Environmental Modelling & Software 26 (2011) 837e844 Contents lists available at ScienceDirect Environmental Modelling & Software journal homepage: www.elsevier.com/locate/envsoft Empirical characterisation of agent behaviours in socio-ecological systems Alex Smajgl a, *, Daniel G. Brown b, Diego Valbuena c, d, Marco G.A. Huigen e a CSIRO Ecosystems Sciences and Climate Adaptation, University Drive, Townsville 4810, Australia University of Michigan, School of Natural Resources and Environment, Ann Arbor, MI, USA c Wageningen University, Department of Environmental Sciences, P.O. Box 47, 6700 AA Wageningen, The Netherlands d International Livestock Research Institute, System-wide Livestock Programme, P.O. Box 5689, Addis Ababa, Ethiopia e Universität Hohenheim, Germany b a r t i c l e i n f o a b s t r a c t Article history: Received 29 July 2010 Received in revised form 14 February 2011 Accepted 17 February 2011 Agent-based modelling has become an important tool to investigate socio-ecological processes. Its use is partially driven by increasing demand from decision makers to provide support for understanding the potential implications of decisions in complex situations. While one of the advantages of agent-based modelling is the ability to simulate the implications of human decision-making processes explicitly, methods for providing empirical support for the representation of the behaviour of human agents have not been structured systematically. This paper develops a framework for the parameterisation of human behaviour in agent-based models and develops twelve distinct sequences for the characterisation and parameterisation of human behaviours. Examples are provided to illustrate the most important sequences. This framework is a first step towards a guide for parameterisation of human behaviour in ABM. A structured discussion within the agent-based community is needed to achieve a more definitive guideline. Ó 2011 Elsevier Ltd. All rights reserved. Keyword: Agent-based modelling 1. Introduction Environmental policy and management increasingly demand the integration of cross-disciplinary knowledge of socio-ecological systems. Understanding human responses to environmental and/or policy changes is critical for understanding socio-ecological systems and the outcomes they produce. As agent-based modelling (ABM) allows for simulating actual decision-making processes of individuals or groups of individuals this modelling technique is gaining importance (Matthews et al., 2007; Parker et al., 2003). This paper is focused on ABM, in which (a) human agents are represented at the scale of households or individuals and (b) the main concern is the understanding of socio-ecological interactions. In empirical contexts, this class of ABM is mostly chosen for its ability to simulate explicit human decision-making processes (mostly described in discontinuous functions) in the context of changes in the bio-physical environment. However, the main strength of ABM can be a key weakness. As the implementation of human decision-making processes is the main strength of ABM, the agent attributes and behavioural response functions that represent these processes require knowledge support from qualitative and/or * Corresponding author. E-mail address: alex.smajgl@csiro.au (A. Smajgl). 1364-8152/$ e see front matter Ó 2011 Elsevier Ltd. All rights reserved. doi:10.1016/j.envsoft.2011.02.011 quantitative empirical sources. Unfortunately, there are no standard approaches to documenting and communicating the empirical support that underlies modelling and design decisions in agentbased models (Berger and Schreinemachers, 2006; Bohensky et al., 2007; Smajgl et al., 2007). Empirical characterisation and parameterisation of human decision-making processes can involve a range of methods that include: expert knowledge, surveys, interviews, and participant observation. We describe a set of specific methods here that, while it does not include all possible options, includes those that are most commonly used in practice and help define what we believe to be a robust set of sequences within which alternative specific methods can be applied. In many cases it is necessary to iteratively combine several empirical methods in a sequence to fully parameterise behavioural responses in ABM. The development of an ABM that aims to simulate behavioural responses of humans requires two fundamental steps in which empirical data are required: the development of behavioural categories and the scaling to the whole population of agents. The effectiveness of different methodological sequences depends on the modelling context (Janssen and Ostrom, 2006; Robinson et al., 2007). Most significantly, the size of the human population in the system to be modelled helps to determine appropriateness of various methods, and can be anything from 20 individuals or households to many million households. Carrying out in-depth 838 A. Smajgl et al. / Environmental Modelling & Software 26 (2011) 837e844 interviews with 20 households is realistic. With larger populations and more complex socio-ecological interactions, it becomes necessary to work with a sample of the population. Sampling often requires characterisation of the existing heterogeneity of agent attributes and behavioural responses in order to provide a simplified description of behavioural realities that can guide sampling and the realistic parameterisation of model agents. The first challenge for the parameterisation of ABMs that represent larger populations is the identification of suitable empirical methods for eliciting behavioural data. Then, the sample-based data need to be translated into behavioural representations for the whole population (i.e. up-scaling). Up-scaling can be performed proportionally or disproportionally. Proportional up-scaling assumes that the sample is representative for the whole population. This can result in under-representation of behavioural minorities. Disproportional up-scaling is needed if these behavioural minorities are likely to determine system dynamics, such as free-riders in common pool situations. Disproportional up-scaling is based on the development of agent typologies. The inverse problem (i.e., down-scaling) occurs when data for a large collection of agents are acquired for aggregated sources (e.g., over census enumeration districts), and populations of agents are generated through some simulation technique that honours the aggregate data (Baker et al., 2010; Brown and Robinson, 2006; Wu et al., 2008). This paper aims to provide an initial structure that can be used within rigorous discussions of parameterisation methods for human behaviour in ABM. This is only a first step towards a guide for parameterisation of human behaviour in ABM as a structured discussion within the agent-based community is needed to achieve a more definite guideline. Thus, we developed a framework for this process followed by a discussion of methods available for what we call the parameterisation sequence. Discussing sequences instead of individual methods allows for consideration of approaches that mix different kinds of methods, a systematic understanding of existing options and evaluation of suitability across a range of modelling contexts. Then, we give examples demonstrating the application of the framework. Most published examples of empirical agent-based models do not disclose methodological details for the parameterisation process that was used. Therefore, we believe that an initial outline of the dimensions of a useful framework along these lines is needed. The systematic application of the framework to a few different types of empirical agent-based models should guide readers in applying the framework and considering its applicability in a wider variety of cases. The paper concludes with a discussion of the need to further tests of the generality of the framework. 