Special Feature Ihpeiin H Oe Aaeesadtesaeof State the and Know, Parameters Nature

Ecology, 84(6), 2003, pp. 1403±1411 ᭧ 2003 by the Ecological Society of America

UNCERTAINTY AND THE MANAGEMENT OF MULTISTATE ECOSYSTEMS: AN APPARENTLY RATIONAL ROUTE TO COLLAPSE

G. D. PETERSON,1 S. R. CARPENTER,1 AND W. A. B ROCK2 1Center for Limnology, 680 N. Park St, University of Wisconsin, Madison, Wisconsin 53706 USA 2Department of Economics, University of Wisconsin, Madison, Wisconsin 53706 USA

Abstract. We use a simple model of ecosystem management to demonstrate that apparently rational management approaches can lead to ecological collapse. Our model of the ecosystem management of lake eutrophication integrates lake dynamics, management decision-making, and learning in a framework that is deliberately simpli®ed to highlight the role of model uncertainty. The simulated lake can switch between alternate eutrophic and oligotrophic states. Managers consider two management models of the lake, one for an oligotrophic lake and the other for a eutrophic lake. As managers observe the lake varying from year to year, they estimate how well each of the two management models is supported by the observed data. Management policies maximize the expected net present value of the lake. Even under optimistic assumptions about environmental variation, learning ability, and management control, conventional decision theory and optimal control approaches fail to stabilize ecological dynamics. Rather, these methods drive ecosystems into cycles of collapse and recovery. We suggest how scientists could help prevent ecosystem management from driving ecosystems toward collapse. Key words: collapse; ecological economics; ecosystem management; eutrophication; model uncertainty; optimal control; resilience. pca Feature Special

