Information Needs for Developing Robust Adaptive Strategies Robert Lempert Director, RAND Pardee Center for Longer Range Global Policy and the Future Human Condition Workshop on Climate Science Needed to Support Robust Adaptation Decisions Georgia Tech February 6, 2014

2 Climate-Related Decisions Poses Both Analytic and Organizational Challenges Climate-related decisions involve: Incomplete information from new, fast-moving, and sometimes irreducibly uncertain science Many different interests and values Long-time scales Near certainty of surprise How to make plans more robust and adaptable while preserving public accountability? Public planning should be: Objective Subject to clear rules and procedures Accountable to public

3 Concept of Decision Support Focuses Attention on How (in Addition to What) Information is Used Decision support: Bridges supply and demand with a focus on decision processes Represents organized efforts to produce, disseminate, and facilitate the use of data and information to improve decisions Includes as key elements: –Recognition that decision processes are at least as important as decision products –Co-production of knowledge between users and producers –Institutional stability –Design for learning Supply and demand of scientific information may be mismatched Sarewitz and Pielke (2007 ) NRC (2009)

4 Should Climate Science Inform Exploratory, Rather Than Consolidative, Models? Consolidative models: –Bring together all relevant knowledge into a single package which, once validated, can be used as a surrogate for the real world –Are often used for prediction Exploratory models: –Map assumptions onto consequences, without privileging any one set of assumptions –Cannot be validated –Can, when used with appropriate decision processes and experimental designs, provide policy-relevant information Bankes (1993) Weaver et. al. (2013 )

5 Consolidative Models Useful for ‘Predict Then Act’ Analysis This traditional approach to risk management works well when the future –Isn’t changing fast –Isn’t hard to predict –Doesn’t generate much disagreement What will future conditions be? Under those conditions, what is best near-term decision? How sensitive is the decision to those conditions? Analytics aims to provide ranking of strategies, which implies consolidative models

6 Predict Then Act Methods Can Backfire in Deeply Uncertain Conditions Uncertainties are underestimated Competing analyses can contribute to gridlock Misplaced concreteness can blind decisionmakers to surprise What will future conditions be? Under those conditions, what is best near-term decision? How sensitive is the decision to those conditions?

7 Believing Forecasts of the Unpredictable Can Contribute to Bad Decisions In the early 1970s forecasters made projections of U.S energy use based on a century of data Gross national product (trillions of 1958 dollars) Energy use (10 15 Btu per year) 0 Historical trend continued Scenarios

8 Believing Forecasts of the Unpredictable Can Contribute to Bad Decisions Gross national product (trillions of 1958 dollars) Energy use (10 15 Btu per year) 0 Historical trend continued Actual Scenarios 2000 Actual In the early 1970s forecasters made projections of U.S energy use based on a century of data … they all were wrong Deep uncertainty occurs when the parties to a decision do not know or do not agree on the likelihood of alternative futures or how actions are related to consequences

9 RDM Manages Deeply Uncertain Conditions by Running the Analysis Backwards 1.Start with a proposed strategy 2.Use multiple model runs to identify conditions that best distinguish futures where strategy does and does not meet its goals 3.Identify steps that can be taken so strategy may succeed over wider range of futures Proposed strategy Identify vulnerabilities of this strategy Develop strategy adaptations to reduce vulnerabilities “RDM (Robust Decision Making) Process” Analytics aims to provide concise summary of key tradeoffs, a useful function of exploratory models

10 Over 15 years, HCMC has planned multi-billion dollar flood investments using best available projections Flood Risk Management Study for Ho Chi Minh City Provides an Example

11 Conditions have diverged from projections and the city is at significant risk

12 Today, HCMC seeks an innovative, integrated flood risk management strategy How Can HCMC Develop This Plan When Today’s Predictions Are No More Likely To Be Accurate? World Bank WPS-6465 (2013)

13 Simulation Model Projects Flood Risk From Estimates of Hazard, Exposure, and Vulnerability Hazard Future rainfall intensity Height of the Saigon River Exposure Population in the study area Urban form Vulnerability Vulnerability of population to flood depth Risk = Expected Number of People Affected By Floods Each year Risk Model