2. Definitions of empirical methods Several methods can be used to parameterise behavioural responses of humans empirically. This set of the most commonly used is not all inclusive, but we posit that it is complete enough to build a fairly comprehensive framework: Expert knowledge: This constitutes a variety of formal and informal methods for capturing the understanding of experts in a field or region and representing that knowledge in ways that it can be used in models. Approaches range from conceptual mapping by experts themselves to informal conversations or focus groups from which expert knowledge can be elicited for subsequent model building. Expert knowledge can also be employed to quantify uncertainties associated with expert judgement (Cooke and Goossens, 2008). Participant observation: Becker (1958) defines participant observation as the process in which the scientist participates in and documents the daily life of communities. Social surveys: Survey instruments consist of a list of questions, each with pre-defined sets of possible answers (Nichols, 1991). Responses are elicited via mail, email, in person, or via phone. Interviews: While a survey comprises of mostly closed questions, interviews are normally less structured (Jupp, 2006) and range from a list of open questions to unstructured dialogues. Census data: While surveys and interviews are conducted with a sample of a population census data is elicited for 100% of a population, normally within national boundaries (Rees et al., 2002). Aggregated census data summarise responses for groups of households within enumeration districts, while disaggregated census data show household-level responses. Field or lab experiments: Experiments are designed to observe how the change in an independent variable affects a dependent variable. Traditionally, this happened in laboratories, which allows for high degrees of control and hence clear causal links in observed outcomes. However, the high levels of control create artificial situations, which can lead to poor applicability of results to more realistic situations (Patzer, 1996). Field experiments aim for less artificialness by placing the experiment in the natural environment (Harrison and List, 2004; Smajgl et al., 2009e). While field experiments allow for more realistic behaviour to be observed the levels of control decrease. Role-playing games: Barreteau (2003) defines Role-Playing Games as “group settings that determine the roles or behavioural patterns of players as well as an imaginary context.” Similar to experimental designs, people are given pre-defined roles and tasks, which they have to perform in interaction with other role players in a pre-defined setting. Cluster analysis: Clustering describes the grouping of subjects based on the similarity of attributes. Based on categorical data, algorithms calculate either (1) the Euclidean distance between the median and each value, or (2) between each pair of values (Rousseeuw, 1987). Depending on proximity, subjects are mapped into clusters. Dasymetric mapping: This method involves a combination of detailed spatial data with aggregated census data, usually about population, to create disaggregated representations of the spatial distribution of population characteristics (Mennis and Hultgren, 2006). Often, for example, land-cover maps based on satellite imagery are used to distribute populations within an area such that (a) population totals within census enumeration districts are preserved and (b) their spatial locations are estimated at a much finer resolution, based on the locations of land-cover types with different population densities. Monte Carlo method: This approach refers to “experiments with algebraic models which involve a stochastic structure” (Martin, 1977). Monte Carlo runs with stochastic agent-based models allow developing uncertainty distributions of output variables. 3. Parameterisation framework ABM requires the systematic representation of three main phenomena: agents, their social networks and the agent environment (Fig. 1). This paper discusses only the parameterisation of human agents; we leave development of empirical frameworks for social networks and the agent environment to others. We acknowledge that any model builder will not be able to separate human agent attributes and behaviours from social networks and the (bio-physical) agent environment, but for the purpose of clarity in this paper we focused on human agents. To achieve agent parameterisation, six different methodological steps were identified (Fig. 1), each with multiple empirical methods available (Table 1). The first step of the framework is to identify different classes of agents (Method 1, Fig. 1). We assume that agents that share the same 839 A. Smajgl et al. / Environmental Modelling & Software 26 (2011) 837e844 Fig. 1. Framework for parameterisation of ABM. sequence of actions are defined in an agent class. The sequences of actions represent the modelled behaviour of the agents and constitute the totality of their actions within the model. The identification of distinct agent classes and their sequences of actions can be based on expert knowledge (EK), participant observation (PO), laboratory experiments (LE), interviews and role-playing games (RPG). We group these methods as Methods 1 (M1) as listed in Table 1. Within each agent class agents can be grouped into agent types according to their behavioural similarity. For instance, farmers would be a different class of agents from urban developers. Among farmers, different behavioural responses might be distinguished based on attributes like wealth or age, and therefore separate farmer types can be created. Types can be developed in three further steps employing methods 2e4. If no agent types are required, attributes and behavioural components can be assigned for the whole agent class employing methods 2 and 3. The second step of the framework can be either to specify the values for agent attributes (M2) or the parameters for the