INTRODUCTION the management of salmon in the Columbia River for example, ecological management has focused on ex- Ecosystem management has a long and diverse his- pensive and sophisticated attempts to assess the cred- tory. It has sometimes been successful (Berkes and ibility of a wide variety of alternate models of eco- Folke 1998), and sometimes not. For example, histor- system functioning (Marmorek and Peters 2001). The ical societies, such as the Maya (Hodell et al. 2001), aim is to discover which model or set of models appears and modern scienti®cally managed resource extraction to best forecast the future behavior of the ecosystem. systems, such as that for the Northern cod (Walters and Although such efforts are certainly worthwhile, there Maguire 1996), have appeared to succeed at ecosystem may be pressure for science to act as a dispassionate management until an abrupt and surprising collapse. tool to assess the relative validity of a selected set of Researchers have tried to explain these collapses. Some alternative models. Too often, the consideration of un- researchers attribute ecological collapse to ignorance certainty focuses only upon the prediction errors of a of gradual ecological change (Martin 1973, Alroy single model while the credibility of the model itself 2001), rational overexploitation (Clark 1973), or the is not assessed (Clark et al. 2001). The credibility of dif®culties of creating effective institutions to manage a given model depends on the other models with which common pool resources (Hardin 1968). Others focus it is compared and on the data available for comparing on the role of social inequality and greed (Blaikie and models. If the data represent only a subset of the po- Brook®eld 1987). Although few people would argue tential behavior of the ecosystem, then the model com- that resource collapses derive from single causes, we believe that downplaying the uncertainty of manage- parison may be biased, or the appropriate model may ment models may be a neglected cause of ecological not even be discovered, because the behaviors of the collapse. ecosystem relevant to the appropriate model have not A large portion of ecological management involves been observed. If an important model is omitted from decision-making under conditions of uncertainty. Cur- the set of models under consideration, substantial errors rent approaches to ecosystem management often cope can occur in assessing the credibility of models, making with uncertainty by ®tting models to data, or using data predictions, and choosing management actions. Thus, to compare competing models (Hilborn and Mangel model uncertainty has critical implications for ecosys- 1997, Clark et al. 2001). In some ecological con¯icts, tem management, but the assessment of model uncertainty is limited by the range of ecosystem behaviors observed in the data, and by the diversity of models Manuscript received 3 December 2001; revised 3 June 2002; accepted 21 June 2002; ®nal version received 3 July 2002. Cor- created by the analyst. responding Editor: J. S. Clark. For reprints of this Special Fea- In this paper, we show that when the family of mod- ture, see footnote 1, p. 1349. els being considered does not adequately capture key 1403 Special Feature ihpeiin h oe aaeesadtesaeof state the and know, parameters nature. to model the managers precision, the with allowing model by on com- focus We misspeci®cation two data. of observed given probabilities models posterior peting the on based pol- chooses icies management adaptive 1986). passive (Walters case, this possible In not man- are adaptive experiments active agement when the practice be management to thought best poli- is select management adaptive to Passive management cies. adaptive passive process management uses a that model We poorly. function will approaches management sophisticated even dynamics, 1404 uentins u giutrlfriiesaeachief a are 1998 al. fertilizers et (Carpenter agricultural source but pro- activities lake nutrients, human a duce of to types (P) Many 1977). phosphorus (Schindler primarily nutrients, of input pro- those 1999). Carpenter than and (Wilson valuable lakes eutrophic more waterfowl, by duced are or ®sh that as recreation, such or resources use, industrial irrigation, consumption, or human for water as ser- such ecosystem vices, produce lakes usu- oligotrophic People that be algae. ®nd also toxic ally of can blooms They produce water. and anoxic murky high and inputs, clear production, nutrient relatively high plant and have lakes production, mod- Eutrophic plant water. to of low inputs, levels nutrient erate low lakes state by Oligotrophic eutrophic characterized 1998). or Smith are 1993, oligotrophic al. an et either (Scheffer in Lakes eutrophication. exist lake of can problem the use we ment cycles to recovery. lead and alternatives collapse of model of set of managed limited dynamics a a the from of that choice dynamics show will the We in ecosystem. with role uncertainty feedbacks the model on local focuses of model of our interaction Rather, drivers. the external as complexi- such ecological ties, and interactions, institutional such sim- complexities, as social analytical omits useful model Our to pli®cations. access lake providing describe to dynamics, equations linear use ecosys- We interaction for dynamics. tem implications its its uncertainty, and to decision-making, model order with on in attention interventions, management focus of and impact managers, the to available information the ecosystem, un- model of role the isolates certainty. an that created management have ecosystem we of model Therefore, uncer- socioecological way. of simple exceptionally clear role the a each isolate in in to tainty dif®cult variation it of and make sources examples situation interacting the multiple of situa- the complexity different in the management However, applied tions. been ecosystem have of that diversity approaches the into insight urpiaini sal asdb h excessive the by caused usually is Eutrophication manage- ecosystem of dynamics the investigate To the of dynamics the simpli®es greatly model Our provide models simulation detailed and studies Case R ATIONALE b .Popou nfertilizer in Phosphorus ). .D EESNE AL. ET PETERSON D. G. idmxn,adteoye otn fde water deep of 1998 content al. sediments, oxygen in et P the (Carpenter of and accumulation mixing, the wind to related behavior is threshold that exhibits Recycling state. eutrophic ser- ecosystem on impose a vices. activities these is costs activities there the polluting and from services, bene®ts ecosystem the de- between of trade-off while value agriculture the of creasing value the Because fertilizer. increases Agricultural of fertilizer use 2001). the al. upon depends et production caus- (Bennett and bodies eutrophication water into es erodes that soil in up builds stebsln aua odn notelake, the into loading P natural baseline the is where yln.TePdnmc nalk r oee ythe by modeled re- are P lake and a equation removal, in difference lake P dynamics P, a P within of The P cycling. loading of the dynamics upon the depends that assume We P. of sections. following the man- in presented ecosystem are our model agement of learning, components lake, management the and of their details about The beliefs models. their management update sci- and year, data Each collect services. lake entists between ecosystem trade-off lake a and of pollution maxi- value upon expected based the mizing to is response management lake's Lake the models management. of these of expectation each their in belief determines of degree Their of lake. models states. the management oligotrophic competing two or have eutrophic Managers produce dy- P can Lake namics 1). (Fig. man- process a and decision-making process, agement learning a dynamics, lake of model tesd o Shfe ta.19,Cretre al. et Carpenter 1993, al. et (Scheffer 1998 not do while addition, others nutrient of sediments. cessation oligotrophic the lake an following to in state return P quickly lakes vege- of eutrophic submerged concentration Some structure, the web and such food tation, lake depth, nu- the area, of of properties reduction as upon a depends following inputs state trient return oligotrophic can al. lake an eutrophic et a to which (Soranno with inputs ease The annual lake 1997). from exceed recycled P may of sediments amount the 1987). P al. lakes, the et high eutrophic in (Schindler In P of recycling enough P years of initiate to accumulation few sediment the a to lead only can that levels shown have lations hshrsrccigwti aecnmiti a maintain can lake a within recycling Phosphorus diinlatrpgncPlaigit h aea time at lake the into loading P anthropogenic additional X erpeettesaeo aeb h concentration the by lake a of state the represent We a of consists management ecosystem of model Our t ϩ S 1 a t ). ϳ ϭ X t N Ά ersnstecnetaino ttime at P of concentration the represents BX BX (0, ttttttt t ␴ ϩ ϩ 2 AM (1) ) r b ϩ ϩ aedynamics Lake b INIMAL l a ϩ ϩ .Eprmna aemanipu- lake Experimental ). l S ϩ M S ODEL clg,Vl 4 o 6 No. 84, Vol. Ecology, if if X X Ն Ͻ crit crit X X l t sthe is ,b t, June 2003 ECOLOGICAL INFERENCE AND FORECASTS 1405

