Managing ecosystems with resist–accept–direct (RAD)

In recent years considerable interest has been generated in a new approach known as resist–accept–direct, or RAD, for managing ecosystems in the face of climate change. Under RAD, strategic responses to climate change are described in terms of three broad categories: resisting climate transformation, accepting the transformation and continuing to manage as best one can, and directing the transformed system toward novel ecological conditions. In particular, the potential for integrating RAD and adaptive management has been broadly considered, though absent a decision‐making framework needed for implementation. We propose a hierarchical decision scheme for RAD that accounts for strategy selection among the three RAD options, as well as adaptive decision making within each option. We use stochastic models and uncertainties about ecosystem processes to account for the dynamics of climate‐transformed ecosystems, and show how these features can be used to inform RAD strategies. Operationally, the approach involves decisions at two levels: one level involves choosing a policy for each strategy, and the second level involves deciding which strategy has the greatest policy value. The structure described here extends recent work in climate change adaptation, by including Markovian decisions under climate change, strategy‐specific policies, and value functions for assessing and selecting RAD strategies. We provide a hierarchical accounting of decisions and responses, and develop rules for the timing of those decisions. Combining RAD and adaptive management can help to organize thinking about ecological conservation under climate change, and focus attention on mechanisms for making decisions. We believe the structure presented here can facilitate conservation efforts under the non‐stationary climate conditions we are sure to face for the foreseeable future.


| INTRODUC TI ON
Climate change is universally recognized as a critical threat worldwide, with unprecedented changes that are projected to continue for the foreseeable future (IPCC, 2023).The severity of climateinduced change will depend on the path of future human activities in mitigating climate causes and limiting their effects on ecosystems (IPCC, 2023).A forceful response to changing climate conditions will benefit from new ways to understand climate change, control its drivers, and deal with its impacts.
A new approach for managing ecosystems in the face of climate change, known as resist-accept-direct (RAD), addresses strategies for dealing with climate-induced transformations.Much of the published literature on RAD builds on the work of Aplet (1999), Millar et al. (2007) and Aplet and Cole (2010) and others, in suggesting that management responses to climate-induced transformation are basically described in terms of three broad categories: resisting the transformation by attempting to maintain ecosystem structure and function; accepting the transformation and continuing to manage with existing policy options as best one can; and directing the transformed system to a novel configuration of structures and functions that can provide preferred ecosystem outputs.Schuurman et al. (2022) liken RAD to a sailboat being pushed away from its home port by strong winds, with the option to attempt a return to home port, move with the winds wherever they lead, or steer the boat to a new and preferred destination.
Put simply, RAD focuses on managing ecosystems transformed by climate, by resisting change, acceding to it or repositioning to a new ecological situation.
RAD is sometimes described as a new ecological management paradigm (e.g.Williams, 2022), though it lacks much of the technical detail necessary for actual decision making.A growing literature documents recent work on RAD by the Federal Navigating Ecological Transformation (fedNET) group and others (e.g.Clifford et al., 2022;Dassow et al., 2022;Feiner et al., 2022;Lynch et al., 2021;Magness et al., 2022;Rogers-Bennett et al., 2022;Schuurman et al., 2022;Thompson et al., 2021) as they attempt to address the threat of climate-induced transformations in aquatic and upland ecosystems.
In particular, a number of tantalizing suggestions have been offered about potential connections of adaptive management and RAD (e.g.Lynch et al., 2022).But many important issues about RAD implementation, and especially the use of adaptive management with RAD, are yet to be resolved.
A worrisome issue in discussions about RAD is an ambiguous meaning of the term "transformation" itself.The term is variably used to mean shifts in state (or regime, or ecosystem), or more generally, shifts of ecosystem characteristics.It has also been defined in terms of ecological thresholds, tipping points, ecological collapse, type conversion and other indices of change (Crausbay et al., 2022).In this paper, we develop some technical linkages between RAD and adaptive management that are needed for conservation with RAD.We use the phrase "climate transformation" to mean a climate-induced change in the processes that drive ecological responses to environmental conditions and iterative management.
