Assessing Adaptive Decision Strategies

Despite the fact that adaptive-decision strategies are likely to be the better approach to climate change, and the approach policy-makers are likely to follow in practice, they have been infrequently considered in the policy literature (see Dowlatabadi, 1999, for one of the few such treatments). One important reason is that the analytic-policy literature has been largely dominated by the approach of finding the optimum response to climate change. This optimization approach is not conducive to assessing adaptive-decision strategies for two reasons. First, the analytic demands of optimization calculations usually require neglecting key feedbacks. Often these feedbacks are precisely those involved in an adaptive-decision strategy.

More profoundly, optimization is the wrong criterion for assessing adaptive-decision strategies because it assumes a level of certainty about the future that, if truly present, would obviate most if not all of the need for such strategies. Optimization finds a unique, best strategy based on specific assumptions about the system model, the prior probability distributions on the parameters of that model, and a loss function which represents society's values. But adaptive-decision strategies are most useful when there is deep uncertainty, that is, when we do not know with confidence the model, probabilities, or societal values, or where different stakeholders disagree about these things. In such cases, disagreements about optimum strategies can quickly reduce to arguments about alternative, unprovable assumptions or differences in goals.

Rather than optimization, the criterion for assessing adaptive-decision strategies ought to be robustness (Lempert and Schlesinger, 2000). Robust strategies are ones that will work reasonably well, at least compared to the alternatives, over a wide range of plausible futures. Robust strategies are advantageous because we can often identify them without specifying, or requiring stakeholders to agree on, specific models, priors, or societal values. In general, we can always identify, post hoc, models, priors, and values that make any given adaptive-decision strategy optimal. But in practice, the robustness criterion is useful precisely because it avoids the need for any prior agreement in those cases where a range of plausible scenarios is the best available representation of the information we have about the future.

The concept of robust strategies has a strong theoretical and practical pedigree. The idea is closely related to Simon's (1959) satisficing strategies and Savage's (1954) idea of minimizing the maximum regret. There is also much evidence that actual decision-makers in practice search for robust strategies rather than optimal ones (March, 1994). While these ideas have been familiar in the decision-analysis literature (Watson and Buede, 1987; Matheson and Howard, 1968), in practice they are infrequently employed on the climate-change problem because they are often difficult to implement for problems of any complexity. In recent years, however, an emerging school of what we might call multi-scenario simulation approaches (Lempert, Schlesinger, and Bankes, 1996, henceforth LSB; Morgan and Dowlatabadi, 1996; Rotmans and de Vries, 1997; Casan, Morgan, and Dowlatabadi, 1999), has begun to exploit the capabilities of new computer technology - fast computer processing; extensive, low-cost memory; and powerful, interactive visualization tools - to consider strategies under deep uncertainty. These methods use simulation models to construct different scenarios and, rather than aggregate the results using a probabilistic weighting, instead make arguments from comparisons of fundamentally different, alternative cases.

Our approach, which we call exploratory modeling (Bankes, 1993), is a multi-scenario simulation technique that explicitly implements these classic ideas about robust strategies. In exploratory modeling we use simulation models to create a large ensemble of plausible future scenarios, where each member of the ensemble represents one guess about how the world works and one choice among many alternative strategies we might adopt to influence the world. We then use search and visualization techniques to extract from this ensemble of scenarios information that is useful in distinguishing among policy choices. These methods are consistent with the traditional, probability-based approaches to uncertainty analysis because when such distributions are available, one can lay them across the scenarios and thus calculate expected values for various strategies, value of information, and the like. However, in situations characterized by deep uncertainty, the exploratory modeling method allows us to systematically find strategies that are robust against a wide range of expectations about the future and a range of valuations of that future.

We now provide a simple example that demonstrates how this concept of robustness and these exploratory modeling methods can be used to assess adaptive-decision strategies. This example, based on our 1996 work in LSB, addresses whether and under what conditions adaptive-decision strategies are a reasonable response to climate change. We begin by comparing the performance of the strategies advocated by the two competing camps in the debate over the Kyoto protocol. We define two alternative strategies, "Do-a-Little" (DAL) and "Emissions-Stabilization" (ES), to represent these camps. Each strategy reflects a prediction-based approach to climate change and is expressed as a given emissions path over the course of the 21st century. The former has little near-term reduction and is similar to the results of many economic analyses that assume relatively small long-term damages due to climate change (less than 2% of GWP) and relatively high abatement costs (more than 2% of GWP) to stabilize atmospheric concentrations of greenhouse gases. The latter, which returns global emissions to their 1990 levels by 2010 and holds them there until mid-century, is similar (though slightly more aggressive) than the reductions paths mandated by the Kyoto protocol. While the proponents of these camps do not explicitly reject adaptive-decision approaches, and the Framework Convention on Climate Change explicitly calls for policies that adjust over time in response to new information, the debate in practice is very much characterized by support or opposition to specific targets for the reductions of greenhouse gases.