FIG. 1. Schematic diagram of the lake ecosystem management model.

t, and r is the amount of P recycled and maintained in Model 1: Xtϩ1 ϭ BXtttϩ b ϩ l ϩ S (3)

the lake. B represents the proportion of P that is retained Feature Special Model 2: Xtϩ1 ϭ BXtttϩ r ϩ b ϩ l ϩ S . (4) in the lake after one year. Xcrit is the critical P concentration at which P recycling begins. St is a random Model 1 corresponds to the belief that P loading drives variation in P loading at time t drawn from a normal lake P dynamics. Model 2 corresponds to the belief distribution with mean 0 and standard deviation ␴. that P recycling and P loading drive lake P dynamics. The proportion of P retained by the lake (B) must For a given P loading, the eutrophic model (Model 2) be between 0 and 1. Both the base (b) and the anthro- predicts a higher observed concentration of P in the pogenic loading (lt) must be equal to or greater than lake than the oligotrophic model. Each of these models zero. Furthermore, for cultural eutrophication to occur accurately describes lake dynamics under a range of the value of Xcrit must exceed b. conditions (Model 1 is correct when X Ͻ Xcrit, while Over a wide range of parameters and P loadings, Model 2 is correct when X Ն Xcrit). However, managers lake dynamics can exhibit two alternate equilibrium are unaware that recycling is a function of P concen- values of X. The following equilibrium points (X1* , tration.

X2*) are found by letting Xtϩ1 ϭ Xt: To focus our analysis on model uncertainty, we allow managers to know correctly and precisely the rate of b ϩ l nutrient retention (B), natural P load (b), the amount X 11crit* ϭ if X* Ͻ X 1 Ϫ B of P recycling (r), and the probability distribution of r ϩ b ϩ l shocks (S). This assumption provides the managers X* ϭ if X* Ͼ X (2) with more information than they would likely have in 22crit1 Ϫ B reality. This simpli®cation places all management un-

If X1* Ͻ Xcrit an oligotrophic equilibrium exists, and if certainty on the choice between the models in Eqs. 3