In our approach to RAD, we assume a pattern of climate change that might be described as "punctuated equilibria," after Eldredge and Gould (1972).Ecological processes are held to be stationary (or nearly so), between transformative climate disruptions, thus allowing learning via adaptive decision making.Decision making will be seen to occur at two levels, one involving a choice among the three RAD strategies and the other involving learning-based decisions within a given strategy.We begin with a decision model for the dynamics of known and understood ecological systems, and then extend assessment to include uncertainties about ecological structure.

| DYNAMIC DECIS ION MAKING
We focus initially on known climate-transformed ecosystems that change over a discrete time horizon in response to fluctuating environmental conditions and time-specific management actions.
The ecosystem state at a particular time is represented here by x, with a the management action at that time and x ′ the subsequent state.System dynamics are represented with single-step transi- This framework implicitly assumes that the structures and functions driving system change are known, perhaps as a result of a lengthy record of decisions and responses under a stable regime of environmental fluctuations.Though uncertainty about system transitions is captured in the transition probabilities P x ′ | x, a , the probabilities are themselves assumed to be known, or can be treated as if they are.They can therefore be used in evaluating and selecting policies.
The model also includes an objective or value function V = V (x) that accumulates discounted returns r(a| x) over some time horizon, starting at some initial time t in state x: V serves as a metric with which to evaluate policy performance and compare the relative effectiveness of different policies.We assume that an extant policy 0 has been used over an extended time in the past, perhaps in an attempt to maximize returns during the period before transformation.
There is a large and well-known body of principles and methods for dealing with dynamic decision making (Bertsekas, 2017;Puterman, 1994).
In recent years, extensions of this work have been developed to address problems such as uncertainties about system structure, partially observable system states and non-stationary transition probabilities that are ubiquitous in ecology (e.g.Williams & Brown, 2023).
Non-stationarity is a particular conservation concern, as unpredictable changes mean that history may no longer repeat itself (Milly et al., 2008).Climate change and other major environmental stressors that alter the trajectory of system dynamics over time can lead to biased projections of future states and consequent management failures (Sutherland, 2006).Some ways of managing ecological systems for non-stationarity could be scenario planning, expert opinion and game theory (Sutherland, 2006), or exploratory techniques such as structured decision making (Martin et al., 2011;Nichols et al., 2011).
In the ensuing sections, we address RAD for known ecological systems and then introduce uncertainty about ecological structure, as related to adaptive decision making.

| R AD WITH KNOWN ECOLOG I C AL PRO CE SS E S
We first consider ecosystem structure and status to be known, and assume that climate change alters ecological conditions by inducing a change from P to Climate-induced change may affect the return function as well, with r(a| x) transforming to r(a| x).According to the RAD paradigm, three strategy options are available to a manager for dealing with such a change.

| Accept
One option is to accept the transformation, and continue to do the best one can to achieve objectives.If A consists of available management actions, the accept strategy calls for accepting and managing the new process P x � | x with policies consisting of actions a ∈ A , so as to obtain where policy consists of the optimizing actions Williams and Johnson (2013).The available policy actions for an accept option are generally the same as those before the transformation, though one now evaluates and selects policy based on the new climate-transformed process P and returns r(a| x).In most instances, the post-transformation process will produce trajectories that differ from pre-transformation trajectories, but no attempt is made to use policy to restructure the new process or replicate an earlier trajectory.Instead, one simply continues to manage in the face of the transformation going forward.
Accept strategies are often described in terms of monitoring, absent immediate management interventions (Magness et al., 2022).
Such a frame of reference may be appropriate when management is limited to the possible expansion of a preservation area that is protected from human intervention (Aplet, 1999;Aplet & Cole, 2010).
Here, we allow for active management in addition to monitoring, with post-transformation strategy based on the climate-transformed process P .Accept essentially means acceptance of the transformation, while continuing to manage as best one can in its presence.