We compare these DAL and ES strategies using a linked system of simple climate and economic models. Emissions of greenhouse gases determine their atmospheric concentrations, which in turn determine the change in global-mean surface temperature. These temperature trajectories determine the trajectory of damage costs, while the emissions trajectories generate a trajectory of abatement costs. We work in a cost-benefit framework, and measure the performance of each strategy as the present value of the sum of the hundred-year time series of damage and abatement costs. In particular, we focus on comparing the performance of these strategies as a function of three key dimensions of uncertainties we face about climate change (Lave and Dowlatabadi, 1993; Nordhaus, 1994), which we express as: (i) the sensitivity of the climate system to increasing concentrations of greenhouse gases, (ii) damages resulting from an increase in global-mean surface temperature, and (iii) the ability of innovation to significantly reduce the costs of abating greenhouse-gas emissions. In each case we define our plausible range of estimates, wherever possible, as the extreme range found in the published, refereed literature (a similar screening was used in Rotmans and de Vries, 1997).

We simulate the impact of uncertainty about the climate sensitivity on the change in global-mean surface temperature due to anthropogenic emissions of greenhouse gases (GHGs) with our energy-balance-climate/upwelling-diffusion-ocean (EBC/UDO) model (Schlesinger and Jiang, 1991; Schlesinger and Ramankutty, 1992,1994a,b, 1995). We allow the climate sensitivity to vary between 0.5°C and 4.5°C. The upper value is the high value from the IPCC; the low value is from Lindzen (1990). We express uncertainty about the damages due to climate change using a simple phenomenological impacts function of changes in global-average temperature, as in the practice of many integrated assessments (Nordhaus, 1994a; Manne and Richels, 1992; Peck and Teisberg, 1993, 1992). We consider cases where total aggregate damages at the end of the next century for a 3°C temperature rise range from 0.5% to 20% of gross world product (GWP), based on a survey of experts conducted by Nordhaus (1994b).

The crude innovation model used in LSB to simulate the consequences of uncertainty about the impacts of innovation, first used in HLS, assumes base-case greenhouse-gas emissions (Houghton et al., 1990) are reduced as non-emitting "fuel-switching" technologies diffuse through the economy with an S-shaped, logistic diffusion curve at some policy-determined rate 1/R. (We will describe a more detailed innovation model below.) The model builds the least-cost mix of emitting and non-emitting technologies to meet the exogenous energy-demand and policy-imposed emissions constraint. The model parameters are chosen so that our basecase innovation case reproduces the results of more detailed models (Manne and Richels, 1991; Nordhaus, 1991). We then assume that innovation can reduce the incremental costs of the non-emitting technologies at some fixed, but currently unknown, annual rate. Abatement costs also depend on the rate of reductions because, in those cases where emissions are reduced sufficiently quickly, emitting capital must be prematurely retired (assuming a normal lifetime of 30 years). This formulation captures in a crude manner the idea of inertia that may affect the choice between early and late action (Grubb, Chapuis, and Ha-Duong, 1995).

Figure 3.1 compares the performance of the DAL and ES strategies as a function of society's expectations about our three, key, climate-change uncertainties: the likelihood of extreme climate sensitivity, the likelihood of extreme damages, and the likelihood of significant innovation. For instance, the lower left-hand corner represents the expectation that the climate is likely to be insensitive to increasing greenhouse-gas concentrations, damages due to any climate change are likely to be small, and that innovation is unlikely to reduce the costs of abatement. In contrast, the upper right-hand corner represents the expectation that the climate system is very sensitive to greenhouse gases, damages are likely to be large, and innovation is likely to radically change the costs of abatement. The curved surface labeled A represents the boundary where we should be indifferent between the two choices, DAL and ES. Formally, this is the indifference surface where the strategies have equal expected values. To the left of the curve, we should prefer DAL; to the right we should prefer ES.

We created this visualization by running a large number of scenarios, each with one of the two strategies and a particular set of assumptions about the three parameters describing our uncertainties. We then laid probability distributions across these parameters as a function of these three expectations, and conducted a computer search to find the surface on which the expected values of the two strategies were equal. Note that this process, similar to the policy-

Probability of extreme Probability of climarte sensitivity extreme damages

Figure 3.1 Surfaces separating the regions in probability space where the expected value of the "Do-a-Little" policy is preferred over the "Emissions-Stabilization" policy, the adaptive strategy is preferred over the "Do-a-Little" policy, and the adaptive strategy is preferred over the "Emissions-Stabilization" policy, as a function of the probability of extreme damages, significant innovation, and extreme climate sensitivity.

Probability of extreme Probability of climarte sensitivity extreme damages

Figure 3.1 Surfaces separating the regions in probability space where the expected value of the "Do-a-Little" policy is preferred over the "Emissions-Stabilization" policy, the adaptive strategy is preferred over the "Do-a-Little" policy, and the adaptive strategy is preferred over the "Emissions-Stabilization" policy, as a function of the probability of extreme damages, significant innovation, and extreme climate sensitivity.

region analysis of Watson and Buede (1987), inverts the standard use of probabilities in decision analysis. Commonly, analysts begin with assumptions about the probability distributions of key uncertainties, which much be gleaned from methods such as expert elicitation (Morgan and Henrion, 1990), and then use these probabilities to find some optimum strategy. Instead, we report key probabilities as outputs of the calculations. That is, we report the probabilities that are implicit in advocating one strategy over another.