X2* Ͼ Xcrit a eutrophic equilibrium also exists. Because and 4. Managers attempt to maximize the utility of the the equilibrium value of X depends upon the P loading, lake by controlling P loading to the lake. Using passive changes in P loading can cause equilibrium point to adaptive management, the managers attempt to im- appear or disappear (Fig. 2). This response to changes prove their management by updating their relative be- in P loading can cause surprising changes in lake belief in these competing models. Each simulated year havior. managers test Model 1 and Model 2 against the lake's observed behavior. Because they are limited to two Learning models, they believe Model 1 with probability p1 and Model 2 with probability p2, where p1 ϩ p2 ϭ 1. This Managers have two alternative models of the lake. type of management action corresponds to management Managers have some degree of belief that the lake is that compares existing models of nature rather than either oligotrophic (Eq. 3) or eutrophic (Eq. 4), but creating new models. management is uncertain about which of these man- The likelihood of a model given a single observation agement models is more accurate: at time t can be calculated as follows: Special Feature tec iese,Es n r eacltdto each recalculated observations. to past are assign their 6 managers upon and based that model 5 probability Eqs. the step, update time each At where 1406 aaesvletelk sasuc feoytmser- ecosystem of source a as lake the value managers o h lgtohcadetohcmngmn models. management eutrophic and oligotrophic the for point) lower (the oligotrophic valuesÐone equilibrium stable two has point). lake higher same (the the loading eutrophic P one higher and a at (B), In equilibrium. rbblt fec oe,uigBys ue(Walters rule 1994): Bayes' Ludwig posterior using and the model, update each to of used probability be can kernels likelihood ttime at 1 F F h aaeetapc ftemdlasmsthat assumes model the of aspect management The IG IG .Epce tlt fdfeetaonso loading of amounts different of utility Expected 3. . oligotrophic stable single a in results loading P low (A), example In model. dynamics P the of illustration An 2. . X 1, t t stevleo ttime at X of value the is Ϫ ,and 1, L L p p 2, 1, 2, 1, t t t t ϭ ϭ ϭ ϭ X exp exp pL pL Management 2, 1, 1, t t t Ϫ Ϫ stesm o oe .These 2. Model for same the is [] [] 11, 11, Ϫ Ϫ pL pL ( ( 2, 1, t t X X t t Ϫ Ϫ ϩ ϩ 2, 1, 12, 11, t t Ϫ Ϫ pL pL 2, 2, ␴ ␴ 2 2 t t t t t X X Ϫ Ϫ obs, obs, rdce yModel by predicted 12, 12, t t ) ) t t 2 2 (6) . .D EESNE AL. ET PETERSON D. G. (5) ytelk elnsa udai ucino aeP lake of function quadratic 8): a (Eq. as concentrations declines produced services lake pollution of the value a the by that as and into 7), loaded lake (Eq. P lake of the the pollution. amount the of P with linearly of value increases sink sink the a that as assume and We water) clean (e.g., vices h pia loading, optimal The where as written be t time at utility net Therefore, flk epne xetduiiyis utility expected response, lake 3). of (Fig. model management each for different load- is optimal ing forecast the Consequently, to dynamics. used P is lake's model the management the upon pends Muellbauer and (Deaton welfare 1980). mea- aggregate valid a of producing of sure problem as dif®cult stakeholders the different as address well of would values analysis heterogeneous economic in the stake Rigorous a have lake. in who the differences people different ignores the the among ecosystem for values an function utility of simple management this of assumption The etduiiisdsone ytedson ae1 ex- rate future discount of the summation by the discounted is utilities which pected lake, the of value the maximize to seek managers The ie fE.9cnb iie by divided be can 9 Eq. both alternatives, of ranking sides of means a is utility Because, hsuiiyfnto sue o lutaiepurposes. illustrative for used is function utility This eas h aaeshv w optn models competing two have managers the Because de- rate loading P given a of utility expected The E U [ , c,t b,t t Јϭ U k t ] E ϭ ϭ U [ V k pE 12 U U tt U 1, ] / l ϩ kU t b,t ,i h iesre fPloadings P of series time the is *, c,t ttt ϭ␦ ϭ U ϭ [ ϭϪ U ͸ ϭ s ϱ kl 1, 0 kl t 1 ϭ ] kX Ϫ t tt s 2 ϩ E 1 kl ϭ X t [ 2 t 2, pE U 2 clg,Vl 4 o 6 No. 84, Vol. Ecology, t (8) . k U Ϫ ϩ 2 Ј xetdntpresent net expected s 2 yielding / .(12) ]. [ kX 2 k U 2, t t . 2 ]. . Ϫ␦ (11) (10) (9) (7) : June 2003 ECOLOGICAL INFERENCE AND FORECASTS 1407 that maximizes the expected net present value of the lake (see Appendix A): k(1 Ϫ B␦) l* ϭϪBX Ϫ b Ϫ rp . (13) 2␦ t 2,t Substituting optimal loading (Eq. 13) into Eqs. 3 and

4 with p2 ϭ 0 and p2 ϭ 1, respectively, yields the expected optimal equilibrium value of X for management model 1 and model 2: k(1 Ϫ B␦) X* ϭ . (14) 2␦

However if Eq. 13 is substituted into Eq. 1 with p2

ϭ 0 and p2 ϭ 1, there are four possible alternate equilibrium values. Two of these equilibria are the expected optimal value of X for each model (Eq. 14), and two are unexpected equilibria values of X. The highest equilibrium value occurs when the lake recycles P, but the managers do not believe that it does (i.e., strong belief in model 1):

X1*r ϭ X* ϩ r. (15) The lowest equilibrium value occurs when the lake

does not recycle P, but managers believe that it prob- Feature Special ably does (i.e., strong belief in model 2):

X2*r ϭ X* Ϫ r.

Because P recycling occurs if X Ͼ Xcrit,X1*r can only exist if there is P recycling; that is if X1*r Ͼ Xcrit. Sim- ilarly, X1*r Ͼ Xcrit Ͼ X*2r can only exist if there is no P recycling, which is true when X2*r Ͻ Xcrit. Consequently, there are two alternate stable states if X1*r Ͼ Xcrit Ͼ