behavioural rules agents follow (M3). While agents in each class share the same sequence of actions, the parameters governing the magnitudes of the actions and the degree to which various sources of information influence those actions can vary. Agents with the same behaviours and similar or identical parameters of behaviour belong to the same agent type (i.e., types are the equivalent of sub-classes). Agent types can be developed from agent attributes or from data on agent behaviours. Thus, data have to be elicited to measure realworld attributes and behavioural responses. Behavioural responses can include observed relationships between qualitative or quantitative contextual variables and the actions or probabilities of actions carried out by agents. We group methods for obtaining agent attributes as Methods 2 (M2) and those for obtaining behavioural responses as Methods 3 (M3). Table 1 lists surveys and census data as M2 methods. Options for step M3 include surveys, interviews, field experiments, participant observation, RPG, timeseries data, and expert knowledge. The sequence of steps within the proposed framework can vary (Fig. 1) because agent types can be developed from agent attributes or from observed behavioural responses. If types are developed from agent attributes, M4a methods can be employed, through the combination of (a) clustering and regression, (b) a mix of correlation and expert knowledge, (c) expert knowledge individually, or (d) dasymetric mapping. The first three options can also be used if developing agent types from behavioural data. If methods M4a are not used, participant observation is an alternative (method group M4b). Once agent types are developed, and their attributes and behavioural response functions characterised, agents in the whole population need to be assigned to the appropriate agent type (M5). If data were elicited from the whole population, this step does not require any up- or down-scaling, as the sub-population (sample) and population are identical (Fig. 1). If there are some assurances that the sample is representative of the population, the scaling can be conducted proportionally. Under such assumptions each sample is assumed to represent a number of other agents, determined by the proportion of the population sampled, and is then cloned, accordingly. Cloning means that each respondent is replicated x times, with x ¼ (population size/sample size). Cloning does not mean that all agents of one clone have to be identical, as ranges of values or measures of variability can be used to introduce some level of heterogeneity. If representativeness of the sample cannot be assumed, agent types can be up- or down-scaled by means of census data or a Monte Carlo approach. These methods are grouped as M5 methods and listed in Table 1. 4. Parameterisation sequences What combinations of M1eM5 methods make sense will depend on the empirical context, including its purpose. We chose to simplify a description of the empirical context using three variables: the size of the population (N), high/low; the behavioural diversity (BD), high/ low; and the pragmatic possibility of creating representative samples. Behavioural diversity describes the relative number of different types of actions that agents can take in a given context. If the options are relatively limited, e.g., to grow rice or not, then behavioural diversity is low. If the agents choose among a large set of options, behavioural diversity is said to be high. Whether or not it is possible to create a representative sample of the population depends on, for instance, availability of funding for the modelling project and ease of accessing the people. We chose not to include model purpose as a dimension in the descriptions of context, though it may have some effect on the empirical choices, because the framework is predicated on the assumption that the interest of the modeller is in creating an empirically realistic model. That assumption necessarily includes some model purposes and excludes others. We have identified twelve different cases or sequences that combine the methods listed in Table 1. These cases can be followed to parameterise the behavioural responses of human agents in ABMs (Table 2). Depending on the actual context this combination of methods might require an iterative approach instead of a strict sequence. There may be other combinations, but we believe this set covers most empirical ABM contexts. Despite being a first step, this is a reasonably comprehensive list of cases that should allow for a structured testing of methods to identify which sequences prove to be effective under various modelling conditions. Such testing could aim for quantifying uncertainty and guide the ABM community towards more robust empirical model development. The Table 1 Overview of methodologies relevant for the parameterisation of behavioural traits of human agents. M1 M2 M3 M4a ATB M4b BT M5 Expert knowledge ParticObser Lab experiment Interviews RPG Survey Census (incl. GIS data) Survey Interviews Field experiment ParticObser RPG Time-series data Expert knowledge Clustering and regression Correlation and expert knowledge Expert knowledge Dasymetric mapping Clustering and regression Correlation þ expert knowledge Expert knowledge ParticObser Proportional Census/GIS based assignment Monte Carlo 840 No up-Scaling CD/GIS based CD/GIS based EK Clustering & Regression Clustering & Regression Census Aggregated Census Disaggregated Census Census EK or PO or RPG EK or PO EK or PO Yes Less relevant Less relevant No Less relevant Less relevant 10 11 12 Less relevant Less relevant Less relevant No No 9 Less relevant Yes 8 Notes: N: agent population; BD: Behavioural diversity; EK: Expert Knowledge; RPG: Role-playing games; PO: Participant observation; CD: Census data; GIS: geographic information systems. EK EK Dasymetric mapping e No up-Scaling EK CD/GIS based No No 7 Less relevant Less relevant Yes 100% Yes No No Less relevant 4 5 6 Yes No access No Yes Yes 3 No Yes Yes 2 No Yes 1 Yes Less relevant EK or PO or Interviews (optional: Experiments EK or PO or Interviews (optional: Experiments EK or PO or Interviews (optional: Experiments EK or PO or RPG EK or PO EK or PO or Interviews (optional: Experiments EK or PO or Interviews (optional: Experiments EK or PO or Interviews (optional: Experiments EK or PO or RPG or RPG) or RPG) or RPG) Survey Census & Surveys (or PO) Disaggregated Census Survey or RPG) Survey OR Census Interviews (PO) or RPG Time-series data AND EK Interviews or Field Experiments or RPG Interviews or Field Experiments or RPG EK or Interviews or Field Experiments or RPG EK or Field Experiments or RPG EK or RPG Time-series data Time-series data e Statistics and EK e PO EK (EK or Clustering) and Regression e Clustering or (Statistics and EK) PO e e Interviews or RPG Survey Survey or RPG) or RPG) No up-scaling CD/GIS based CD/GIS based CD/GIS based CD/GIS based (EK or Clustering) and Regression