There are two important arguments we can make based on the comparison shown in Figure 3.1. First, the figure provides what we call a landscape of plausible futures (Park and Lempert, 1998), because it provides an explicit overview of the implications of the full range of uncertainties. Such landscapes are useful because they provide a common framework in which contending stakeholders to the climate-change debate can position themselves. Each stakeholder can find a portion of the landscape that reflects their view of the world and thus agree that the models we are using capture their assumptions and can reproduce the arguments they make from those assumptions. This process helps stakeholders find a language in which to express their disagreements. It also helps them buy-in to the analysis. We have shown this "climate cube" to audiences including both oil-company executives and ardent environmentalists and convinced them that their divergent views are captured within different regions in the cube.

Second, this figure shows why DAL and ES provide an unsatisfactory set of options. We cannot predict which point in this cube best represents the state of the world. Little scientific evidence is currently available, or is likely to be available anytime soon, to convince a reasonable individual whose beliefs place her on one side of surface A that she in fact should switch to the other camp. For instance, someone who believes the new, emissions-reducing technologies are likely to make future greenhouse-gas emission reductions inexpensive can offer dozens of examples of technologies that have in the past, and others that may in the future, dramatically change the cost of emissions reduction and other pollution prevention. Alternatively, someone who believes such emissions-reducing technologies are unlikely can point to dozens of stories about over-optimistic forecasts and technologies that failed to change the world. Because the technological future is fundamentally unpredictable, there is no evidence we can gather nor theorems we can evoke that will definitely sway this argument one way or the other. Thus, we cannot really know which side of the surface A in Figure 3.1 we are on. We can show, however, that guessing wrong can be very costly. That is, if we chose ES in a world where it turns out DAL would have been best, or DAL in an ES world, the costs can be significant. Thus, it should be no surprise that framing the climate-change problem around competing predictions of an unpredictable future should foster a hostile and unresolvable debate.

We next consider how an adaptive-decision strategy performs in comparison to these two static strategies. We posit a very simple, threshold-based adaptive-decision strategy that observes the damage and abatement costs once a decade and can make one correction based on these observations. The strategy begins with slow emissions reductions, as shown in Figure 3.2. If the damage exceeds a particular threshold after ten years, the strategy switches to draconian emission reductions. If not, the strategy waits another ten years. If the damage then exceeds the threshold, or if the abatement costs are below a threshold, then the strategy switches to rapid reductions. If not, it continues checking the damage and abatement costs every decade until the mid-twenty first century. We assume that once the strategy makes a mid-course correction, it makes no further observations nor reduction-rate adjustments. This particular adaptive-decision strategy is one, very simple exemplar, not necessarily the best such strategy, but sufficient for our purposes here.

Note that we express this strategy, not as any particular outcome for emissions-reductions, but as a specific process that will produce different levels of

Figure 3.2 Flow chart describing a very simple adaptive-decision strategy.

emissions reductions in alternative scenarios. Thus, in a crude way, we can examine the type of process the decision-makers ought to use rather than the particular outcomes they ought to achieve. This question is different from those generally addressed by optimization approaches which, by definition, assume that the decision-maker conducts a particular, idealized process. Our claim, however, is that modeling the process of decision-making not only helps answer questions more relevant and credible to decision-makers faced with deep uncertainty, it also (and probably not coincidentally) provides solutions that perform better across a wide range of scenarios than those based on optimum outcomes.

We compare the performance of this adaptive-decision strategy to that of the DAL and ES strategies, with the surfaces labeled B and C in Figure 3.1. Surface B, in the lower left-hand corner of the cube, is the indifference surface between the adaptive-decision strategy and DAL. If society's expectations about the future put us in this lower left-hand region, we should prefer DAL; otherwise we should prefer adaptive-decision. Surface C, in the corner of the cube closest to the viewer, is the indifference surface between the adaptive-decision strategy and ES. If society's expectations about the future put us in this upper-right-hand region, then we should prefer ES; otherwise we should prefer adaptive-decision. In addition, the adaptive-decision strategy never makes large errors, so the costs of choosing it incorrectly - where DAL or ES in fact turn out to be the right answer - are not large.

The adaptive-decision strategy performs better than DAL and ES over a wide range of plausible expectations about the future, and thus provides a more robust response than either of the two static strategies. For those expectations where the adaptive-decision strategy does not perform better than DAL or ES, the cost of choosing it compared to the best option is not large. (Later in this chapter we will formalize this notion of robustness.) In addition, if we convinced a group of opposing stakeholders that the different poles of the climate-change debate are captured within this cube, then they should now agree that the adaptive-decision strategy is the proper way to address the climate-change problem, whether or not they previously resided in the DAL or ES camp.

Was this article helpful?

0 0

Post a comment