X2*r. Implementation Parameters of the model were chosen so that alternate stable states existed and natural P loading was less than Xcrit (i.e., plausible P loading rates). Furthermore, for ease of presentation, the model was parameterized so that maximum values of X and net utility are less FIG. 4. Expected and actual P equilibria values when lake than 1. At a high discount rate (e.g., 0.5) the future is managed (A) as if model 1 is true, and (B) as if model 2 ␦ϭ is true. is not highly valued and managers will sacri®ce the future state of the lake for the bene®ts of disposing of P in the present. Such high discount rates are both uninteresting and atypical. To avoid illustrating the ef- expected equilibrium does not exist, and the system fects of discounting we choose a relatively low discount will be attracted towards an equilibrium that is higher rate. Parameters chosen for an example model are b ϭ than expected (Fig. 4A). Similarly, if the managers be- 2 0.02, B ϭ 0.1, r ϭ 0.2, Xcrit ϭ 0.7, ␴ ϭ 0.02, k ϭ 1.5 lieve Model 2, a lower optimal loading will be chosen, and ␦ϭ0.99. The model was implemented in Microsoft producing a stable equilibrium. However, in this case, Excel (Microsoft, Redmond, Washington, USA). there is also another lower stable equilibrium (the oligotrophic state) to which a shock can move the eco- RESULTS system (Fig. 4B). The optimal loading chosen by a manager depends While changes in belief and loading vary smoothly, upon the degree of belief in each management model. changes in utility and lake P are occasionally abrupt However, because each management model implies a providing both positive and negative surprises. This single equilibrium, lake managers can be surprised by movement between one equilibrium and another results the switches between alternate equilibria that occur in in an aperiodic cycling in belief in different models, the lake. When the lake is managed using Model 1, the lake P concentration, P loading and net utility (Fig. 5). Special Feature 1408 hnerqie osittelk rmoesaet an- of to state amount Because one the 1973). from lake (Holling as the other shift de®ned to be required change can eutrophic Resilience and oligotrophic states. the between transitions for C hne nuiiypoue ytelk ayarpl,wietelk aae' xetduiiy( time. utility over expected manager's cyclically lake vary the models while management abruptly, competing vary the lake in the belief by in produced Changes utility (B) in dramatically. Changes (C) more vary the within ec fteetohcsaeutltelk butyre- again. begins abruptly cycle the lake and the state oligotrophic until an to managers state verts eutrophic resil- state, the the eutrophic decreases of gradually ience a which to loading, P shift reduce rapidly In the increases. drops to state utility eutrophic response the state, of the resilience eutrophic When the state. a while eutrophic to shock a a shifts to until shift lake resilience, to low lake and the utility causes high a per- at then sists loading and P decreases, as gradually 6A, resilience Fig. increases, In simulation. same the of namics occur. states between eutrophic re- transitions and alter abrupt oligotrophic models why reveals management helps in silience belief and loading P ftelk,bcuete r nwr htPrecycling P that unaware when are resilience they occurs the because of lake, the unaware of are Managers resilience. of etsaeo h lake cur- the the of between state distance rent the states, eutrophic and phic F eiinecnb sda nidxo h potential the of index an as used be can Resilience i.6sostodfeetprpcie ntedy- the on perspectives different two shows 6 Fig. IG .Asnl elzto ftemdltm yais A odn otelk aisgaulywielvl fP of levels while gradually varies lake the to loading P (A) dynamics. time model the of realization single A 5. . X t exceeds X t and X crit X crit xmnn o hfsin shifts how Examining . a eue sameasure a as used be can X crit eaae oligotro- separates .D EESNE AL. ET PETERSON D. G. h lgtohcsae eifi h urpi model eutrophic the rapidly. entering declines in 2) once belief (Model However, state, lake. oligotrophic the the of state be- current than temporarily rather the observations increases past upon 2 depends in- belief Model resilience cause its in lake state Belief oligotrophic the creases. a to When stable back no equilibrium. switches be- is there management prediction because is eutrophic increases, This accurate. further less comes lake in belief eutrophic as However, pre- the accurate. model more and declines become model dictions oligotrophic the and in be- prediction lief predictions, poor in several error After becomes surprising increases. resilience lake a is oligotrophic resil- there the but eutrophic de- When accurate, oligotrophic declines. is more ience lake become the that predictions belief dynamics clines, As future and lake. the resilience the forecasts of to management compared (that well are 2 Model how eutrophic) in is belief lake in changes the 6B, Fig. competing In to models. assign managers probabilities what upon ainl cet® aaeetlaigt ylsof cycles to leading management scienti®c rational, h odn htmngr pl oalk depends lake a to apply managers that loading P The u oe sasml,tasaetprbeabout parable transparent simple, a is model Our D ISCUSSION clg,Vl 4 o 6 No. 84, Vol. Ecology, E [ U )vre slowly. varies ]) June 2003 ECOLOGICAL INFERENCE AND FORECASTS 1409

FIG. 6. An example of the dynamics of the model. (A) High P loading increases the utility of the lake but decreases lake resilience, until the lake ¯ips to a eutrophic state. Utility of the lake rapidly declines. As the loading of the eutrophic lake decreases, the resilience of that state gradually decreases, while utility remains low. Eventually the lake once more becomes oligotrophic, and the cycle begins again. (B) Belief in a model increases as prediction error (surprise) declines. However, as certainty increases, the lake management causes lake resilience to decrease, increasing the chance that a lake will change state, producing a management surprise. When such a surprise occurs, the degree of belief in a model rapidly declines, causing belief in the other model to increase.

