wisdom of conceptual model to numerical model (involves many judgments that beg for transparency - we might be solving the wrong problem - we could be iterating and solving a more appropriate and/or tractable one 15 Uncertainties in the decision pathway Identification Decision prompt Scope • Boundaries of analysis Monitor & evaluate • Trea.ng emerging concerns • Iden.fying need for change Development & Evaluation Iden=fy data Choose and knowledge methodology • Measurement error • Point of view • Representa.vity • Limita.ons • Imprecision • Assump.ons • Inaccuracy • Technical issues Implement • Adop.on • Compliance Commitment to ac+on Frame Search Deliberate Analyse • How inputs and outputs are represented in the model • ACtudes and rela.ons between stakeholders • Communica.on • Ranking of and trade-‐offs between objec.ves • Avoiding locally op.mal solu.ons • Missed alterna.ves For integrated modelling: • Model structure • Model parameters • Calibra.on method • Valida.on method • Technical • Integra.on Action Decision pathway step • Corresponding uncertain.es Some useful developing trends • Distinguishing between quantitative evaluation of the model, and evaluation of the process and outcomes (surveying) • Towards increased debate in framing the problem and weighing aspects and uncertainties • Use of many models and methods 17 Four useful concepts/techniques • Boundary Critique (Midgley 2011, Ulrich 2005) • Hypothesis testing in the broad • Sensitivity analysis • Robust decision making Guillaume et al. (draft): Beyond quantifying uncertainty: a review of concepts for predicting uncertain outcomes 18 Boundary Critique • Judgments define which models are and aren t plausible – inc. assumptions wrt structure, parameters and performance measures ie feasible sets • Secondarily, BC determines what problem the modelling is trying to address: as important as deciding feasibility • Debate and agreement where to fix the boundaries leading to a core ensemble of agreed models (ie inputs, structures, parameter values), and a marginal set of contested models • Examining areas of dissenting opinion help to avoid surprise failures • Reducing uncertainty is eliminating an unacceptable model from the ensemble 19 Hypothesis Testing: beyond the classical • Gives problem definition a central, broader and more explicit role • There are limits to probabilities w.r.t. epistemic uncertainties and stakeholder participation • Testing whether a stated conclusion could be false; whether the model survives an agreed testing process • Counter examples can be used for falsification • Encourages a designed approach: explicit statement of a problem and test conditions and the need to be critical 20 Sensitivity Analysis • Not just (first) what values of variables make a difference, but (second) trying to understand what (variables) cause a change • Methods for the second: eg global sensitivity analysis • Methods for the first: eg set inversion and scenario discovery • POMORE: estimates how far parameters have to change to obtain a different conclusion • Link with BC: ability to discuss model scenarios within a feasible set facilitates judgements 21 Robust Decision Making • Can extend RDM concept from avoiding ‘actions’ that lead to failure to making robust decisions about ‘knowledge’ • Expecting a problem definition to change • Identify model ‘scenarios’ that would cause high regret – anticipating surprise and/or searching for breaking ‘scenarios’ (hypothesis testing links here) • Becomes an iterative process addressing the boundary critique of feasible models and the problem definition 22 How do these four techniques support UA? BC • Constantly questioning assumptions • Constantly questioning problem definition HT • Testing whether uncertainty could change conclusions • Designing analysis with possible conclusions in mind SA • Trying to understand possible causes of failure of a prediction RDM • Making changes to counter possible causes of failure of a prediction 23 Main messages • Analyse models and errors for crucial weaknesses, reducing unnecessary complexity or building it rigorously • Focus on purpose, make limits of knowledge explicit, generate stress-testing model scenarios • Emphasize possibilities rather than probabilities • Evaluation can be a holistic process, not just a reporting exercise; focus on defining boundaries, scenarios and agreed feasibility to refine margins of uncertainty • Emphasise learning from evaluating scenarios • Uncertainty assessment will never be complete but opportunity is there for wholesale improvements • Develop a workplan among the communities to address model evaluation systematically 24 References • • • • • • • • Bankes, S. C., R. J. Lempert and S. W. Popper (2001). "Computer-assisted reasoning." Computing in Science & Engineering 3(2): 71-77 DOI: 10.1109/5992.909006 Guillaume J.H.A, Pierce S.A., Jakeman A.J. (2010) Managing uncertainty in determining sustainable aquifer yield. Groundwater 2010, Canberra, Australia, 31 October – 4 November 2010, from http://www.groundwater2010.com/documents/ GuillaumeJoseph_000.pdf Guillaume, J.H.A, Qureshi E.M. and Jakeman, A.J. (2012) A structured analysis of uncertainty surrounding modeled impacts of groundwater extraction rules, Hydrogeology Journal Lempert, R. J. (2002). "A new decision sciences for complex systems." Proceedings of the National Academy of Sciences of the United States of America 99(Suppl 3): 7309-7313, from http://www.pnas.org/content/99/suppl.3/7309.short Norton, J. P. (1996). "Roles for deterministic bounding in environmental modelling." Ecological Modelling 86(2-3): 157-161 DOI: 10.1016/0304-3800(95)00045-3 Reason, J. (2000). Human error: models and management. BMJ 320(7237): 768-770 DOI: 10.1136/bmj.320.7237.768 Refsgaard, J. C., J. P. van der Sluijs, J. Brown and P. van der Keur (2006). "A framework for dealing with uncertainty due to model structure error." Advances in Water Resources 29(11): 1586-1597 DOI: 10.1016/j.advwatres.2005.11.013 Rittel, H. W. J. and M. M. Webber (1973). "Dilemmas in a general theory of planning." Policy Sciences 4(2): 155-169 DOI: 10.1007/bf01405730