In some ways an accept strategy is the simplest RAD strategy to implement, in that it does not pursue new goals or novel ecological dynamics.Instead, one continues to manage as effectively as possible with the transformed process, on the basis of accepted objectives and available policy options.A simple example that illustrates the approach is grazing-land management in the face of climate change.In this instance, management actions might consist of different levels of duration and intensity of grazing at different times during a grazing season, and policies constructed with these alternatives would be applied under different ecological conditions (Vallentine, 2001;Williams, 1985Williams, , 1986).An accept strategy involves sorting through and evaluating policy options based on these alternatives, using a new process P for grazing-land dynamics that has been climatically transformed from P .
In this example, the accept strategy continues to manage grazing as effectively as possible with available policy options, but given new environmental and ecological conditions.Though the trajectory of the new process model may veer away from previous patterns, no attempt would be made to recover the pattern of change.An analogue with preservation lands under climate change would be to continue a preservation strategy, perhaps with enhanced monitoring, and allow the system to develop without human interference.

| Resist
A second option is to resist climate-induced change, by attempting to restore dynamics back to those for the transitions P 0 x ′ | x that preceded transformation.That is, one recognizes the transformed process P x � | x and seeks a management policy 1 for the new process so as to approximate the pre-transformed situation: Note that policy 1 is not tied to the enhancement of value per se; instead, it is chosen to drive system processes to their preclimate conditions.That said, several policies may be available to achieve those conditions, and if so, a secondary goal might be to enhance value with policy selected from the constrained policy set.
Resistance can thus be viewed as a search for value based on a system transformed by climate change, where the search is limited to policies that seek pre-change system features.The manager essentially attempts to reverse the transformation and thereby sustain historical system dynamics.Only in that conditional sense is management used to enhance value.
A resist strategy is goal directed but also retrospective, in that it seeks to restore or recover ecological structure and function.
Depending on the ecosystem and level of disturbance, policies could include a mix of different kinds of actions, each with different levels of activity.Consider, for example, a situation in which limited financial resources are available for several activities in managing wildlife populations, for example, stocking to restore depleted species populations; restricting hunting to prevent further population declines; rebuilding habitat to sustain ecological diversity; or eliminating physical barriers to ecological resilience (Nichols et al., 2007;Silvy, 2012).
Within budget constraints, any combination of these activities could be included as a policy action for a given system state x.A comparative policy assessment would identify high-value policies as indicated above, but only for policy options that are designed to resist change and restore ecological functions and dynamics.

| Direct
Finally, the third RAD option is to direct the change from a climatetransformed process P to one with novel preferred features, by seeking a policy 1 so as to produce a target process P * with In this case, policy 1 is used neither to enhance value nor to recover pre-transformation conditions; instead, it is chosen to drive the system to a target pattern of system dynamics.If several policies are available to achieve the target, a conditional goal might be to seek value with policy selected from the constrained policy set.
Like the resist strategy, the direct option can be seen as a search over a constrained set of policy options that are limited by an overarching goal for system behaviour.But here the manager uses system controls to guide and direct the transformed system to a new configuration with a desired pattern of system dynamics.Only in that conditional sense would management be used to enhance value.Also like resistance, the direct strategy is goal-directed.However, in this case it is prospective rather than retrospective, seeking to drive the transformed system toward a preferred structure and behaviours by means of well-chosen policies.An available policy for this purpose would be composed of forward-directed actions at different levels, applied differentially for different system states.
For example, management actions might include the use of fire and revegetation to create new understory conditions; control of predators to stabilize prey population dynamics; translocation of populations to unused habitat patches to increase biodiversity resilience; or construction of dams and other physical structures to create new ecological conditions (Ausden, 2008;Lopez et al., 2024;Pereira & Hansen, 2003).Given budget constraints, any combination of these activities could be included as a policy action for a given system state, with the intended effect of creating novel conditions and preferred ecosystem services not experienced historically.A comparative assessment of the available policy options could be used for a secondary goal of identifying high-value policies.