e Interviews or Field Experiments or RPG or EK Interviews or RPG Survey M3 M2 M1 BD high? Sample representative? N high? When Case Table 2 Overview of techniques for the parameterisation of human agent behavioural responses. e e M4a e M4b M5 (Cloning or Monte Carlo) and Spatial references CD/GIS based A. Smajgl et al. / Environmental Modelling & Software 26 (2011) 837e844 characterisation of sequences should facilitate description and documentation of ABM applications. While model description is clearly important (Grimm et al., 2006, 2010), we argue that good description of the empirical support for models in specific contexts is equally important in allowing scientific peers to make good judgements about the value of the model results. We describe the modelling contexts within which the different cases make sense and emphasise different routes through the parameterisation processes (Fig. 1) depending on these contextual variables. The first case represents a modelling context with a large population and the possibility to develop representative samples (Case 1). An understanding of agents and their actions (M1) can be developed using expert knowledge (EK), participant observation (PO), or interviews. Detailed agent attribute data (M2) can be elicited by conducting sample surveys, while behavioural data (M3) can be obtained from interviews, field experiments, role-playing games (RPG), or expert knowledge (EK). Assuming that the sample used to generate attributes and behavioural parameters is representative, proportional up-scaling can be carried out, in which each data point is cloned to generate the whole population (M5). This step needs to include spatial referencing if agent locations are an element of the model design. This sequence was implemented in an example that is explained in the following section. In cases where the behavioural diversity is high and assumptions on the representativeness of samples do not hold, variations to the first sequence have to be introduced. While steps M1 and M2 can remain unchanged, the behavioural complexity makes field experiments or expert knowledge less suitable in step M3. Instead, two alternative sequences that require the development of agent types (for disproportional up-scaling) can be used. Firstly, interviews can be conducted to generate behavioural data for development of agent types (Case 2). For each type, distributions of agent attribute values can be identified by surveys. This means that the actual sequence of this approach (following Fig. 1) is M1 / M3 / M4b / M2 / M5. Instead of developing agent types from behavioural data, types could be constructed from agent attributes (Case 3). This means that a survey could be conducted to elicit attributes and clustering techniques applied to develop agent types. Then, interviews can be targeted to people who match the profile of agent types. The behavioural data would then be disproportionally up-scaled to the whole population based on a process that assigns each agent to a type by mapping type-specific characteristics against census data. This approach sequences methods as M1 / M2 / M4a / M3 / M5. Cases 2 and 3 require census data to allow for up-scaling (M5). This type of up-scaling requires the identification of the agent types using census data, in order to determine the proportions of the population in each type. While both of these parameterisation sequences reduce uncertainty resulting from sample non-representativeness, both approaches introduce uncertainty from the development of agent types by clustering (i.e. disproportional upscaling). Below examples are given for these two approaches, the Dutch model and the SimPaSI model for East Kalimantan. In situations where representative samples can be developed in spite of high behavioural diversity (Case 4), expert knowledge or participant observation can provide broad systems understanding (M1). Participant observation allows for developing agent types based on attribute data (census or surveys) and behavioural data (interviews or role-playing games). The final up-scaling could be guided by census data. The MamaLuke model is described below to give an example for this particular sequence. Case 5 describes a situation where the actors are not accessible for interviews or surveys (e.g., because they are not responsive to surveys or due to funding limitations). In such cases agent classes can be identified based on expert knowledge or participant observation. This sequence requires the availability of disaggregated data A. Smajgl et al. / Environmental Modelling & Software 26 (2011) 837e844 to provide agent attributes. Assumptions on behavioural responses can be developed from statistical analysis of time-series data or expert knowledge. Experts could then guide the mapping of behavioural assumptions into the agent population. In the next section, the DEED model gives an example for such an approach. Case 6 is another variation of the first sequence. This approach assumes that instead of a proportional sample, the development of agent types is based on behavioural data (M4b). Disproportional up-scaling could be performed with a census-based mapping, as described for Case 2. Case 7 involves relatively small populations. It is very similar to Case 1 but does not require any up-scaling (M5) because the whole population can be described, see for instance (Dray et al., 2006). Case 8 is another variation of Case 1 (and similar to Case 3). This case represents a large population with low behavioural diversity. It assumes that disaggregated census data are sufficiently available for providing agent attribute information. In such a case agent types could be statistically developed from census data or surveys (e.g., for disproportional up-scaling). Then, interviews could be conducted to obtain behavioural response data (M3) before finally behavioural data are mapped via census data into the agent population. The sequence would be M1 / M2 / M4a / M3 / M5. In cases with small populations and available census data, two different sequences can be used, depending on the behavioural diversity. In case of low behavioural diversity (Case 9) agent classes can be identified using expert knowledge, participant observation or role-playing games. Disaggregated census data would provide agent attributes. Agent attributes could further be characterised by adding expert knowledge (M4a). Then, behavioural data can be developed for each agent based on expert knowledge, field experiments, or roleplaying games. Experts could guide the process of refining behavioural assumptions for each agent. While publications on agent-based models mostly lack sufficient detail on parameterisation methods it seems as if Case 9 is quite common (Barnaud et al., 2010; Dung et al., 2009; Mathevet et al., 2003; Smajgl et al., 2009e). The case of high behavioural diversity (Case 10) only differs from Case 9 in that field experiments are likely to become