Methods Integrated Water Resources Management (IWRM) is inherently multi-objective. Traditional approaches to such problems such as Cost-Benefit Analysis (CBA) relied on the aggregation of objectives into a single, parameterized goal function (e.g. (Lahmeyer 1998)). This was achieved by valuation. Valuation establishes a common metric between changes and therefore allows straight-forward comparison of the individual components of such changes. The weighted sum of values finally is the measure based upon which allocation decisions will be made by the decision makers (Heal 2004). CBA are based most often on financial valuation, i.e. the quantification of costs and benefits directly related to an allocation decision. Traditional CBA did neither pay attention to external effects nor take into account non-use values of natural resources. Clearly, this is a critical issue since not valuing some services is equivalent to assigning those services a zero value. This shortcoming was addressed by the development of the Total Economic Valuation Framework (TEV). TEV is the process of quantifying the economic value of particular changes in the levels of a goods and/or services in a holistic way (see Figure 1). Thus, it includes every accountable item as well as externalities that affect the wider economy (Cropper and Oates 1992; Turner 1993). Out of several reasons, values to these changes are difficult to assess. First and especially in a context of subsistence and developing economies, many of these goods and services are not traded in markets and a large part of the value may stem from nonuse value. Second, catchment approaches involve large spatial scales that cut across political boundaries. However, economic valuation techniques are difficult to carry out on such scales when an exhaustive list of services provided by natural systems should be incorporated and regional heterogeneities in the political/institutional realm exist (National Research Council (United States) Committee on Valuing Ground Water 1997; Acharya and Barbier 2002; Heal 2004). o Pr du Natural Resources Structure & Function io ct n Natural Resources Goods & Services Allocation Decision Human Actions (Private / Public) Decisionmakers Values Use Values Non-Use Values Figure 1: Valuating goods and services from natural resources (Heal 2004). In the context of this study, the notion natural resources refers to surface water, groundwater and related resources. Values consist of use and non-use values to extent the narrow utilitarian approach. Due to the various issues related to valuation and CBA, we suggest a novel approach to IWRM that does not rely on the aggregation of objectives (see Figure 2). Rather, we propose to combine methods from the fields of Multi-Criteria Decision Making Models (MCDM) and Operations Research (OR) as well as the field of Artificial Intelligence (AI) respectively to provide a systematic means for comparing tradeoffs and selecting policy alternatives that best satisfy the decision maker’s objectives. MCDM are methods that have proved successful in both, the linking of natural sciences / engineering and social sciences research as well as bridging the gap between scientific knowledge and the identification of commonly agreed upon compromise solutions that lead to allocation decisions. Although variations exist, the basic steps of MCDM involve: 1) Definition of the decision problem and identification of stakeholders; 2) Identification of objectives and attributes; 3) Identification and pre-selection of alternatives to allocation; 4) Prediction of outcomes; 5) Quantification of preferences of stakeholders for outcomes; 6) Ranking of alternatives; 7) Assessment of results; 8) Implementation and Monitoring (Heathcote 1998; Reichert, Borsuk et al. 2004). The use of MCDM in natural resources management has generated a substantial body of literature in recent years (Steiguer, Liberti et al. 2003). Traditional MCDM suffer from major shortcomings that are particularly relevant in the context developing countries. First, the methods rely on early involvement of decision makers and stakeholders. They thus work best in well structured institutional and political setups with an established culture of negotiation. Furthermore, non-cooperative strategic environments as well as latent conflict situations are not conducive to solution finding. Second, a well defined set of allocation alternatives is under consideration in the process of decision making. However, this set is not exhaustive and therefore does not cover the whole set of feasible policies that can be of potential interest. Due to short-sightedness, the presence of vested interests as well as the inherent complexity of the management policies under consideration, non-obvious intelligent management policies might very