For the foregoing RAD framework, best management policy and state-specific values can be identified for each of the three RAD strategies by means of dynamic programming and its extensions for adaptive optimization (Bellman, 1957;Marescot et al., 2013;Williams, 2011b;Williams et al., 2002).Choosing an appropriate RAD strategy as the system transitions among states then becomes a relatively straightforward matter of selecting the strategy with the largest state-specific value.Once selected, a RAD strategy is maintained until the system transitions to a state for which a different RAD strategy attains a higher value.At that time, the new strategy is adopted and maintained in turn until further system transitions lead to yet another state for which another strategy attains a higher value.In this way, the trajectory of RAD strategies essentially tracks the trajectory of state transitions.Adaptive selection of a RAD strategy is discussed in more detail in Section 5.2.

| ACCOUNTING FOR UN CERTAINT Y ABOUT ECOLOG IC AL PRO CE SS E S
The framework described in Section 2 for dynamic systems assumes that returns and transition probabilities are known, which is unlikely in most ecological situations.Relaxing this assumption allows for a more realistic and robust assessment of decision making, and in particular promotes the integration of adaptive management and RAD.It is the incorporation of uncertainty about system structure in decision making that defines adaptive management (Conroy & Peterson, 2013;Westgate et al., 2013) and allows for adaptive decision making in RAD.The uncertainty about ecosystem structure, function or processes is denoted as "structural uncertainty," or alternatively as "process uncertainty" (Williams, 2011a;Williams & Brown, 2014).
A familiar characterization of uncertainty uses multiple models with transition probabilities P k x ′ | x, a , k = 1, … , K, along with associated likelihood measures q(k) with ∑ k q(k) = 1 (Williams & Johnson, 2013).Each model describes the ecological system somewhat differently, with a corresponding likelihood that expresses one's relative confidence in that model as the most appropriate descriptor of system dynamics.As decisions are made over time and the system responds, the likelihoods are updated to reflect each model's effectiveness in representing system change (Williams & Brown, 2016).The value function V (x, q) includes an average of the models, along with the model state q ′ in V x ′ , q ′ obtained by recursive updating of q = q(1), q(2), … , q(K) with Bayes' theorem (McCarthy, 2007).
The argument (x, q) for the value function reflects the dependence of value on both the system state x and process state q.
Operationally, policy selection with V (x, q) accounts for both factors, as well the updating of the model state as information accumulates about system responses (Williams & Brown, 2016).The optimization produces an optimal policy consisting of the actions identified in the optimization (Williams & Johnson, 2013).
The foregoing is in fact a technical description of optimal adaptive management, a well-established protocol for decision making that explicitly deals with uncertainty and learning in managed ecological systems (Holling, 1978;Walters, 1986;Williams & Brown, 2016).By accumulating management returns and tracking system uncertainties over time, adaptive decision making pursues both management and learning.One important consequence is that structural uncertainty is expected to decline over time, as the process likelihoods gradually converge on the most appropriate model describing system dynamics.It is in this sense that adaptive management is held to be learning-based (Walters, 1986;Williams & Brown, 2016).
Adaptive decision-making under climate transformation is conceptually straightforward, though somewhat challenging in its details.Of primary importance is the suite of models used to characterize system dynamics, which must be updated to reflect the transformation.Let Pk x � | x, a and rk (a| x) represent one of K ′ processes characterizing the transformed system, noting that K ′ is not necessarily the same K in pre-transformation.That is, the number of models describing uncertainty about system processes after transformation may differ from the number describing the untransformed system.
With these new climate-induced features, decision making under uncertainty proceeds as above in assessing value and identifying policy.Thus, model-specific value functions are averaged to get for combinations (x, q) of system and model states.Because the process models Pk x � | x, a and return functions rk (a| x) after transformation differ from those prior to transformation, preferred policies prior and posterior to transformation are likely to differ, perhaps substantially.

| R AD UNDER UN CERTAINT Y ABOUT ECOLOG I C AL PRO CE SS E S
By extending the decision framework to include uncertainty about system processes, we can consider the reconciliation of RAD and adaptive management.There are actually two levels of adaptive management to be considered, one involving decisions about which of the three RAD strategies is selected and the other involving learning-based decisions within each strategy.