unrealistic for M3 due to complications in designing experiments (Smajgl et al., 2008). The final two cases start with expert knowledge or participant observation to design system properties. Behavioural parameters are then developed through analysis of time-series data (e.g., An et al., 2011), while attribute data are available in the form of aggregated or disaggregated census data. Statistical techniques such as clustering or regression methods can be employed to process the time-series data into agent types (M4b). Aggregated census data has to be disaggregated through dasymetric mapping (Case 11) before behavioural data can be assigned for the whole agent population (i.e., down-scaling). If disaggregated census data are available they can be directly used for proportional up-scaling assumptions on agent behaviour (Case 12). The following section provides empirical examples for the first five cases listed in Table 2 5. Case examples 5.1. SimPaSI model for Central Java Purpose of the model: The SimPaSI model (Simulating Pathways to Sustainability in Indonesia) was implemented for the districts Demak, Jepara and Pati in Central Java. The goal was to provide capacity to analyse and foster a cross-scale dialogue across multiple levels of resource governance about the implications of plans to change fuel subsidies. Main model features: The SimPaSI model (Smajgl et al., 2009a,b) simulates in daily time steps livelihood related activities of about 841 3.2 million people. Within the model, these decisions take place in an environment that is created from best available (polygon-based) GIS data (elevation, land use) and with representations of hydrological dynamics (surface water, groundwater, soil moisture) as well as crop growth (trees, fruit, rice, tapioca, bamboo) and ecological elements (mangroves, corals). The model design document (including pseudo code) can be downloaded from http:// www.cse.csiro.au/downloads/DesignDoc_JaTeng.pdf. The user can vary the price for petrol, kerosene, and electricity, and the level of quarterly cash payments to people below the poverty line. Results are presented (in maps and charts) for poverty, unemployment, high water level risk, rice production, forest cover, coral area, and mangrove area. Sequence applied with details on approach: Empirical application of this model followed the sequence outlined in Case 1. Artificial households and their members were parameterised based on survey data. In total, 3500 households were surveyed to elicit demographic data, perceptions of importance of natural and social resources, and behavioural responses to a total of eight scenarios (i.e., fuel-price changes). The underlying assumption is that the stratified sample is representative and that behavioural responses can be proportionally up-scaled to initialise the total population. This means that each surveyed household is cloned about 350 times to create attributes for model agents. Technically, survey data is read during the initialisation process and instead of creating one agent from each entry line attributes of 350 agents are created and characterised. To assign values for many of these attributes (i.e. wages), value ranges (based on the minimum and maximum from the survey) were used instead of point values. As a result, households and their members are heterogeneous, even within the clusters of similar clones. 5.2. Dutch agriculture model Purpose of the model: the objective of this ABM was to analyse and explore how the response of farmers to changes in policy, expressed through socio-economic scenarios, affect the changes in the use, cover and structure of the land in a Dutch rural region (Valbuena et al., 2010a,b). In this region, farmers’ decisions included processes of farm cessation, farm expansion and protection of hedgerows and tree lines. Main features: In order to represent farm(er)s, 2700 agents were created on a grid that represented 1 ha. Each agent owned one or several fields and each field was formed by one or several pixels. Agents’ decisions were parameterised based on a probabilistic rather than a deterministic approach, in which a probability for a certain choice was assigned to each agent. A run of the model consisted of 15 time steps each representing the actions taken in one year. Agents could stop farming, sell or buy land and protect or clear hedgerows (Valbuena et al., 2010a,b). The outputs include spatially explicit changes in the landscape structure of the rural region. Sequence applied with details on approach: Empirical application of this model followed that sequence outline for Case 2. Only one agent class was identified (i.e., farmers) based on expert knowledge (M1). A sample survey of 300 farmers was used to elicit agent attributes (e.g., farm size, agribusiness type and age) and behavioural components (i.e., agents’ willingness to make certain choices). Both attributes and components were combined to build an agent typology, to describe these differences (M4a and M4b). Although the typology was built based on expert knowledge, it was supported by quantitative analyses (e.g., cluster analysis and classification regression trees). Based on this typology and on the census data, the attributes and the behavioural parameters were assigned to each agent of the whole population (Valbuena et al., 2008). 842 A. Smajgl et al. / Environmental Modelling & Software 26 (2011) 837e844 5.3. SimPaSI model for East Kalimantan Purpose of the model: In this case study, the SimPaSI model was implemented for six districts of the Indonesian province of East Kalimantan. The goal was to provide a simulation tool that facilitates a dialogue across governmental levels in the context of potential fuel subsidy reductions. The underlying idea was to simulate the impacts that central government interventions could have on poverty in different parts of Indonesia. The two selected regions were Central Java and East Kalimantan. Main model features: This implementation of the SimPaSI model (Smajgl et al., 2009c,d) is similar to the Central Java implementation described above. However, land uses (trees, rubber, rattan, fruit) and ecological variables (fish, honey, deer, dolphins, hornbill) differ. Sequence applied with details on approach: Empirical application of this model followed that sequence outline for Case 3. The development of a dataset for parameterising households and their members was based on the assumption that representative sample could not be gathered. While about 53% of the 1.8 million people to be included in the model live in (peri-)urban areas, the rest live in 99% of the area, which is mostly quite remote. Additionally, the complexity of the context required a larger set of data than for the first case example, which translated into interviews of unrealistic length (up to 4 h based on pilot tests). Therefore, field work was structured to identify all possible types in a first step and elicit behavioural response functions for each type separately (Bohensky et al., 2007; Smajgl et al., 2007; Trébuil et al., 1997). In a first step, a survey of 3000 people elicited data on demography, values, and livelihood strategies. In a second step, a K-means cluster analysis was performed based on regression trees, given that most of the variables were categorical. This enabled the development of a decision tree (with a limited set of variables), through which a type could be assigned to each household in the area. This step was followed by validation workshops to test the plausibility of clusters. Then, interviews with 530 households were carried out to elicit behavioural responses to eight scenarios (i.e. fuel prices). These households were identified based on the variables with highest discriminatory power for each cluster. The behavioural responses were scaled up disproportionally and conducted based on (disaggregated) household-level census data, with behavioural response data (defined for types) assigned to each household (in the census). 