* Abstract

Using ethnography to build agent-based models may result in more empirically grounded simulations. Our study on innovation practice and culture in the Westland horticulture sector served to explore what information and data from ethnographic analysis could be used in models and how. MAIA, a framework for agent-based model development of social systems, is our starting point for structuring and translating said knowledge into a model. The data that was collected through an ethnographic process served as input to the agent-based model. We also used the theoretical analysis performed on the data to define outcome variables for the simulation. We conclude by proposing an initial methodology that describes the use of ethnography in modelling.

Ethnography, Institutional Analysis, Survey, Qualitative Data, MAIA, Conceptual Modelling

* Introduction

Building empirically-grounded artificial societies of agents requires qualitative and quantitative data to inform individual behaviour and reasoning, and document macro level emerging patterns (Robinson et al. 2007). While quantitative data can be collected through surveys, literature and other available sources, gathering qualitative data to design the behaviour of the agents, their decision making process and their forms of interaction is not a straight-forward task (Janssen & Ostrom 2006). Likewise, macro-level data for model validation requires theoretical analysis about the system that is being modelled (Robinson et al. 2007).

Modellers commonly use behavioural and social theories, and desk research to cover the qualitative aspects of agent-based models. They may also use surveys and statistical analysis to understand the decision making behaviour of individuals (Sanchez & Lucas 2002; Dia 2002).

One field of research that can also be used to collect data for agent-based models is ethnography (Bharwani 2004). Ethnography is a research method covering many approaches in anthropology. The data is gathered through interviews and field surveys which are then 'coded'[1] for theoretical analysis. The collected data is a rich set for understanding human behaviour and interaction which is also a good source to build artificial humans or agents. Furthermore, the theoretical analysis that is performed on ethnographic data could be a good source of macro level data for model validation by observing whether the same mechanism and patterns concluded from the analysis result from the simulation (Robinson et al. 2007).

Since ethnography provides a rich set of data about the system and its entities, we anticipate it can be used to make richer agent-based models populating them with empirically grounded data. However, this data, although coded for theoretical analysis, is difficult to interpret and decompose in order to build agents and their behavioural rules. Ethnographic data is normally in textual format obtained from interviews, fieldwork, participant observation or formal documents (Yang & Gilbert 2008).

The difficulty in making use of ethnographic information for agent-based modelling and simulation (ABMS) is due to the fact, that in qualitative ethnographic research the interviewees are normally allowed to talk about their concerns in an open manner, which may lead to an overload of information that may also be immensely rich and diverse in terms of content. In addition, the researcher and the interviewees each have their own world-view, which leads to bias, as abstraction and generalization is required to arrive at specifications of behaviour and characteristics suitable for building agent-based models.

The most complete research in the intersection between ABMS and Ethnography is Bharwani (2004). Bharwani (2004) provides a detailed procedure for the fieldwork process which describes how ethnographic data is collected and formalized. Bharwani (2004) used knowledge engineering techniques in the process, allowing a continued engagement with the interviewees. She designed a specific ontology (i.e., architecture) for her particular domain namely, Agro-Climatic systems, to decompose the ethnographic information into a model. Yang and Gilbert (2008) discuss the differences and similarities between ethnographic data and ABMS and propose recommendations for modellers when using ethnographic data. They emphasize on the requirement for computer-aided qualitative analysis to manage and structure the data. Another requirement indicated by them is a model of data to represent relationships among actors (Yang & Gilbert 2008).

There are also case specific examples of using qualitative data in agent- based models. Geller and Moss (2008) present a model of solidarity networks in Afghanistan, informing agents' structures, behaviour and cognition by qualitative data. They use an evidence-based approach following rules according to which agents behaviours are directly drawn from empirical studies. Moore et al. (2009) use a combination of ethnography and ABMS to study psychostimulant use and related harms. They also indicate the difficulty in generalizing ethnographic information to build agent-based models. They built a model called SimAmph as a shared ontology to combine ethnography and ABMS for their particular case, which proved to be useful in making the connection between the two domains as well as in facilitating collaborative model development and analysis.

Thus, from the literature, it appears that a shared ontology or a conceptual framework is one of the main requirements for generalizing and structuring qualitative information, especially ethnographic data for ABMS. To address this requirement, in this research, we use an ABMS framework called MAIA (Ghorbani et al. 2013) which provides a shared ontology for social systems, covering a diversity of social, institutional, physical and operational concepts that are required for building agent-based models. Using MAIA as a template of required concepts may help collect and structure ethnographic data for building agent-based models. Therefore, in this research, we explore this possibility by using this modelling framework to structure ethnographic data collected from interviews, fieldwork and formal documents to build an agent-based model. To underpin this possibility, we use a case study on innovation practices in the Dutch horticulture sector.

The remainder of this paper is as follows. In Section 2, we give a brief overview on ethnography and introduce the MAIA framework. In Section 3, we introduce the horticulture case study. In Section 4, we explain the methodological process of integrating ethnographic processes into ABMS. In Section 5, we discuss the lesson learnt from this process and analyse our methodological process. Finally, we conclude in Section 6.

* Background

The goal of this research is to propose a methodology for using ethnography to build agent-based models. In this section, we will first explain ethnography. Then, we will introduce the MAIA framework, which will be used as the tool for this methodological process.