| Learning-based decisions within RAD strategies
From Section 3, each RAD strategy can be viewed in terms of a search among potential policies that are delimited by the particular RAD option.For example, an accept strategy may allow access to most or all actions available prior to transformation.Letting A 1 represent these actions and Π 1 the set of all policies that can be constructed with them, adaptive management under the accept strategy involves a policy search over Π 1 for enhanced value, with optimal policy consisting of the actions Though value and learning are sought within the general range of policy options that were available prior to climate transformation, the transformed process and return function are likely to project different consequences of policy into the future.In consequence, preferred policies for pre-and post-transformation are likely to differ, perhaps substantially.
A resist strategy has the same general features.Letting A 2 represent feasible restoration actions and Π 2 the set of all policies that can be constructed with them, adaptive management involves a policy search over Π 2 , again using the transformed features P x � | x, a, q and r(a | x, q) to project future policy consequences.
Finally, a direct strategy has the same general features as accept and resist strategies, but with different restrictions on available actions and policy options.Letting A 3 represent directed actions and Π 3 the set of all policies that can be constructed with them, adaptive management under the direct option involves a policy search over The bottom line is that within each of the three RAD strategies, adaptive management is implemented by identifying the range of available actions for a given RAD strategy and using P x � | x, a, q and r(a | x, q) to determine optimal values and policies over the corresponding policy set.

| Adaptive selection of a RAD strategy
In addition to choice of a policy within each of the RAD strategies, selection of a particular strategy can itself be viewed adaptively, where strategy values are used to inform decisions about which RAD option to use and when.An approach mirrors the two-stage optimization of dynamic programming, wherein future value is optimized conditional on a current action, then the current action is selected to optimize a combination of immediate return and future optimal value (Marescot et al., 2013).
By analogy, the current action consists of the choice of a RAD strategy, given that optimal values V 1 [x, q], V 2 [x, q], and V 3 [x, q] are determined as described in Section 5.1.With a two-step strategy, for a given combination (x, q) of ecological state and uncertainties one need only determine which of the three values is optimal, then implement the corresponding RAD strategy.An implementation rule retains that strategy as further actions are taken and (x, q) changes over time, until a combination of system state and process uncertainties is reached for which a different RAD strategy becomes optimal.At that point, the new RAD strategy is adopted, and is continued thereafter as (x, q) changes until yet another change in RAD strategy is indicated.
This decision approach is essentially hierarchical, with decisions about RAD strategy building on adaptive decisions that themselves are based on changing system status and learning.Let A 1 , A 2 and A 3 represent feasible actions for acceptance, resistance, and directed response respectively, with Π 1 , Π 2 and Π 3 the corresponding sets of policies constructed with them.Then RAD valuation can be expressed in terms of decisions at both levels: where optimal policy consists of the actions We note that the trajectory of the system state as well as the accumulation of evidence play key roles in the process of switching among RAD strategies.This can be seen by the presence of structural uncertainty q and system state x as arguments in the value above.It is the relative change in strategy values, rather than the accumulation of evidence alone or the path of a state trajectory alone, that is the actual driver of a change in strategy.
In discussing RAD, Williams (2022) points out that for any given system a manager may change strategies over time, perhaps resisting change as long as feasible, then accepting change and working with it, then later directing change toward new outcomes as opportunities are perceived and valued.The approach outlined here provides a technical context for the direction and timing of such changes.
It is useful to consider the information needs for RAD.In the absence of uncertainty about system processes, basic information for a RAD strategy is the first loop of standard double-loop learning (Argyris & Schön, 1978;Pahl-Wostl, 2009;Williams & Brown, 2018).
In general, double-loop learning can be described in terms of (1) a set-up or deliberative phase that identifies decision elements (prediction models, objectives, management actions and other decision elements); (2) an iterative phase that uses these elements in an ongoing cycle of managing and learning about system structure and the influence of management; and (3) an institutional learning phase involving the periodic reconsideration of the decision elements (Figure 1).