5.4. MameLuke settlement model Purpose of the model: The MameLuke Settlement model, which has been constructed using the MameLuke (multi-actor modelling of land use and land-cover changes) framework for the development of agent-based models (Huigen, 2004), was implemented for the Isabela province in the Philippines. The aim of the model is to understand the settlement processes in the area. Understanding the settlement process and the spatial effects of population growth in the area are of pivotal importance for the future land use and land-cover change (Huigen, 2004; Huigen et al., 2006). Main model features: The settlement model for the Isabela province simulates all demographic processes relevant to population growth, including migration. The ethnicity is a major driver for the location of settlement of a household. Throughout the simulated years households settle following different rules defined by their ethnic preference structure. Furthermore, the model simulates spatially explicit household land-use behaviour (including deforestation) over a period of 100 years (starting in 1900) and measures changes in land use and cover. Key outcomes of the model are measured using an ‘ethnic spread’ fitness function. Sequence applied with details on approach: Empirical application of this model followed that sequence outline for Case 4. Participant observation was the dominant technique used for the parameterisation of human agents. Participant observation was used to create a principle systems understanding (M1), which includes the identification of ethnic groups as a major driver for the creation of settlements. In total 44 unstructured qualitative interviews in combination with 7 focus group sessions were used to obtain empirical data. The interviews were stratified on age, gender and ethnicity. Attributes and behavioural traits of agents were defined iteratively. While Fig. 1 could be interpreted as linear sequences of methods, the MameLuke implementation was conducted through iteration among steps M2, M3 and M4a/b, using participant observation. The iterative approach allowed for refining the consistency of assumptions on agent behaviour and attributes. Unstructured interviews were conducted to elicit the base structure of an agent typology. This structure was then used as a reference point in focus group discussions. Up-scaling to the total 3000 households that live in the area has been done using the aforementioned agent-type identification technique. Based on the attribute ethnicity in combination with the age of the household (at the specified simulated year) a rule-set for settlement decisions and for land use decisions were generated using the empirical datasets. 5.5. DEED model of ex-urban developments in South-eastern Michigan Purpose of the model: The Dynamic Ex-urban Ecological Development (DEED) was developed to understand the evolution of exurban landscapes in South-eastern Michigan, USA, and identify opportunities to introduce landscape designs that provide enhanced ecosystem services in these heterogeneous humandominated landscapes. Main model features: A key feature of the DEED model is that it represents the interacting behaviours of land developers and residential home buyers (Brown et al., 2008; Robinson and Brown, 2009; Zellner et al., 2009). Land developers are included because of their importance in structuring the supply of residential landscapes to residential land buyers. Each of three different developers builds one of three different development types. The development types are distinguished by their density, where they locate relative to roads and aesthetic features, and their effects on the amount of tree cover on the landscape, thereby providing both aesthetic value to residents and ecological services. Each time a developer creates a development, lots within the development become available to residents, who are created by the model at a user-specified rate per time step. A given developer cannot create a new development until a specified threshold percentage of lots in his existing developments are sold to residents. Sequence applied with details on approach: Empirical support for the developers in this model followed that sequence outline for Case 5. Empirical support for residents was described elsewhere (Brown and Robinson, 2006). The development types were created as a working hypothesis about differentiation among development types within South-eastern Michigan based on expert knowledge. Empirical information was needed to test the differentiation among the development types and to represent the developer behaviours related to selection of sites for developments. The first type of empirical information was largely definitional (i.e., do these types exist?) and was observed by identifying multiple developments in parcel data, geographically overlaying the subdivisions on digital aerial imagery to observe differences in landscape characteristics, and accessing tax records for selected subdivisions to observe differences in house and price characteristics (Brown et al., 2008). A. Smajgl et al. / Environmental Modelling & Software 26 (2011) 837e844 Once the differences among development types were confirmed, important behavioural information about the development types, and by inference the developers of those different types, was observed through analysis of time-series information on a sample of developments created between 1950 and 2000 (An and Brown, 2008). Subdivisions were identified in GIS-based parcel data and labelled for type based on parcel sizes and layouts as well as landscape characteristics observed in scoincident aerial photography. The dates of each development were determined using historical aerial photography, collected approximately decadally over the study period. Survival analysis was used to determine the locational characteristics that determined the likelihood of given farm being converted to each of the three different development types. The coefficients estimated by survival analysis were used as parameters in the developer models (i.e., one for each developer type). 