A framework for linking advanced simulation models with interactive cognitive maps

https://doi.org/10.1016/j.envsoft.2008.02.006Get rights and content

Abstract

There is a dichotomy between advanced simulation models and flexible, simple tools for supporting policy-making. The former is difficult to use for policy-makers and the latter lacks in analytical value. It is a step forward to link these two types of tools in a way that enables the analytical value of the advanced models, while retaining the flexibility and comprehensibility of the simple tools. This paper presents a framework for such a linkage. The framework is based on an interactive cognitive mapping tool, which uses the qualitative probabilistic network (QPN) formalism to make qualitative (sign-based) calculations. This paper shows that there are several differences that need to be bridged. Each of these is discussed and approaches are presented. It is shown that (1) QPNs can be linked consistently to models with deterministic functions and continuous variables; (2) it is possible to deal with spatially and temporally explicit information; (3) despite the fact that QPNs must be a-cyclic, it is possible to capture feedback loops in a QPN-based tool. To prevent that negative feedback loops automatically result in ambiguous influences, we used a heuristic approach. The framework has been illustrated by analysing two models from literature with the QPN-based method.

Introduction

The field of environmental modelling and software has a long history of use in the study of water, air and soil quality, ecology and biodiversity, hydrology, geomorphology, as well as all kinds of coupled models. They can be either statistical, empirical models or theoretical, mechanistic models (Clark, 2006). Most of these instruments are advanced simulation models developed and used by the scientific community. In general, these models are not designed for use by policy-makers. This contributes to a gap between science and policy, since much of this model-based knowledge and information is under-utilised by policy-makers (Boogerd, 2005, Dahinden et al., 2000, Funtowicz and Ravetz, 1990, Harris, 2002, in 't Veld, 2000, Parker et al., 2002, Toth and Hizsnyik, 1998). Some instruments have been developed with the specific aim of being interactively accessible and adjustable by policy-makers. Examples include Topic (RIKS, 2006) and the Estuary Decision Support System (RA, 2006). These flexible and comprehensible instruments facilitate by interactively structuring the problem at hand. However, the analytical value of these policy-tools is limited, since they are not designed for impact assessment (van Kouwen et al., 2005, van Kouwen et al., 2008). It seems clear that there is a gap between simulation models developed by researchers and the interactive, relatively simple tools used for supporting policy-making.

In order to explain this gap, the information level is a key factor (van Koningsveld, 2003); policy-makers and managers are working on a higher level of aggregation than modellers and researchers (Boogerd, 2005, van Koningsveld et al., 2003). Environmental policy-making deals with a higher level of aggregation where fundamental ‘laws of nature’ are not appropriate (Wainwright and Mulligan, 2003). In order to connect the worlds of science and policy, it will be necessary to aggregate the information of any simulation model to the level of policy-making. An approach known as cognitive mapping has an information level suited to policy-making (Axelrod, 1976; see Section 2.2). It is based on cause-and-effect relationships, graphically represented with box-and-arrow diagrams (directed graphs). Cognitive mapping is valued as an interactive method for problem structuring and decision-making in groups (Axelrod, 1976, Eden and Ackermann, 2004, Vennix, 1996). This paper presents a framework for multilevel modelling, in an attempt to bridge the gap between cognitive mapping and environmental simulation models.

The formalism of qualitative probabilistic networks (QPNs) provides a formal basis for an interactive computer tool which facilitates cognitive mapping (Wellman, 1994). As such, QPNs can be an interface for policy-makers. The formalism of QPNs was developed in the early 1990s and is basically a qualitative abstraction of Bayesian belief networks (van Kouwen et al., in press-b, Wellman, 1990). Bayesian belief networks were developed to deal with uncertain or incomplete knowledge, and comprise a compact representation of a joint probability distribution on a set of variables (Pearl, 1988).

There are three main reasons why a Bayesian network approach is useful in the domain of environmental problems. First, a Bayesian network can explicitly deal with uncertainty, which is inevitably connected to environmental problems and sustainability issues (van Asselt, 2000, van der Sluijs, 1997). Second, the complexity of a system can be represented without the need to integrate processes at different spatial and/or temporal scales (Ticehurst et al., 2007). Third, these networks can be used to bridge the gap between statistical and theoretical models (Clark and Gelfland, 2006). Compared to regular Bayesian networks, the qualitative nature of QPNs makes them less specific. In the field of environmental problems, quantitative information is usually lacking (Kuipers, 1994, McIntosh, 2003). This implies that certain processes can only be modelled by means of a qualitative approach. The QPN technique is such a qualitative modelling approach, which can also be used to facilitate problem structuring dialogues between policy-makers, researchers and stakeholders (van Kouwen et al., in press-a).