In the iterative phase of adaptive management actions are taken, outputs are assessed, and understanding of processes is updated and integrated into future decision making.Iterative decisions and learning are captured in and which can be seen to incorporate process models P k x ′ | x, a , management alternatives in A, objectives as specified in V(x, q), and monitoring of system responses (Williams & Brown, 2018).The sequence of policy identification and learning occurs over a timeframe during which the decision elements are thought to be fixed.
The second loop of double-loop learning includes revisiting and possibly changing these decision elements.The transformative nature of climate change, coupled with a natural tendency for objectives, options, and other elements to evolve as understanding accumulates and stakeholder perspectives and priorities change, underscores the need for periodic review and adjustment of the elements.A learning approach would include the recognition of when decision elements should be revisited, which elements should be adjusted, and how alternatives can be identified and incorporated on the basis of experience and management performance (Williams & Brown, 2018).Of special interest here are possible changes in models, management actions and objectives.
By accounting for the transformation of ecosystem processes induced by climate change, RAD offers a motivation and operational conditions for double-loop learning.Climate change transforms the processes that drive system responses, and RAD relies on an accounting of these transformations.As described in Section 5, differences among RAD strategies are generated in large part by differences in the policy actions available for formulating policy under each strategy, along with differences in the goals for resisting, accepting and directing climate-induced transformation.Variation in these features is a fundamental component of the RAD framework, Adaptive management displayed in terms of (1) a deliberative phase that identifies key decision elements (objectives, management actions, prediction models, etc.); (2) an iterative phase that includes monitoring, learning and adjustment of management strategy; and (3) an institutional learning phase that involves periodic reconsideration of the decision elements (from Williams & Brown, 2014).

| DISCUSS ION
In this paper, we describe the integration of adaptive management and RAD in a hierarchical decision framework for managing ecosystems in the face of climate transformations.We build on the work of the fedNET group and others, especially Lynch et al. (2022), who describe the basic problem formulation and discuss some issues and challenges in effecting the integration.We extend their development by considering technical elements that are needed to implement RAD through adaptive decision making.In particular, we incorporate stochastic models and uncertainties about system processes that underlie adaptive management, and show how they can be framed in terms of the three decision strategies of RAD.The operational effect is hierarchical decision making, involving policy selection for each RAD strategy along with strategy selection based on policy outcomes.Once a particular RAD strategy is selected, it is continued as decisions are made and the system responds over time, until a combination of system status and process uncertainty occurs for which a different RAD strategy offers greater value.
This framework advances the application of RAD for decision making by including explicit decision models, state-based policies and policy-based values and rules for the timing of RAD decisions within an apparatus for decision making.In particular, the recognition of value for different strategies provides a mechanism for determining when switching among the RAD alternatives would be appropriate.
Our technical description of RAD emphasizes ecosystem stochasticities and uncertainties in the movement to a new ecological configuration.This differs from some descriptions, which seem to suggest that RAD trajectories start after transformation and unfold from there in some deterministic or nearly deterministic way to reach a targeted ecological condition.Our scheme accounts for stochastic transitions among states, even if controlled by policy, so that state trajectories are themselves random and exhibit random variation as they unfold.The value function guiding policy selection accounts for this variation by calculating expectations of change over time, where policy selection for a RAD strategy is based on those expectations.
A special application of RAD involves a more classical experimental approach, in which management is implemented on aggregates of separate spatial units.In this situation, it becomes possible to apply different RAD strategies simultaneously rather than sequentially.Aplet and McKinley (2017) discuss unit features that are amenable to each of the RAD strategies.If multiple units are available in the three RAD groupings, one could use different policies on units in an experimental context, so as to spread the risk over uncertain outcomes and accelerate the learning about strategy effectiveness.Aplet and McKinley (2017) call such a strategy a "portfolio approach" to managing ecological risks of global change.