well never have been considered within step 3) of a MCDM. In other words, policies might be under consideration for which there exist better ones in a Pareto-sense. Both of the above issues can be dealt with by state of the art multi-objective optimization methods. Multi-objective optimization denotes the task to find a set Q* of Pareto-optimal allocation policies p* Î Q in the sense that an improvement in one component of the objective vector f ( p* ) can be achieved only at the expense of another component. More precisely, Q * := {p* Î Q $/ pˆ Î Q : f ( pˆ ) f ( p*)} Traditional approaches of solving such problems relied on the aggregation of the objectives into a single, parameterized goal function. The optimizer would then systematically vary the parameters to achieve a set of solutions that approximate the Pareto–optimal set. These approaches however are sensitive to the shape of the Pareto– optimal front and generally do not exploit synergies between solutions that could help to reduce computational time of the search (Deb 2001). This is a major drawback in cases where the search space is conceivably complex as in the case of IWRM. Multi-objective evolutionary algorithms (MOEA) on the contrary do not rely on the availability of a common metric which makes individual goods, services and with that objectives commensurable. These methods perform well in difficult search spaces (Zitzler and Thiele 1999; Zitzler, Thiele et al. 2003; Siegfried 2004). The genetic algorithm models the evolution of a population of policies p through successive generations using three probabilistic operators: reproduction, crossover, and mutation. The reproduction process serves to retain those strategies with high fitness i.e. favorable objective f ( p ) , the cross over operator seeks to improve the design by combining the high-fitness strategies, and the mutation operator protects against convergence to a local minimum by adding random noise. The effect of these operators is that designs with high fitness will persist as the management policies evolve from one generation to the next to the final set of Pareto-optimal solutions. Groundwater Soil Natural Resources Structure & Function Allocation Decision Human Actions (Private / Public) Identification of Suitable Compromise Solution Ecosystems Pr Fu odu n c ct tio ion ns Strategic Environment Institutions / Political System Natural Resources Goods & Services on ati fic eoffs i t an d Qu f Tra O Set of Pareto-Optimal Allocation Policies Identification of Objectives Surface Water Decisionmakers Tradeoff Analysis Preferences Figure 2: Identification and quantification of policy options. Boxes denote spatially distributed system models and their outcomes. The blue curved lines denote linkages by fluxes between system components. Figure 2 shows a sketch of our approach to IWRM. Step 3 and 4 of the traditional MCDM are replaced by multi-objective optimization. Converse to the MCDM, our approach therefore carries out search before decision making. That is, optimization is performed without any preference information by the decision-makers given (Hwang and Masud 1979; Horn 1997)1. The set of Pareto-optimal solutions then provides the necessary input for the subsequent decision-making process that is further support by the methods from MCDM. The impacts of a given allocation policy on goods and services from natural resources and the objectives under consideration are difficult to assess since the natural resources are complex, dynamic, variable, interconnected and show nonlinear response. Therefore, the individual systems within the investigated catchment and their linkages shall be approached by a component-based modeling approach (see Error! Reference source not found. for an example of a set of relevant hydrological system components of a particular catchment). Such top-down modeling has major advantages such as quick identification of relevant processes on the catchment scale, modular component updating once more detailed system models become available during the research project and ease of coordination between individual work packages and across disciplines. Zambezi River Kafue River Itezhitezhi Dam Lake Malawi Kafue Flat Wetland Luangwa River Kafue Dam Shire River 700 m3/s 700 m3/s Zambezi Delta 320 m3/s 1060 m3/s 2100 m3/s Kariba Dam 3500 m3/s Cahora Bassa Dam Figure 3: Schematic systems representation of the major features in the Zambezi catchment. Blue triangles denote reservoirs and lakes whereas the green triangles show wetlands. The fluxes are annual averages. Only major tributaries to the Zambezi are shown. As with MCDM, our approach requires first the identification of objectives by all the stakeholders involved in policymaking and those affected by such policy or with vested interests respectively. Without