collapse and renewal for a managed ecosystem. In our Our model is an abstraction of the general situation model, rational scienti®c management is represented in which a manager must choose input rates to an eco- Feature Special by a passive adaptive management process in which system of a substance with unknown persistence, re- no effort is made to enlarge the model space. Man- cycling, or transformations. Many ecological processes agement actions are rational in the sense that data and exhibit time-lagged dynamics, delays in response to competing models are used to make decisions that max- intervention, or irreversible change. Examples include imize expected utility. Cycles arise from the dynamics carbon dioxide enrichment of the atmosphere, persis- of the beliefs of managers as supported by actual lake tent organic pollutants, and heavy metal pollution. behavior. As the weight of evidence builds for a par- Thus, in a general way, we expect that our ®ndings ticular management model, policies become ®xed and about uncertainty, learning, and ecosystem dynamics belief in the model fossilizes as no contradictory evi- would arise in more detailed and realistic models for dence is gathered. Eventually, a surprise provides new managing inputs of a wide variety of substances to a evidence that supports the alternative model, increasing wide variety of ecosystems. uncertainty, and resulting in changes in policy. Evi- The route to collapse shown here provides an alter- dence accumulates in support of the alternative model, native to other explanations of ecological collapse in- leading to the ®xation of a different policy, until the cluding inability to detect gradual change (Martin system produces another surprise. Because manage- 1973, Alroy 2001), the tragedy of the commons (Hardin ment actions depend upon what managers know at a 1968), command and control management (Holling and given time and the models that are available to them, Meffe 1996), social inequality (Blaikie and Brook®eld even if a system has exhibited cycles in the past, man- 1987), or rational overexploitation (Clark 1973). The agers will attribute these cycles to some low probability mechanism demonstrated here is one plausible cause events. Breaking away from these cycles requires cre- of the cycles seen in more complex models of ecosys- ating a new management model of the system that ad- tem management (Carpenter et al. 1998b, Janssen and equately captures the dynamics of the system. Carpenter 1999, Janssen 2001). In respect to its information dynamics, our model is The surprises that cause policy shifts in this model related to work in economics that explores learning and are analogous to those seen in many case studies of its limits. Rational behavior can create chaos in markets ecosystem management (Gunderson et al. 1995). For (Brock and Hommes 1997) and cyclic behavior in econ- example, cultural eutrophication of lakes in North omies (Nyarko 1991). Economists have also explored America, initially recognized following World War II, the adaptive management of simpler systems, in which was a surprise (Hasler 1947), as was the emergence of the system being managed neither evolves dynamically dead zones and harmful algal blooms as serious en- (Easley and Kiefer 1988), nor exhibits alternative sta- vironmental problems for coastal oceans (Downing et ble states (Beck and Wieland 2002). Future research al. 1999). In these cases, long periods of stasis rein- could extend our work to examine active adaptive eco- forced the notion that ``dilution was the solution to system management. pollution,'' however this belief was eventually made Special Feature oeta fteeoytm edn oteunderesti- the scales dynamic at to Experimentation full uncertainty. leading model the of ecosystem, mation reveal the not of do ®x- potential that from policies follows pre- on beliefs model ation the of In fossilization data. here, the available models sented of out test range point to the experiments can beyond informative Scientists safe, (3) of 2003 value al. et 1996; Heijden Peterson der van 1985, (Wack example analysis for model scenario unknown, when in even or out high carried quite is be uncertainty can the policies of of implica- explorations robustness Such different behavior. ecosystem very be for with tions can models policies under candidate examined of are collapse consequences of The costs the high. when example weight for non-negligible the decision, carry cases in how models be for will novel There such models future. where the novel in imagine change might Scien- to system (2) help ignored. can be cannot tists it uncertainty, infor- of objective discomforting aspect this putatively Despite ecosystem. to the about be- unrelated mation and are the attitudes of that on observation liefs depends prior on therefore part It under in ecosystem). construct depends models mental it of a if is set (even models the of that set of out This property point consideration. can a Scientists is (1) ways. uncertainty three management least to ecosystem at in of contribute worldview can manage- the scientists ecosystem broadening speci®cally, during More tested ment. and can that models stakeholders developed and of information be of set set diverse the increase more can a include decision-making to the process Opening consideration. mod- of under range els the expanding involve all They collapse. regime. new ecosys- a when, to and shift if, dynamics surprise tem for stage of the solidi®cation sets best apparently model The this around model. choices policy and prevailing evidence the biased in This belief increases range. behaviors actual potential ecosystem's its the of of constrained sample portion is small behavior a ecosystem within and point apparent an optimal familiar at ®xed that become ultimately within Policies regime. models dynamics ecosystem of explain support Thus, that in policymakers. accumulates and evidence be- scientists scienti®c that to a regime familiar around a comes within ¯uctuates remains behavior patternÐit ecosystem repeated time, the of ways. surprising in behave can strikingly social±ecologi- systems two large-scale cal how provide hole of AIDS examples ozone of different the and of Antarctica emergence over The 2001). ®sh- limited (Bundy severely ing of decade have a populations despite cod recovered mod- not Unexpectedly, inadequate ®shery. on the part, 1982) of in els Hart least at and blamed, been (Pitcher have ®sheries achoveta and 1996) Maguire Peruvian and re- (Walters cod Northern in in lapses col- changes surprising Similarly, massive 1992). (Likens and waters ceiving surprising by untenable 1410 hr r aywy oaodtertoa ot to route rational the avoid to ways many are There Most events. infrequent are surprises de®nition, By .D EESNE AL. ET PETERSON D. G. in o csse aaeeti e rafrfuture for area research. key a de- is institutional management effective ecosystem for of signs study the Thus 2001). to ters Pe- barriers and greater Marmorek 1997, far (Walters implementation are their schemes such complexities non- organizing political of and remain economic challenges social, management the technical adaptive trivial, while apply that to suggest attempted have eco- who those simulated of experiences of the However, management. success system the adaptive in¯uence active and issues management diversity these model assess how to plan examining we by nor work, easy future neither In is free. cost management adaptive in or making re- states. could ecosystem that desirable valuable practices inforce be innovative may with experiment it to possible, are unfavorable Never- dynamics and caution. ecosystem surprising where with situations approached in be theless, human must support being that large- well course, ecosystems Of on trap. the experiments out scale way ecosys- one for is models behavior tem alternative testing for appropriate apne,S . .Blre,R .Ltrp .A tw T. Stow, A. C. Lathrop, C. R. Bolgrien, D. R., S. Carpenter, ud,A 01 ihn neoytm:teitrlyof interplay the ecosystems: on Fishing ran- to 2001. route A. rational A Bundy, 1997. Hommes. C. and and W., degradation Brock, Land 1987. Brook®eld. H. and P., Blaikie, ecological Linking 1998. editors. Folke, 2001. C. Caraco. and F., F. Berkes, N. and Carpenter, R. S. M., E. a Bennett, in control and Learning 2002. Wieland. V. and G., Beck, end- the of simulation overkill multispecies A 2001. J. Alroy, omnsfo .Cx .Bnet .Cak n w anon- two and by Clark, J. improved Bennett, was referees. E. ymous It Cox, Foundation, S. Foundation. McDonnell from Science S. comments National James the the from and grant a through h osdrto fdvremdl ndecision- in models diverse of consideration The apne,S . .F aao .L orl,R .Howarth, W. R. Correll, L. D. Caraco, F. N. R., S. Carpenter, etn . n .Melae.18.Eooisadcon- and Economics 1980. Muellbauer. emerging J. an and forecasts: A., Ecological Deaton, 2001. al. et S., J. Clark, Science over-exploitation. of economics The 1973. C. Clark, ed n .A isn 1998 Wilson. A. M. and Reed, 1167. inJunlo ihre n qai Sciences Aquatic and Fisheries Cana- of Newfoundland±Labrador. Journal in dian predation and ®shing Econometrica domness. USA. York, New York, New Routledge, society. Cam- Press, University Cambridge UK. bridge, systems. social and BioScience perspective. global eutrophication: and a phosphorus erodable on impact Human Control and Dy- Economic namics of Journal environment. economic changing Science extinction. mass 1893±1896. megafaunal Pleistocene hswr a upre yteRslec Alliance, Resilience the by supported was work This .N hrly n .H mt.1998 Smith. H. V. and Sharpley, N. A. Ecology of Aus- pollution. Journal nonpoint tralian by eutrophication lake of analysis uino ufc aeswt hshrsadntoe.Eco- Applications nitrogen. and logical phosphorus with waters surface of lution ue eair abig nvriyPes odn UK. London, Press, University Cambridge behavior. sumer Science imperative. 181 :630±634. A L CKNOWLEDGMENTS 26 ITERATURE 293 8 :1359±1377. :559±568. 65 :657±660. :1059±1095. 23 :68±79. a clgcladeconomic and Ecological . C 51 ITED :227±234. clg,Vl 4 o 6 No. 84, Vol. Ecology, b opitpol- Nonpoint . 58 :1153± 292 : June 2003 ECOLOGICAL INFERENCE AND FORECASTS 1411