This paper presents a framework that allows one to make a linkage between a qualitative, QPN-based cognitive mapping tool and quantitative simulation models. To do so, we will show that the information within the quantitative models can be aggregated to a level which matches that of the QPN. Three major differences between QPNs and simulation models will need to be bridged. First, QPNs define probabilistic functions between variables that are required to be discrete. Many simulation models are based on deterministic functions and continuous variables. Second, QPNs are formally not spatially or temporally explicit, unlike many simulation models. Third, whereas many simulation models are dynamic systems with feedback mechanisms, traditional QPNs are not allowed to deal with feedback loops. We will show that for any advanced simulation model, a QPN model can be derived for it. In this transformation, the aim is to keep the QPN consistent with the advanced model. We will illustrate our approach with some examples of simulation models from literature that were aggregated to a QPN-based model.

Section snippets

Preliminaries

Similar to a Bayesian belief network, a qualitative probabilistic network includes a graphical representation of the independencies between variables. In the graph, a node represents each variable. We will first give a definition for this graphical part. Next, we will elaborate on causal loop diagrams and cognitive mapping, and a definition for the QPN formalism will be given.

Differences bridged

There are a number of differences that must be bridged between QPNs and simulation models. Apart from QPNs being qualitative, variables of QPNs must be discrete, whereas most simulation models have continuous variables. In addition, QPNs describe probabilistic functions whereas most mechanistic simulation models have deterministic functions (usually defined with differential equations). Many relationships between variables in simulation models are optimum curves and have to deal with

Application of the framework in practice

Since we implemented the QPN-based cognitive mapping tool in Java™, we were able to do practical experiments. We will start with a model concerning the eutrophication in lakes as presented by Scheffer et al. (1993), which is shown in Fig. 9. The diagram in Fig. 9 only contains positive feedback loops. One arrow does not directly fit into the QPN formalism; it points at another arrow instead of a variable. The relationship between fish and zooplankton is negative, and the vegetation has a

Conclusions

The aim of this paper was to present a framework that allows one to bridge the distinct levels of policy-oriented cognitive mapping and science-oriented environmental modelling with simulation models. It was shown that (1) QPNs can be linked consistently to models with deterministic functions and continuous variables, (2) it is possible to deal with spatially and temporally explicit information (3) despite the fact that QPNs must be a-cyclic, it is possible to capture feedback loops in a

Discussion

Cognitive mapping is a flexible and easy-to-use instrument which is helpful for identifying the structure of a complex problem (Vennix, 1996). As such, it can help in exploring what needs to be modelled. To reach the ultimate explanatory goals of modelling, we need to provide the means of finding optimal model structures (Wainwright and Mulligan, 2003). We have shown that QPN models can be derived from existing simulation models. This can be useful in making the model easier to understand for

Acknowledgements

We would like to thank Sjors Provoost for his M.Sc. research on qualitative modelling. We would also like to thank Peter de Ruiter, Stefan Dekker and Max Rietkerk for having fruitful discussions on dynamic modelling and feedback mechanisms. We are grateful to Remko Holtkamp, Jasper van Belle and Maarten Eppinga for allowing us to make qualitative diagrams of their models and to compare these qualitative results with their own model outcomes. Finally, we thank Gabriel Olsen and Chris Cooper for

References (61)

  • R. Nash

    Causal network methodology – Tourism Research Applications

    Annals of Tourism Research

    (2006)
  • P. Parker

    Progress in integrated assessment and modelling

    Environmental Modelling and Software

    (2002)
  • A. Perdicoúlis et al.

    Causal networks in EIA

    Environmental Impact Assessment Review

    (2006)
  • M. Scheffer et al.

    Alternative equilibria in shallow lakes

    Trends in Ecology and Evolution

    (1993)
  • K. Soini

    Exploring human dimensions of multifunctional landscapes through mapping and map-making

    Landscape and Urban Planning

    (2001)
  • J.L. Ticehurst et al.

    A Bayesian network approach for assessing the sustainability of coastal lakes in New South Wales, Australia

    Environmental Modelling and Software

    (2007)
  • M.P. Wellman

    Fundamental concepts of qualitative probabilistic networks

    Artificial Intelligence

    (1990)
  • M.P. Wellman

    Inference in cognitive maps

    Mathematics and Computers in Simulation

    (1994)
  • R. Axelrod

    Structure of Decision – The Cognitive Maps of Political Elites

    (1976)
  • F. Ackermann et al.