14 Model Projections Depend on Values of Six Deeply Uncertain Parameters Rainfall Increase +0% + 35% Increase River Height 20 cm 100 cm Population 7.4 M 19.1 M Urban Form Growth in Outskirts Growth in Center Poverty Rate 2.4 % 25 % Vulnerability Not Vulnerable Very Vulnerable

15 Traditional Planning Asks “What Will The Future Bring?” Infrastructure performance under best estimate future Risk to the Non-Poor Risk to the Poor Risk Model Best Estimate Future

16 Risk to the Poor Infrastructure Under Alternative Future Risk Model We Run HCMC’s Infrastructure Plan Through 1000 Different Combinations of Conditions Alternative Future Future Risk to the Non-Poor What factors best distinguish this region, where risk is reduced for both poor and non-poor?

17 6% increase 45cm rise Infrastructure 1. Under What Future Conditions Is HCMC’s Infrastructure Vulnerable? Infrastructure fails to reduce risk when: > 6% increase in rainfall intensity, > 45 cm increase in river height, OR > 5% poverty rate 6% 45cm 5% In this region, risk is reduced for both poor and non-poor

18 6% increase 45cm rise 2. Are Those Conditions Sufficiently Likely To Warrant Improving HCMC’s Plan? Infrastructure IPCC SREX Mid Value (20%) IPCC SREX High Value (35%) MONRE SLR Estimate 30 cm NOAA SLR + SCFC Subsidence 75 cm Even stakeholders with diverse views can agree that its worth considering improvements to current plan

19 We Consider A Range of Options for Integrated Flood Risk Management 1. Rely on current infrastructure 5. Capture Rain Water “Soft Options” 3. Relocate Areas 4. Manage Groundwater 2. Raise Homes

20 Infrastructure NOAA SLR + SCFC Subsidence (75 cm) IPCC SREX Mid Value (20%) IPCC SREX High Value (35%) Elevate (23%, 55cm) + How Will Adding “Soft” Options Improve Our Strategy? MONRE SLR Estimate 30 cm

21 NOAA SLR + SCFC Subsidence (75 cm) What Are Tradeoffs Between Robustness And Cost? Stakeholders debate about how much robustness they can afford (which is much more useful than debating what the future will be) Relocate (17%, 100cm) High Cost Low Cost Rainwater (10%, 100cm) Elevate (23%, 55cm) Groundwater (7%, 55cm) MONRE SLR Estimate 30 cm IPCC SREX Mid Value (20%) IPCC SREX High Value (35%) x Infrastructure

22 RDM Approach Also Used to Help Develop New Plans for Managing Colorado River 2012 Bureau of Reclamation study, in collaboration with seven states and other users: Assessed future water supply and demand imbalances over the next 50 years Developed and evaluated opportunities for resolving imbalances

23 RDM Embeds Analytics in a “Deliberation with Analysis” Decision Support Process Scenarios that Illuminate Vulnerabilities Robust Strategy Deliberation Analysis Deliberation with Analysis Participatory Scoping Scenario Exploration and Discovery Key products include: 1.Scenarios that illuminate vulnerabilities of plans 2.New or modified plans that address these vulnerabilities 3.Tradeoff curves that help decision makers choose robust strategies New Options Case Generation Tradeoff Analysis

24 Decision Structuring: Work with Decision Stakeholders to Define Objectives/Parameters 1. Decision Structuring Deliberation with Stakeholders Metrics that reflect decision makers’ goals Management strategies (levers) considered to pursue goals Uncertain factors that may affect ability to reach goals Relationships among metrics, levers, and uncertainties Information needed to organize simulation modeling Also called “XLRM”

25 Case Generation: Evaluate Strategy in Each of Many Plausible Futures 2. Case Generation Simulating Futures Large database of simulations model results (each element shows performance of a strategy in one future) Strategy Plausible assumptions Potential outcomes 100s/1000s of cases

26 Scenario Generation: Mine the Database of Cases to Identify Policy-Relevant Scenarios Scenarios that illuminate vulnerabilities of proposed strategy 3. Scenario Discovery Uncertain input variable 2 1.Indicate policy-relevant cases in database of simulation results 2.Statistical analysis finds low- dimensional clusters with high density of these cases 3.Clusters represent scenarios and driving forces of interest to decisionmakers Uncertain input variable 1 Parameter 1 Parameter 2 Strategy successful Strategy less successful