Downing, J. A., J. L. Baker, R. J. Diaz, T. Prato, N. N. Ra- Peterson, G. D. G. S. Cumming, and S. R. Carpenter. 2003. balais, and R. J. Zimmerman. 1999. Gulf of Mexico hyp- Scenario planning: a tool for conservation in an uncertain oxia: land and sea interactions. Council for Agricultural world. Conservation Biology 17(2):1±10. Science and Technology, Ames, Iowa, USA. Pitcher, T., and P. J. B. Hart. 1982. Fisheries ecology. Chap- Easley, D., and N. M. Kiefer. 1988. Controlling a stochastic man and Hall, London, UK. process with unknown parameters. Econometrica 56:1045± Scheffer, M., S. H. Hosper, M.-L. Meijer, B. Moss, and E. 1064. Jeppesen. 1993. Alternative equilibria in shallow lakes. Gunderson, L., C. Holling, and S. Light, editors. 1995. Bar- Trends in Ecology and Evolution 8:275±279. riers and bridges to the renewal of ecosystems and insti- Schindler, D. W. 1977. Evolution of phosphorus limitation tutions. Columbia University Press, New York, New York, in lakes. Science 195:260±262. USA. Schindler, D. W., R. H. Hesslein, and M. A. Turner. 1987. Hardin, G. 1968. The tragedy of the commons. Science 162: Exchange of nutrients between sediments and water after 1243±1247. 15 years of experimental eutrophication. Canadian Journal Hasler, A. D. 1947. Eutrophication of lakes by domestic of Fisheries and Aquatic Sciences 44:26±33. drainage. Ecology 28:383±395. Smith, V. H. 1998. Cultural eutrophication of inland, estu- Hilborn, R., and M. Mangel. 1997. The ecological detective: arine, and coastal waters. Pages 7±49 in M. L. Pace and P. confronting models with data. Princeton University Press, M. Groffman, editors. Successes, limitations, and frontiers Princeton, New Jersey, USA. in ecosystem science. Springer-Verlag, New York, New Hodell, D. A., M. Brenner, J. H. Curtis, and T. Guilderson. York, USA. 2001. Solar forcing of drought frequency in the Maya low- Soranno, P. A., S. R. Carpenter, and R. C. Lathrop. 1997. lands. Science 292:1367±1370. Internal phosphorus loading in Lake Mendota: response to Holling, C. S. 1973. Resilience and stability of ecological external loads and weather. Canadian Journal of Fisheries systems. Annual Revue of Ecology and Systematics 4:1± and Aquatic Sciences 54:1883±1893. 23. van der Heijden, K. 1996. Scenarios: the art of strategic Holling, C. S., and G. K. Meffe. 1996. Command and control conversation. John Wiley and Sons, New York, New York, and the pathology of natural resource management. Con- servation Biology 10:328±337. USA. Janssen, M. A. 2001. An exploratory integrated model to Wack, P. 1985. ScenariosÐshooting the rapids. Harvard assess management of lake eutrophication. Ecological Business Review 63:139±150. Walters, C. J. 1986. Adaptive management of renewable re-