being an exhaustive list, objectives may include 1 project costs [monetary unit] (min); Δ hydropower production [kWh] (max); Δ extent of irrigated area [km2] (max); Δ crop production [t/a] (max) ; Depending on how the optimization and search process are combined, multi-objective optimization methods can be broadly classified into: 1) Decision making before search; 2) Search before decision making and 3) Decision making during search. Δ extension of flooded area in existing wetlands [km2] (max); Δ average depth to groundwater in specific locations that are sensitive to salination [m] (max); number of people to be resettled [-] (min); Δ mean annual downstream runoff [m3/a] (max); sedimentation rate in reservoir [m3/a] (min) estimated loss in biodiversity of downstream ecosystems [# species] (min) ... The definition of objectives can be supported by preliminary results from the componentbased model as well as a stakeholder analysis as performed in step 1). As in the case of the economic valuation approach, failure to include essential objectives in the optimization approach can possibly lead to unacceptable results on the basin-scale. Note that there exists no limitation with regard to the number of objectives under consideration. One of the major difficulties in the presence of n objectives with n ³ 2 is the task of visualization of the results for the purpose of communication optimization outcomes. Due to the stochastic nature of the systems under consideration, the quantification of uncertainty is an important task so as to identify robust management policies. To search for such policies, evolutionary algorithms should work on an expected objective vector function by means of Monte Carlo integration (Jin and Branke 2005). Care, however, must be taken that the problem remains computationally tractable since the evolutionary optimization approaches require intense computational efforts. Therefore, the inherent parallel nature of evolutionary computation should be exploited. Furthermore, subsequent model evaluation results can be utilized to construct systems response functions. These functions mimic the actual system models in an approximative manner. Iterative updating of the response functions guarantees an increased quality of the approximation of the individual approximations to the systems. This method is very flexible and capable of handling most complex resource management problems (Rogers and Dowla 1994; Aly and Peralta 1999; Zheng and Wang 2002). In an ideal decision-making process of resource utilization, i.e. in case of complete cooperation, decisions on regional, national or even supranational scales have to be based on local resource characteristics and the aggregate knowledge about the resource. Such commonly shared information then is determining local policy decisions of the distribution of economic activity throughout the catchment. In case of fragmentation however, that is in the prevalence of non-cooperative behavior and resource utilization policies, information on the state of resources gets a private good since only partial knowledge on the systems state within ones own political border can be obtained. Furthermore, the other decision makers strategies remain uncertain. Even under the truly hypothetical assumption of complete knowledge with regard to the state of the natural systems and their dynamics, the resources appear stochastic in the absence of full cooperation and information sharing (Varis 1997; Varis 1998). From the decision makers perspective, the value of information therefore becomes an important aspect in IWRM. Decision makers not only choose between policies based on the amount of reward they get or on how these policies change the states of the systems but also on how much information they provide (Kaelbling, Littman et al. 1998). This implies that there is no analytical distinction to be drawn between individual actions taken to change the systems states (such as the installation of a groundwater supply system to cover demand) and actions taken to gain information (such as geoexploration). To properly account for uncertainty and, with that, the value of information. Namely, stochastic state transition models within an agent-based optimization have to be implemented. In other words, future states are not known by certainty to a particular decision-maker but rather perceived as probability distributions over states. State transitions models in the case of a great number of agents (decision makers) can be complex. For a compact representation of the former, dynamic Bayesian networks can be utilized (Pearl 2001). Several studies in the field of AI have shown that in case of an agent representation that incorporates probabilities for the unobserved aspects of the resource state and the actions of the competitors, better