6. Concluding remarks This paper developed a framework for the parameterisation of attributes and behavioural response functions of human agents in agent-based models for a socio-ecological context. This is a first step to allow for systematic testing, documentation and communication of empirical parameterisation methods that can be compared across various modelling contexts. We identified multiple methods that can be used at each of five iterative or sequential steps in the process of creating an empirically informed agent-based model. Until now, methodological choices have often been made in an ad hoc manner, depending on the researcher’s conceptual understanding, methodological training, and access to data. We classified the various modelling contexts according to the size of the agent population, their behavioural diversity, and the ability of researchers to acquire a representative sample of agent attributes or behaviours. This classification provides a structure that can be used to evaluate the appropriateness of various combinations of methods in different contexts, and can formalize these combinations in a way that can guide choices and facilitate communication of those choices among modellers and to practitioners who use model results. Such an effort to systematically understand the effectiveness of methodologies employed in ABM for the parameterisation of behavioural aspects is critical because the (technical) strength of ABM, which is its flexibility and capacity in representing human decision making, has to be based on a robust empirical support. Without translating this advantage into empirical contexts, the applied value of ABM remains limited and the focus remains on hypothetical and theoretical analyses. The framework for and describing different parameterisation sequences should also be tested in various modelling (Janssen and Ostrom, 2006; Robinson et al., 2007) and other application contexts (Matthews et al., 2007) such as the growing domain of Social Epidemiology in public health research (Kaplan, 2004) or agent-based economics (Luna and Perrone, 2002). Further refinement and testing of this framework would allow for reducing model uncertainty, for provision of systematic guidelines for building and applying empirical ABMs, and for facilitation of the model comparisons (Parker et al., 2008a,b). Particular sequences could be assessed by comparing their performance, which links directly into the related topic of model validation (Grimm et al., 2005; Moss, 2008; Smajgl et al., 2011). The framework described in this paper is a first step in this process. The community of ABM researchers needs to assess the framework and whether it is sufficient to cover all existing or potential approaches for empirical agent-based models. That evaluation should include an assessment of situations in which the most appropriate sequence of methods can be determined at the beginning of a project, based on the characteristics of the problem at hand, and when appropriate sequencing might not 843 be clear at the beginning, but needs to emerge through iteration across the various steps. Opportunities for extension of the framework might appear as methods for the parameterisation for the social networks of agents and the agent environment are refined (Fig.1). For example, the requirement for real-world data on social network might change the approach to obtaining behavioural data for individual agents, as individual behaviours might be dependent on how these individuals interact with each other. References An, L., Brown, D.G., 2008. Survival analysis in land-change science: integrating with GIS and remote sensing in Southeastern Michigan. Annals of the Association of American Geographers 98 (2), 323e344. An, L., Brown, D.G., Nassauer, J.I., Low, B., 2011. Variations in development of exurban residential landscapes: timing, location, and driving forces. Journal of Land Use Science 6 (1), 13e32. Baker, T.J., Miller, S.N., Prager, S.D., Legg, D.E., 2010. Disaggregating human population for improved land use management in Kenya. Journal of Land Use Science 5 (4), 237e257. Barnaud, C., Van Paassen, A., Trébuil, G., Promburom, T., Bousquet, F., 2010. Dealing with power games in a companion modelling process: lessons from community water management in Thailand Highlands. The Journal of Agricultural Education and Extension 16 (1), 55e74. Barreteau, O., 2003. The joint use of role-playing games and models regarding negotiation processes: characterization of associations. JASSS e The Journal of Artificial Societies and Social Simulation 6 (2). http://jasss.soc.surrey.ac.uk/6/2/ 3.html. Becker, H.S., 1958. Problems with inference and proof in participant observation. American Sociological Review 23 (6), 652e660. Berger, T., Schreinemachers, P., 2006. Creating agents and landscapes for multiagent systems from random samples. Ecology and Society 11 (2). Bohensky, E., Smajgl, A., Herr, A., 2007. Calibrating behavioural variables in agent-based models: insights from a case study in East Kalimantan, Indonesia. Paper Presented at the MODSIM 2007 International Congress on Modelling and Simulation. Brown, D.G., Robinson, D.T., 2006. Effects of heterogeneity in residential preferences on an agent-based model of Urban Sprawl. Ecology and Society 11 (2). Brown, D.G., Robinson, D.T., An, L., Nassauer, J.I., Zellner, M., Rand, W., et al., 2008. Exurbia from the bottom-up: confronting empirical challenges to characterizing complex systems. GeoForum 39 (2), 805e818. Cooke, R.M., Goossens, L.L.H.J., 2008. TU Delft expert judgment data base. Reliability Engineering & System Safety 93 (5), 657e674. Dray, A., Perez, P., Jones, N., Le Page, C., D’Aquino, P., White, I., et al., 2006. The AtollGame experience: from knowledge engineering to a computer-assisted role playing game. JASSS e The Journal of Artificial Societies and Social Simulation 9 (1), 6. http://jasss.soc.surrey.ac.uk/9/1/6.html. Dung, L.C., Hoanh, C.T., Le Page, C., Bousquet, F., Gajaseni, N., 2009. Facilitating dialogue between aquaculture and agriculture: lessons from role-playing games with farmers in the Mekong Delta. Vietnam Water Policy 11 (S1), 80e93. Grimm, V., Berger, U., Bastiansen, F., Eliassen, S., Ginot, V., Giske, J., et al., 2006. A standard protocol for describing individual-based and agent-based models. Ecological Modelling 198 (1e2), 115e126. Grimm, V., Berger, U., DeAngelis, D.L., Polhill, J.G., Giske, J., Railsback, S.F., 2010. The ODD protocol: a review and first update. Ecological Modelling 221 (23), 2760e2768. Grimm, V., Revilla, E., Berger, U., Jeltsch, F., Mooij, W.M., Railsback, S.F., et al., 2005. Pattern-oriented modeling of agent-based complex systems: lessons from ecology. Science 310 (5750), 987e991. Harrison, G., List, J., 