1. Analysis and Management of Vertebrate Populations & Communities March 6, 2015 1 Introduction to Adaptive Management

2. Outline 2 What do we do in the face of uncertainty?  Make decisions anyway  Conduct research to reduce uncertainty (then make a decision later)  But how do we determine the value of new information?  Both, simultaneously  Adaptive management  The roles of monitoring

3. Information 3 As scientists, we have a strong tendency to ask for more information But we can ask, will that information change our decision and enhance our performance. Remember, information comes at a price.  Is the information relevant to the decision?

4. Value of information 4 Expected value of perfect information (EVPI)  Analytical technique  Allows you to assess how much your management might improve if you can resolve structural uncertainty (or the loss in management performance in the continued face of uncertainty)  Can help you decide if it’s worth the cost of gathering information Expected value of partial information (EVPXI)  Expected gain with a reduction in some sources of uncertainty  Thus, it is the marginal gain in performance Expected value of sample information (EVSI)  The gain resulting from a set of observations that do not completely eliminate the uncertainty

5. EVPI Example 5 Decision: build artificial spawning channels? Max net economic value of the sockeye fishery ($M) Costs to build included in “Build” option Options WtModels Do not Build Build Channel 0.5 No response 240135 0.5 Good response 240564 from Walters (1986) In absence of new information: V(“do not build”) = (0.5)(240) + (0.5)(240) = 240 V(“build”) = (0.5)(135) + (0.5)(564) = 349.5* If you can resolve uncertainty: V = (0.5)(240) + (0.5)(564) = 402 EVPI = 402 – 349.5 = $52.5M

6. Class exercise 6 Runge et al. (2011) explored the implications of uncertainty for restoring an eastern population of whooping cranes Several hypotheses were considered to explain reproductive failures; (H1) nutrient limitation and (H2) human disturbance; we are completely unsure of which is best Several management actions were designed to increase demographic rates: (1) swap eggs in nests for those further along in incubation (from a captive flock); (2) restore meadows; and (3) conduct spring draw- downs Objective values (normalized and weighted) for each action and each hypothesis: ActionH1: nutrientsH2: disturbance Swap eggs0.2540.740 Restore meadows0.9920.363 Draw-downs0.8630.216

7. Class exercise 7 With certainty = 0.992(0.5) + 0.740(0.5) = 0.866 With uncertainty = 0.678 EVPI = 0.866 – 0.678 = 0.188 or +28% Action H1: nutrients H2: disturbance exp. value Swap eggs0.2540.740 0.497 Restore meadows 0.9920.363 0.678 Draw-downs0.8630.216 0.540 max0.9920.7400.678

8. EVPI: effect of model state 8

9. EVPXI 9 H2: harassment of incubating birds by black flies H8: human disturbance of nesting pairs H7: egg salvage H8: human rearing of cranes

10. ADAPTIVE MANAGEMENT 10 Reducing uncertainty while making decisions

11. Dynamic decisions 11 We approach recurrent decisions using SDM, but there are nuanced differences:  Decisions are linked over time (today’s decision affects future system state and thus future decisions)  Decisions can also be adaptive (the best state- dependent choice can change in response to what is learned)

12. Dynamic decisions 12 System state (t) System state (t+1) Return (t) Action (t) System state (t+2) Return (t+1) Action (t+1) Return (t+2) Action (t+2) System state (T) Return (T) Maximize cumulative sum of returns What does my model have to predict?

13. Dynamic optimization 13 Martin et al. 2011. An adaptive-management framework for optimal control of hiking near golden eagle nests in Denali National Park. Conservation Biology.

14. Dynamic decisions under uncertainty 14 System state (t) System state (t+1) Return (t) Action (t) System state (t+2) Return (t+1) Action (t+1) Return (t+2) Action (t+2) System state (T) Return (T) Maximize cumulative sum of returns environmental variation Uncertain system dynamics partial controllability partial observability

15. Adaptation 15 System Model Prediction Monitoring Observation System Model* Learning Adapt Action

16. A MANAGEMENT CONTEXT 16 Monitoring

17. Monitoring 17 Purposes  To determine if the objectives are being met  To assess the state of the system  To resolve uncertainty The development of the monitoring system should be tailored to these needs & driven by the decision context

18. Monitoring 18 (1) Progress toward objectives Maintain open canopy (<60% closure) pine stand, with understory vegetation cover of 15-25% pinegrass,  5% elk sedge, <1% exotics. (2) Management Trigger A management prescription calls for thinning a Ponderosa pine stand when the basal area is greater than 85 ft 2 /acre. (3) Learning What are the differential effects of mechanical thinning vs. prescribed understory fire on vegetation composition?

19. ADAPTIVE MANAGEMENT 19 Putting It All Together

20. Adaptive management 20 Seeks to optimize management decisions in the face of uncertainty, using learning at one stage to influence decisions at subsequent stages, while considering the acquisition of information in the optimization.

21. Adaptive Management 21 t = t + 1 Decision Survey Gather other data Revise Model Weights Make Predictions Calculate Utilities MonitorDecide Model Learn Alternative Actions Objective Function Monitoring System System Models

22. Adaptive (learning) 22 t = t + 1 Decision Survey Gather other data Revise Model Weights Make Predictions Calculate Utilities MonitorDecide Model Learn Alternative Actions Objective Function Monitoring System System Models

23. Management (optimization) 23 t = t + 1 Decision Survey Gather other data Revise Model Weights Make Predictions Calculate Utilities MonitorDecide Model Learn Alternative Actions Objective Function Monitoring System System Models

24. Optimizing 24

25. Learning 25

26. Adapting 26

27. Forms of adaptive management 27 Passive (learning a unplanned by-product)  Use best system model to make decision; then refine model through monitoring of outcomes  Use model averaging to make decisions; then update model weights based on a comparison of predicted & observed outcomes Active (taking informative mgmt actions)  Learn then do: set up experiment to reduce uncertainty; then modify mgmt based on what is learned (suboptimal)  Learn while doing (the problem of “dual control”): occasional probing of the system, with intent to balance need to learn with desire for maximum mgmt performance (optimal)

28. Double-loop learning 28 Often, problem framing is iterative  Start with a prototype structure  Perform some initial analysis  Revise the prototype  Implement & gain experience  Revise the structure… It is sometimes difficult to understand the core issues of a problem until you’ve implemented a prototype structured approach

29. Double-loop learning 29

30. Cycles of learning 30 ContextFramesActionsOutcomes Single-loop: incremental improvement of established routines Double-loop: reframing Triple-loop: transforming Pahl-Wostl, 2009, Global Environmental Change 19:354-365

31. Cycles of learning 31 Single-loop learning Are we doing things right? Double-loop learning Are we doing the right things? Triple-loop learning Who has the right?

32. What to remember 32 The value of information is useful for motivating an AM program  EVPI is the expected gain in performance with the elimination of uncertainty, or equivalently as the loss in expected management performance due to uncertainty  EVPI is the expected value of eliminating all (specified) uncertainty; EVPXI is the value of eliminating a subset of the uncertainties; EVSI is the value of reducing (but not eliminating) uncertainty Monitoring is useful for  Tracking progress toward objectives  Make state-dependent decisions (dynamic decision making)  Comparing predicted and observed consequences to learn Adaptive management is useful  For dynamic problems  With uncertainty that is important to management  Where actions are differentially informative  When monitoring is sufficiently precise to discern the most appropriate models