Modelling 140:111±124. Feature Special Janssen, M. A., and S. R. Carpenter. 1999. Managing the sources. McGraw Hill, New York, New York, USA. resilience of lakes: a multi-agent modeling approach. Con- Walters, C. 1997. Challenges in adaptive management of ri- servation Ecology 3:15. parian and coastal ecosystems. Conservation Ecology 1(1). Likens, G. E. 1992. The ecosystem approach: its use and (Online, URL: ͗http://www.consecol.org/vol1/iss2/art1͘.) abuse. Ecology Institute, Oldendorf-Luhe, Germany. Walters, C., and D. Ludwig. 1994. Calculation of Bayes pos- Marmorek, D., and C. Peters. 2001. Finding a PATH towards terior probability distributions for key population param- scienti®c collaboration: insights from the Columbia River eters. Canadian Journal of Fisheries and Aquatic Sciences Basin. Conservation Ecology 5(2) [Available from the In- 51:713±722. ternet] URL: ͗http://www.consecol.org/vol5/iss2/art8͘. Walters, C., and J.-J. Maguire. 1996. Lessons for stock as- Martin, P. S. 1973. The discovery of America. Science 179: sessment from the northern cod collapse. Reviews in Fish 969±974. Biology and Fisheries 6:125±137. Nyarko, Y. 1991. Learning in mis-speci®ed models and the Wilson, M. A., and S. R. Carpenter. 1999. Economic valu- possibility of cycles. Journal of Economic Theory 55:416± ation of freshwater ecosystem services in the United States: 427. 1971±1997. Ecological Applications 9:772±783.

APPENDIX A description of the calculation of optimal loading is available in ESA's Electronic Data Archive: Ecological Archives E084-034-A1.