As mentioned earlier, our approach to RAD in this paper focuses on a pattern of climate change in which ecological processes are held to be stationary between periods of transformative climate dis- Additional ambiguities arise with the interchangeable use of terms like ecosystem change, trajectory, pathway and transformation rate in describing RAD, raising questions about the nature and pattern of ecological responses to management over time.Concerning adaptive management in a context of RAD, clarification is needed as to the role of ecological uncertainty in describing ecosystem processes, specifically how learning occurs over time and how it actually factors into learning-based management in RAD.Such concerns no doubt come with early attempts to merge two different and complex approaches.The potential and value of such a merger nevertheless seems apparent.
where action a is one of a sequence of actions over the time horizon and z represents environmental variation.Demographic stochasticities and randomness in z induce Markovian transition probabilities, which can be portrayed with a Markov decision process P with probabilities, Policy in this expression maps state x to action a by (x) = a and encodes the pattern of ecological management by specifying what actions to take under what circumstances as the system fluctuates over time.
structural uncertainty into decision making can be accomplished by means of a simple averaging of transition probabilities and returns.Thus, model-specific returns r k (a| x) are averaged with the model confidence measures to get and model-specific value functions are averaged to get requirements include a post-transformation process model P x � | x, a and return function r(a| x).With this information and a range of allowable actions, one can determine policies and their corresponding values for possible ecological states.Extending the decision framework to account for structural uncertainty expands the information requirements: Now one needs multiple process models Pk x � | x, a and (possibly) return functions rk (a|x).With this information one can then determine policies and values across the range of ecological states and uncertainties.Note that RAD increases the computational burden by requiring three different policy configurations and assessments, but does not increase model information needs.Marescot et al. (2013) described approaches for computing optimal values and policies for Markov decision processes.Williams (2011b), Chadès et al. (2014) and Williams and Brown (2023) and others have discussed computing issues and approaches in adaptive decision making.6 | R AD AND DOUB LE-LOOP LE ARNING Lynch et al. (2022) portray the use of adaptive management in a RAD context in terms of nested decision loops, where technical learning for considering changes in the elements of adaptive management.Double-loop learning could lead to a change in RAD strategy via the process of management itself.It is ongoing decision making that exposes inadequacies of a particular element of decision making, and it is that exposure (i.e. the learning element of doubleloop learning) that motivates a change in the decision elements.For example, as monitoring data accumulate, it may become clear that none of the models in a model set track the system's actual performance adequately(Runge et al., 2016).More broadly, in the face of climate transformation, the need for a change in models, objectives, potential actions, or other elements of decision making is learned as management progresses, which may then indicate the need for a different RAD strategy.Lynch et al. (2022) distinguish between the periodic revisitation of decision elements as prompted by climate transformation, versus revisitation of the elements in its absence as an ongoing part of double-loop learning.Such a distinction is motivated by the linkages of goals, policy options, and process models in the RAD strategies.However, it is worth noting that at least in concept, both cases represent an interruption of a technical learning cycle, pursuant to possible changes in the architecture of decision making.
ruptions.An alternative to this pattern would have climate varying continuously over an extended time horizon.In such a situation, a different approach to adaptive decision making would be required, involving continuous non-stationarity in the processes themselves.Whether and how a RAD framework could apply is an open question, but in any case the analytic and computational demands would increase substantially.Ecological literature is limited on the subject of decision making for dynamic systems that undergo continuous structural change (e.g.McDonald-Madden et al., 2011;Nicol et al., 2015;Tucker & Runge, 2021).Little or none of this work specifically mentions RAD; however, it does address non-stationarity that may be driven by climate change.Beyond experimentation and the treatment of non-stationarity, a large number of technical and computational issues offer avenues for further research.Crausbay et al. (2022) suggest a research agenda that includes the extent to which ecological transformation necessitates new management policy; the identification of plausible ecological futures; the range of available management alternatives; the expected consequences of each alternative and its uncertainty; and the process of decision making itself.Ongoing research in these and other areas offers promise for the future of RAD implementation.We have focused primarily on technical issues with RAD, especially as they involve the merger of RAD and adaptive management.But the human-dimensions side of both RAD and adaptive management, including stakeholder and rightsholder engagement and governance infrastructure, must not be overlooked.Both paradigms emphasize the importance of embedding ecosystem management in a social-ecological milieu of norms, values and governance arrangements, where social relations frame and influence decisions.Magness et al. (2022) emphasize these and other factors in RAD