2004. Field experiments. Journal of Economic Literature 42 (4), 1009e1055. Huigen, M.G.A., 2004. First principles of the MameLuke multi-actor modelling framework for land use change, illustrated with a Philippine case study. Journal of Environmental Management 72 (1e2), 5e21. Huigen, M.G.A., Overmars, K.P., de Groot, W.T., 2006. Multiactor modeling of settling decisions and behavior in the San Mariano Watershed, the Philippines: a first application with the MameLuke framework. Ecology and Society 11 (2). Janssen, M., Ostrom, E., 2006. Empirically based, agent-based models. Ecology and Society 11 (2). http://www.ecologyandsociety.org/vol11/iss32/art37/ Art. 37. Jupp, V., 2006. The Sage Dictionary of Social Research Methods. Sage Publications, London. Kaplan, G.A., 2004. What’s wrong with social epidemiology, and how can we make it better? Epidemiologic Reviews 26, 124e135. Luna, F., Perrone, A., 2002. Agent-based Modelling in Economics and Finance: Simulations in SWARM. Kluwer, Massachusetts, Dordrecht. Martin, L.R., 1977. Survey of Agriculture Economics Literature: Quantitative Methods in Agricultural Economics, 1940’se1970’s, vol. 4. University of Minnesota Press, Minneapolis. Mathevet, R., Bousquet, F., Le Page, C., Antona, M., 2003. Agent-based simulations of interactions between duck population, farming decisions and leasing of hunting rights in the Camargue (Southern France). Ecological Modelling 165 (2e3),107e126. Matthews, R., Gilbert, N., Roach, A., Polhill, J., Gotts, N., 2007. Agent-based land-use models: a review of applications. Landscape Ecology 22 (10), 1447e1459. 844 A. Smajgl et al. / Environmental Modelling & Software 26 (2011) 837e844 Mennis, J., Hultgren, T., 2006. Intelligent dasymetric mapping and its application to areal interpolation. Cartography and Geographic Information Science 33 (3), 179e194. Moss, S., 2008. Alternative approaches to the empirical validation of agent-based models. Journal of Artificial Societies and Social Simulation 11 (1), 5. http:// jasss.soc.surrey.ac.uk/11/1/5.html. Nichols, P., 1991. Social Survey Methods: A Fieldguide for Development Workers. Oxfam, Oxford. Parker, D.C., Brown, D., Polhill, J., Deadman, P.J., Manson, S.M., 2008a. Illustrating a new ‘conceptual design pattern’ for agent-based models of land use via five case studies e the Mr. Potatohead framework. In: López Paredes, A., Hernández Iglesias, C. (Eds.), Agent-Based Modelling in Natural Resource Management. Pearson Education, Upper Saddle River, New Jersey, pp. 23e51. Parker, D.C., Entwisle, B., Rindfuss, R.R., Vanwey, L.K., Manson, S.M., Moran, E., et al., 2008b. Case studies, cross-site comparisons, and the challenge of generalization: comparing agent-based models of land-use change in frontier regions. Journal of Land Use Science 3 (1), 41e72. Parker, D.C., Manson, M.S., Janssen, M.A., Hoffmann, M.J., Deadman, P., 2003. Multiagent system models for the simulation of land-use and land-cover change: a review. Annals of the Association of American Geographers 93 (2). Patzer, G.L., 1996. Experiment-Research Methodology in Marketing: Types and Applications. Quorum Books, Westport, CT. Rees, P., Martin, D., Williamson, P., 2002. The Census Data System. John Wiley and Sons, New York. Robinson, D.T., Brown, D.G., 2009. Evaluating the effects of land-use development policies on ex-urban forest cover: an integrated agent-based GIS approach. International Journal of Geographical Information Science 23 (9), 1211e1232. Robinson, D.T., Brown, D.G., Schreinemachers, P., Janssen, M.A., Huigen, M., Wittmer, H., et al., 2007. Comparison of empirical methods for building agentbased models in land use science. Journal of Land Use Science 2 (1), 31e55. Rousseeuw, P.J., 1987. Silhouettes: a graphical aid to the interpretation and validation of cluster analysis. Journal of Computational and Applied Mathematics 20, 53e65. Smajgl, A., Bohensky, E., Bohnet, I., 2007. Capturing heterogeneity in empirical agent-based models: a guideline. Paper Presented at the Fifth ESSA Conference, Toulouse. Smajgl, A., Carlin, G., Bohensky, E., House, A., Kurnia, A.S., Butler, J., et al., 2009a. Assessing Impacts of Fuel Subsidy Decisions on Poverty and Fish Catch in Central Java, Indonesia: An Agent-based Analysis. CSIRO Sustainable Ecosystems, Townsville. Smajgl, A., Carlin, G., House, A., Kurnia, A.S., Butler, J., Sugiyanto, C., et al., 2009b. Design Document for Agent-based Model SimPaSI Jawa Tengah. CSIRO Sustainable Ecosystems, Townsville. http://www.cse.csiro.au/downloads/ DesignDoc_JaTeng.pdf. Smajgl, A., Carlin, G., House, A., Pambudhi, F., Butler, J., Bohensky, E., et al., 2009c. Simulating Pathways to Sustainability in Indonesia: Design Document for Agent-based Model SimPaSI East Kalimantan. CSIRO Sustainable Ecosystems, Townsville. http://www.cse.csiro.au/downloads/DesignDoc_JaTeng.pdf. Smajgl, A., Carlin, G., Pambudhi, F., Bohensky, E., House, A., Butler, J., et al., 2009d. Assessing Impacts of Fuel Subsidy Decisions on Poverty and Forest Depletion in East Kalimantan, Indonesia: An Agent-based Analysis, vol. 2. CSIRO. Smajgl, A., Heckbert, S., Ward, J., Straton, A., 2009e. Simulating impacts of water trading in an institutional perspective. Environmental Modelling and Software 24, 191e201. Smajgl, A., Heckbert, S., Bohnet, I., Carlin, G., Hartcher, M., McIvor, J., 2008. Simulating the Grazing System in the Bowen Broken Catchment: An Agent-based Modelling Approach. CSIRO, Townsville. Smajgl, A., House, A., Butler, J., 2011. Implications of ecological data constraints for integrated policy and livelihoods modelling: an example from East Kalimantan, Indonesia. Ecological Modelling 222, 888e896. Trébuil, G., Kam, S.P., TurkelBoom, F., Shinawatra, B., Teng, P.S., Kropff, M.J., et al., 1997. Systems diagnoses at field, farm and watershed levels in diversifying upland agroecosystems: towards comprehensive solutions to farmers’ problems. In: Applications of Systems Approaches at the Farm and Regional Levels. Kluwer Academic, The Hague. 99e114. Valbuena, D., Verburg, P.H., Bregt, A.K., 2008. A method to define a typology for agent-based analysis in regional land-use research. Agriculture, Ecosystems & Environment 128 (1e2), 27e36. Valbuena, D., Verburg, P.H., Bregt, A.K., Ligtenberg, A., 2010a. An agent-based approach to model land-use change at a regional scale. Landscape Ecology 25, 185e199. Valbuena, D., Verburg, P.H., Veldkamp, A., Bregt, A.K., Ligtenberg, A., 2010b. Effects of farmers’ decisions on landscape structure of a Dutch rural region: an agentbased approach. Landscape Urban Planning 97, 98e110. Wu, B.M., Birkin, M.H., Rees, P.H., 2008. A spatial microsimulation model with student agents. Computers, Environment & Urban Systems 32 (6), 415e488. Zellner, M.L., Page, S.E., Rand, W., Brown, D.G., Robinson, D.T., Nassauer, J., et al., 2009. The emergence of zoning policy games in exurban jurisdictions: informing collective action theory. Land Use Policy 26 (2), 256e367.