    Using causal mapping with group support systems to elicit an understanding of failure in complex projects: some implications for organizational research

    Group Decision and Negotiation

    (2005)
  • D.A. Bessler et al.

    Money and prices: U.S. data 1869–1914 (a study with directed graphs)

    Empirical Economics

    (2002)
  • Blalock

    Causal Models in the Social Sciences

    (1974)
  • Boogerd, J.L.M., 2005. Van Droge Kennis naar Natte Natuur. PhD thesis, Vrije Universiteit,...
  • M.C. Bootsma et al.

    Rehabilitation of acidified floating fens by means of addition of buffered surface water

    Restoration Ecology

    (2002)
  • J.S. Clark

    Models for Ecological Data

    (2006)
  • H. Couclelis

    Where has the future gone?” Rethinking the role of integrated land-use models in spatial planning

    Environment and Planning A

    (2005)
  • B.F.W. Croke

    Integrated assessment of water resources: Australian experiences

    Water Resources Management

    (2007)
  • U. Dahinden et al.

    Exploring the use of computer models in participatory integrated assessment – experiences and recommendations for further steps

    Integrated Assessment

    (2000)
  • H. de Mars et al.

    Redox potentials in relation to water levels in different mire types in the Netherlands and Poland

    Plant Ecology

    (1999)
  • M.J. Druzdzel et al.

    Efficient reasoning in qualitative probabilistic networks

  • Cited by (21)

    • A methodological framework to support the initiation, design and institutionalization of participatory modeling processes in water resources management

      2018, Journal of Hydrology
      Citation Excerpt :

      Fuzzy cognitive mapping allows for assessing the plausibility of cognitive maps and generating scenarios (e.g., van Vliet et al., 2010; Jetter and Kok, 2014). Causal loop diagrams can also be used in the design of qualitative and quantitative Bayesian networks (van Kouwen et al., 2008; Giordano et al., 2013). The modeling process usually proceeds in an iterative manner.

    • Inference reasoning on fishers' knowledge using Bayesian causal maps

      2015, Ecological Informatics
      Citation Excerpt :

      But the authors insist on the need to adequately train the stakeholders (regarding probability theory) which is deemed unnecessary with our method. Finally, a similar approach to ours can be found in (Kouwen et al., 2008), where qualitative Bayesian networks are built, in which inference is enabled thanks to sign-propagation algorithms. But unfortunately, sign propagation does not solve the problem of ambiguity in qualitative diagrams.

    • Formalizing knowledge on international environmental regimes: A first step towards integrating political science in integrated assessments of global environmental change

      2013, Environmental Modelling and Software
      Citation Excerpt :

      These methods also focus on a more global characterization of the system, in terms of relationships between parts/components/subsystems, combining quantitative as well as qualitative knowledge on these characteristics and using some form of inference rules, being probabilistic (as in BBNs) or logico-linguistic (as in FL). E.g. Qualitative Bayesian Belief Networks have been used to study the role and effects of institutional settings in complex water management issues (Saravanan, 2008), and enable also linkage with system dynamic modelling (van Kouwen et al., 2008). Given that our conceptual framework is based on rather qualitative rules expressing to what extent certain determinants influence the likelihood of regime formation and implementation, the use of such a qualitative probabilistic modelling set up might seem rather appropriate to put our conceptual framework into a model.

    • Bringing diverse knowledge sources together - A meta-model for supporting integrated catchment management

      2012, Journal of Environmental Management
      Citation Excerpt :

      Thus, a variety of different meta-models can be constructed in phase 4 for different purposes, each derived from the same complex coupled model created in phase 3. Bayesian Networks (BNs) provide a promising route for implementing the meta-model, and are particularly attractive in the context of representing uncertainties (van Kouwen et al., 2008). The BN approach is based on a directional graph representing cause–effect relationships in the system.

    • Modelling with stakeholders

      2010, Environmental Modelling and Software
      Citation Excerpt :

      They can help frame the discussions with stakeholders, whose expert opinions can be used to put together the initial maps (Özesmi and Özesmi, 2004; van Vliet et al., 2010). Kouwen et al. (2008) suggest that there is much promise in integrating cognitive mapping and environmental simulation models, using qualitative probabilistic networks such as BNs. FCMapper is a free software package to process FCMs and calculate important metrics for them.

    View all citing articles on Scopus
    View full text