27 Tradeoff Analysis: Help Decisionmakers to Compare Tradeoff Among Strategies Robust strategy or information to enable decision- makers to make more robust strategy 4. Tradeoff Analysis Visualization helps decisionmakers compare strategies

28 Analysis with Colorado System Simulations Reveal Key Vulnerabilities ,000 climate projections Recent historic Paleo records Model projections Paleo-adjusted model projections Several demand projections 24,000 climate projections Recent historic Paleo records Model projections Paleo-adjusted model projections Several demand projections Baseline strategy Reclamation’s Colorado River Simulation System + + RAND, RR-242-BOR

29 Used Climate Information to Assess Vulnerabilities of Current Management Plans Current plans fall short if: Temperature greater than 2°F Any decrease in precipitation

30 Initial Actions (dependent on beliefs) Initial Actions Contingent Actions Common Options Strategy Analysis Can Support Deliberation Regarding Near- and Longer-Term Actions

31 What about probabilities?

32 RDM Considers Sets of Alterative Probability Distributions Expected value of strategy s for distribution  (x) is given by HARD 1.Choosing what strategies to consider 2.Choosing what futures to consider 3.Calculating the performance of strategy s in some future x 4.Knowing – and convincing other people that you know – the true probability distribution EASY Calculating the integral for any once you have 1-3 above Thus RDM considers many probability distributions over the set of futures x -- NOT a uniform distribution

33 Some Strategies Are Robust Over a Wide Range of Probability Estimates This chart: Shows expected cost to taxpayers from re-authorizing U.S. Terrorism Risk Insurance Act Quoted on floor of US Senate by a proponent Called “insidious” by opponents Usefully informed Congressional debate CBO, Treasury Assumption RAND, MG-679-CTRMP

34 Observations Predict-the-act approaches promote restricted view of demand for climate information –Generally seek to aggregate information into consolidative models aimed at prediction “Backwards” analyses like RDM effectively employ a much wider range of products from climate science –Integrated models of entire system (exploratory and consolidative) –Models of individual components of full system –Scientific understanding not well-embodied in models –Any probabilistic information, however imprecise RDM creates demand for decision support methods and tools for managing and summarizing large and diverse sets of information in a decision-relevant context.

35 More Information Thank you!

36

37 Flexible Options Can Prove Effective

38 Research Portfolio Includes Developing and Evaluating Improved Tools Tools to draw meaning from information Users Tools to represent information Analytic tools to facilitate RDM Exploratory modeling methods run computer simulation models many times to create a database that links a wide range of assumptions to their consequences Scenario Discovery methods uses cluster analysis on these databases of model results to simply characterize the future conditions where a the proposed strategy does not meet its goals Tool development efforts include: 1.Facilitating design of strategies with “pareto-satisficing surfaces” 2.Higher dimensional scenario discovery 3.Adaptive sampling

39 Portfolio Analysis Reveals Key Tradeoffs Percent of SequencesVulnerable in Lowest Streamflow Conditions Lee Ferry Deficit VulnerabilityLake Mead Vulnerability

40 “XLRM” Framework for the Colorado River Basin Study 40 Uncertain Factors (X)Options and Strategies (L) Demand Conditions (6) Supply Conditions (4) Observed Resampled (103 traces) Paleo Resampled (1,244 traces) Paleo Conditioned (500 traces) Downscaled GCM Projected (112 traces) System Operations Conditions (2) Options for demand reduction and supply augmentation (40) Portfolios of many options designed to adjust over time in response to new information (4) Near-term actions Signposts Contingent actions Relationships or Models (R)Performance Metrics (M) Colorado River Simulation System (CRSS) Water delivery (5) Electric power (3), Recreation (11), Ecological (5), Water quality (1), and Flood control (1)

41 Colorado Basin Study Followed This Process Deliberate: Participants to decision define objections, options, and other parameters Analysis: Participants work with experts to generate and interpret decision- relevant information Dozens of workshops with many stakeholders over course of study Interactive visualizations Revised instructions 1.Start with proposed policy and its goals 2.Identify futures where policy fails to meet its goals 3.Identify policies that address these vulnerabilities 4.Evaluate whether new policies are worth adopting