Elsevier Science Publishers B.V., Amsterdam - Printed in The Netherlands
A NONLINEAR PROGRAMMING APPROACH TO THE ANALYSIS OF PERTURBED MARINE ECOSYSTEMS UNDER MODEL PARAMETER UNCERTAINTY C. ALLEN A T K I N S O N
System Science Applications, Inc., 121 Via Pasqual, Redondo Beach, CA 90277 (U.S.A.) (Accepted 18 November 1985)
ABSTRACT Atkinson, C.A., 1987. A nonlinear programming approach to the analysis of perturbed marine ecosystems under model parameter uncertainty. Ecol. Modelling, 35: 1-28. A general class of applied marine ecosystem problems is modeled using nonlinear programming (NLP) techniques. These problems involve multiple interacting population groups, one or more of which has been perturbed from some approximately known state via man-induced or natural phenomena. Critical ecosystem response characteristics and optimum management strategies are to be determined. The NLP modeling approach is shown to have substantial advantages for treating complex ecosystems and, in particular, for dealing with the practical issue of parameter uncertainty in the dynamic modeling. The NLP modeling approach is demonstrated by analyzing the collapse of the sardine fishery off the Pacific Coast of North America during the 1930's and 1940's. A priori predictions are shown to bound the measured sardine ecosystem response and provide evidence of overfishing as the reason for the collapse.
INTRODUCTION
In reviewing world fisheries problems, the Food and Agriculture Organization of the United Nations (1978) stressed the need for multispecies modeling approaches. Traditional single-species fish models (e.g., Beverton and Holt, 1957; Ricker, 1958) are usually inadequate since the exploited species interact strongly with predators, prey, and competitors within the ecosystem. Examples of multispecies models are described in papers by Riffenburgh (1960), Saila and Parrish (1972), Steele (1979), and May et al. (1979). The ability to estimate model parameters, which is essential to the practical application of all ecosystem models, is particularly critical in multispecies models. Data is not generally available on the temporal and spatial scale needed to estimate population interaction terms with any degree of accuracy (Goodall, 1972). Unless parameter estimates can be 0304-3800/87/$03.50
tightly bounded, multispecies models will predict a wide spectrum of dynamic behaviour even with relatively few population groups represented (Parrish 1975; Steele 1979). The problem is exasperated as the number of ecosystem components represented in the model increases, since both the potential interaction terms and model degrees-of-freedom increase geometrically. Examples of extremely complex multi-component models are described by Anderson and Ursin (1977) and Laevastu and Favorite (1978). The present nonlinear programming (NLP) approach has been developed to deal effectively with the issue of model parameter uncertainty in applied ecosystem studies. Model parameters are treated as variables and constraint equations are formulated to relate all 6xplicit and implicit information that limits their values. By incorporating these constraint equations directly as elements of the NLP problem statement, the predictive power of the model, i.e., the resolution of the possible response modes, is greatly enhanced. The NLP solution procedures also provide parameter sensitivity results in the form of Lagrange multipliers (Luenberger, 1973). This information can be used to guide field studies to better estimate critical model parameters and to further resolve the ecosystem's dynamic character. System optimization techniques have had limited use in applied ecological problems. The most common application is optimal control theory for questions pertaining to the management of fishery resources (Palm, 1975; Vincent et al., 1975; Agnew, 1979; Ludwig, 1979). Mathematical programming techniques, which include nonlinear as well as linear and dynamic programming, have not been used as extensively (Jaquette, 1972; Jeffers, 1972; Nakamura, 1973; Mendelssohn, 1976). However, Parrish (1975) and the Oak Ridge Systems Ecology Group (1975) have suggested the potential applicability of such procedures. A general class of applied ecosystem problems are formulated as nonlinear programs in the present paper. The NLP approach is then applied to the question of the sardine fishery collapse off the West Coast of North America. The objectives are both to demonstrate the general modeling procedure and to explore the nature of the ecosystem and the reason for the sardine collapse. PROBLEM SCENARIO F O R NLP APPLICATION
The NLP approach will be applied to the problem of predicting the response characteristics of ecosystem populations. While no restrictions need be placed on the nature of the populations or the type of ecosystem, the proposed application is for marine ecosystems and higher trophic-level organisms such as fish and marine mammal species. Predictions are sought for some future perturbations to these populations or to their environment that can be expected to have significant, and typically complex interplay