overall performance can be achieved than in the case of an implementation of purely logical agents which are situated in a deterministic environment (for a comment on that see (Russell and Norvig 2003)). In summary, by the combination of spatially distributed systems modeling with state of the art optimization techniques and methods from the field of AI as well as MCDM, we aim at the development of an Assessment Toolbox for large, water related projects on the catchment scale. In particular, it will serve as a means to assess and quantify the implications of different policy option from environmental, social, economic and political viewpoints. allow to test working procedures for integrating scientific assessment into stakeholder analysis and political. be utilized as a descriptive as well as prescriptive tool. Descriptive because it helps to explain the emergence of likely allocation policies within a certain institutional and political environment in the catchment. Prescriptive in the sense of a normative tool which can identify Pareto–optimal policy designs that, compared to a baseline, are beneficiary for the rational decision makers if adopted and thus constitutes a indispensable input to decision making. References Acharya, G. and E. Barbier (2002). "Using domestic water analysis to value groundwater recharge in the Hadejia-Jama'are floodplain, Northern Nigeria." American Journal of Agricultural Economics 84(2): 415-426. Aly, A. H. and R. C. Peralta (1999). "Optimal design of aquifer cleanup systems under uncertainty using a neural network and a genetic algorithm." Water Resources Research 35(8): 2523-2532. Cropper, M. L. and W. E. Oates (1992). "Environmental Economics: A Survey." Journal of Economic Literature 30(2): 675-740. Deb, K. (2001). Multi-objective optimization using evolutionary algorithms. Chichester, John Wiley & Sons. Heal, G. (2004). Valuing Ecosystem Services, Toward Better Environmental DecisionMaking, National Research Council. Heathcote, I. W. (1998). Integrated watershed management principles and practice. New York, Wiley. Horn, J. (1997). Multicriteria Decision Making. Handbook of evolutionary Computation. T. Bäck, D. B. Fogel and Z. Michalewicz. Bristol (UK), Institute of Physics Publishing. Hwang, C. L. and A. M. Masud (1979). Multiple objective decision making methods and applications a state-of-the-art survey. Berlin a.o., Springer. Jin, Y. and H. Branke (2005). "Evolutionary optimization in uncertain environments - A survey." Ieee Transactions on Evolutionary Computation 9(3): 303-317. Kaelbling, L. P., M. L. Littman, et al. (1998). "Planning and acting in partially observable stochastic domains." Artificial Intelligence 101(1-2): 99-134. Lahmeyer, I. (1998). Cost-Benefit Analysis in Dam Construction and Operation. National Research Council (United States) Committee on Valuing Ground Water (1997). Valuing ground water economic concepts and approaches. Washington D.C., National Academy Press. Pearl, J. (2001). Causality models, reasoning and inference. Cambridge, Cambridge University Press. Reichert, P., M. Borsuk, et al. (2004). Concepts of Decision Support for River Rehabilitation. EAWAG Manuscript. Dübendorf. Rogers, L. L. and F. U. Dowla (1994). "Optimization of ground water remediation using artificial neural networks with parallel solute modelling." Water Resources Research 30(2): 457-481. Russell, S. J. and P. Norvig (2003). Artificial intelligence a modern approach. Upper Saddle River, Prentice Hall. Siegfried, T. U. (2004). Optimal utilization of a non-renewable transboundary groundwater resource - Methodology, case study and policy implications. Zürich. Steiguer, J. E. d., L. Liberti, et al. (2003). Multi-Criteria Decision Models for Forestry and Natural Resources Management: An Annotated Bibliography. USDA, Northeastern Research Station. Turner, R. K. (1993). Sustainable environmental economics and management principles and practice. Chichester etc., Wiley. Varis, O. (1997). "Bayesian decision analysis for environmental and resource management." Environmental Modelling & Software 12(2-3): 177-185. Varis, O. (1998). "A belief network approach to optimization and parameter estimation: application to resource and environmental management." Artificial Intelligence 101(1-2): 135-163. Zheng, C. and P. Wang (2002). "A Field Demonstration of the Simulation Optimization Approach for Remediation System Design." Ground Water 40(3): 258-265. Zitzler, E. and L. Thiele (1999). "Multiobjective evolutionary algorithms: A comparative case study and the Strength Pareto approach." Ieee Transactions on Evolutionary Computation 3(4): 257-271. Zitzler, E., L. Thiele, et al. (2003). "Performance assessment of multiobjective optimizers: An analysis and review." Ieee Transactions on Evolutionary Computation 7(2): 117-132.