between ecosystem components. Underlying the scenario is the further practical assumption that knowledge of the ecosystem's dynamic response characteristics is limited. An ecosystem dynamics model can be postulated in some general form, such as coupled differential or difference equations, but data are lacking to adequately resolve the model parameters. The above problem scenario, restated simply as the prediction of marine ecosystem response dynamics under model parameter uncertainty, is the focus for the NLP approach. The dynamics equations will be assumed by hypothesis to adequately represent the first-order features of complex population interactions and processes provided appropriate parameter values are input. The estimated range of model parameter values is assumed to be quite large, spanning both the measurement uncertainty and any inherent uncertainty associated with the ecosystem conceptualization and mathematical structuring. The NLP approach addresses the fundamental dynamic characteristics of the ecosystem and uses deterministic functions to describe them. Stochastic treatments, while offering a potentially better representation of ecosystem characteristics, further add to the estimation problem as both the nature and parameters of its probability distribution must be defined. It is assumed that the stochastic features of the modeled ecosystem can be superimposed on the NLP which determine its fundamental character. Kremer (1983) discusses various stochastic approaches used in ecosystem models. The structure of nonlinear programming is such that extreme (minimum or maximum) response characteristics will be sought. In a practical application, this information might be used to show that worst case results are acceptable or that, in fact, additional data are needed to better estimate parameters and resolve the ecosystem response. In another application, the optimum strategy for maximizing or minimizing a particular population's response might be of interest given that there is some control mechanism available. Further insight into the use of NLP procedures in applied ecosystem problems is provided by the following general development and the subsequent example in which the sardine population collapse off western North America is analyzed. NLP FORMULATION
OF APPLIED ECOSYSTEM PROBLEMS
The perturbed ecosystem problem can be stated as the following general nonlinear program: minimize • (x) subject D( x ) <~0 G(x) = 0 (1) X 0 ~< X > X c¢
I"I
PERTURBED DYNAMICS MODEL
(PARAMETER~k VALUES) J J
~.__....~ ~[
RESPONSE
OPTIMIZATION SEARCH
.l' J i
MODEL PARAMETER Ir I CONSTRAINTS F I
I
OBJECTIVE FUNCTION
Fig. 1. Flow logic showing nonlinear program elements for the perturbed ecosystem problem. Inputs to the NLP problem are represented by the ovals, while computational components are represented by the rectangles. where x is an n-dimensional vector of problem variables; ~ ( x ) is the objective function; D(x) and G(x) are vectors of implicit constraint functions that are distinguished from each other by the type of information encompassed (see below); and x ° and x ~ are vector constants defining explicit bounds on variable space. The problem reduces to a linear program if the objective and constraint functions are all linear. This will not be the case for the practical problems of interest here. The elements of the perturbed ecosystem problem are structured as an N L P in the flow diagram of Fig. 1. A description of these problem elements follows.
System dynamics model Arbitrary functions can be used to represent system response dynamics in the N L P formulation. However, for present purposes, a coupled set of difference equations will be assumed for representing the changes in ecosystem populations over time:
P(t + 1 ) = E[P(t), a, 8]
(2)
where E is a vector function relating numbers or biomass of the population vector, P, at time t + 1 to that at time t. The vector of ecological parame-
ters, a, are distinguished from the vector of perturbation terms, 8, in this representation. Components of the a vector correspond to competitor and predator interaction coefficients; to fecundity, growth and mortality rates; and to other ecological factors that quantify the response of a population to its environment and to other populations. These parameters represent variables in the context of the NLP formulation since there is a large degree of uncertainty associated with their values and, hence, in the resulting ecosystem dynamics. The components of the perturbation vector, 6, represent disturbances to the ecosystem, the effects of which are to be predicted. For example, a perturbation parameter could be defined to represent the catch rate of a new fishery to be introduced to the ecosystem. This parameter could be either estimated or treated as an additional unknown depending on the issue at hand. Before the perturbation, the population dynamics are represented by equation (2) with the vector 6 = 0. After the perturbation, 6 4:0 and the perturbed ecosystem response is determined.
Model parameter constraints Explicit constraints bounding the ecological parameters, a, can generally be stated based on ecological theory, empirical observations, or practical considerations. If the disturbance terms, 8, are treated as variables, similar bounds can be defined. These relationships are expressed as: a ° ~
8