Ethnography is a field of science that spans many methods and schools of approaches in anthropology. The power of ethnographic research is that real people are studied at the level of small communities/groups or individuals, and at the societal level, while the mutual interaction is also considered. This qualitative research aims to address complex phenomena by analysing and interpreting the system from the participants' point of view. Ethnography is often exploratory in nature, using observations to construct the analysis from 'bottom-up'. Together, this appears to be what is needed for developing agent-based models, in order to characterize the interaction of the individual and the system:
Ethnographic research can range from a realist perspective in which behaviour is observed to a constructivist perspective where understanding is socially constructed by the researcher and subjects. Research can range from an objectivist account of fixed, observable behaviours to an interpretivist narrative describing "the interplay of individual agency and social structure." Critical theory researchers address "issues of power within the researcher- researched relationships and the links between knowledge and power (Ybema et al. 2010).

In ethnography there are several types of methodologies, which can broadly be categorized as either inductive or deductive. An inductive approach to ethnography formulates theories from the 'bottom-up' rather than from the 'top-down'. This means that the researcher starts by observing the community and by looking for repeated patterns of behaviour. If certain themes continue to appear, the researcher can develop a tentative hypothesis that is then verified and which may be turned into a theory. This may require the collection of more corroborating data from other communities within the same society[2]. 'Grounded theory' is an inductive method of analysis commonly applied in ethnography to help scientists generate theories (Corbin & Strauss 2008). Unlike other theories, grounded theory does not start by hypotheses for social behaviour but concludes with them. The grounded theory approach is an iterative process where the analysis of the data may raise new questions that stimulate new data collection (Neumann 2014). While this describes inductive research, some anthropologists also take the deductive approach, using prefixed questionnaires, hypothesis, quantitative data and statistics etc.

The inductive approach is more flexible, however, when it comes to addressing human societies, as it helps the researchers let go of their own preconceived (and often culturally biased) ideas of what the society they are studying is like. While the inductive approach is still used in cultural anthropology today, currently this theory has shifted from 'start fieldwork and wait for answers' to 'start field work with a few general questions to answer'. This would provide enough frameworks to focus the research, but would leave the questions general enough to allow for the flexibility that studying human culture needs. Some methods play a central role in this inductive approach:
  • Open-ended and semi-structured interviewing: semi-structured interviews are open-ended, but the interview is guided by a list of topics[3]. Such interviews allow discussions that have not been prepared for, while the list guides the discussion. Together, this renders the interview to be both efficient and effective.
  • Participant observation and field work: this method is the foundation of cultural anthropology, and entails the residence of the researcher in a field setting, where the observer blends into the daily life of the people and may closely monitor their activities.

The data produced in ethnography is a combination of written interviews, recordings, documents and personal notes. Structuring, analysing, interpreting and presenting the data is therefore an important step. The richness of data from ethnographic studies can be organized in programs like Atlas.ti [4]. In the analysis process, the next step is to generate categories, themes and patterns from the organized data. The processed and organized data can then be inspected and interpreted, and theories can be used to frame and analyse the data to elucidate patterns and give meaning and explanation to the data.

The MAIA Framework

MAIA (Modelling Agent systems based on Institutional Analysis) is a modelling framework that structures and conceptualizes an agent-based model in a high level modelling language (Ghorbani et al. 2013). The concepts in the framework are a formalization of the Institutional Analysis and Development (IAD) framework of Elinor Ostrom (2009), extended with concepts from other social science theories (Structuration (Giddens 1984), Social mechanisms (Hedström & Swedberg 1996) and Actor-centered institutionalism (Scharpf 1997).

MAIA has been designed to support the participatory development of agent-based simulations. Since its concepts are taken from various theories, this modelling framework can be used by inexperienced modellers and those who are not familiar with programming skills. Furthermore, an online tool[5] supports the conceptualization process of agent-based models. In this tool, the MAIA model (i.e., the conceptual model developed using MAIA) is observable and traceable through cards and diagrams and can therefore be used for communication with domain experts and problem owners for concept verification. MAIA has been evaluated in several projects (e.g., transition in consumer lighting, the wood-fuel market, e-waste recycling sector, and manure-based bio-gas energy system) (Ghorbani 2013).

The framework provides a guideline to arrive at a comprehensive overview if not model of a social system by defining five interrelated structures that group related concepts:
  1. In the Collective structure actors are defined as agents by capturing their characteristics and decision criteria based on their perceptions and goals.
  2. The Constitutional structure defines roles and institutions. Actors can take multiple roles in social systems. These roles are formalized as unique sets of objectives and capabilities. Roles allow efficient modelling of heterogeneous agents who perform similar tasks. Institutions are defined as the set of rules devised to organize repetitive activities and shape human interaction (Ostrom 1991). In MAIA, institutions are defined using "ADICO grammar of institutions" proposed by Crawford and Ostrom (1995). In ADICO, 'A' is the attribute or the actor who is the subject of the institution, 'D' is the deontic type of the institution (prohibition, obligation, permission), 'I' is the aim of the institution, 'C' is the condition under which the institutional statement holds and 'O' is the sanction for non-compliance to the institution.
  3. The Physical structure is the non-social environment that the agents are embedded in. Its building blocks are physical components.
  4. The Operational structure is viewed as an action arena where different situations take place, in which participants interact as they are affected by the environment. These produce outcomes that in turn affect the environment. The agents, influenced by the social and physical setting of the system, perform their actions in the action arena. The action arena contains all the entity actions, ordered by plans, which are in turn ordered by action situations.
  5. The Evaluative structure provides concepts with the help of which the modeller can indicate what patterns of interaction, evaluation, and outcomes she is interested in. The modeller identifies those variables that can serve as indicators for model validity (is it sufficiently realistic?) and model usability (will its implementation help me to explore the question(s) I set out to address?).
Figure 2 at the end of this article shows the concepts in MAIA. Extensive specification of MAIA can be found in Ghorbani et al. (2013).

* Case Study: Horticulture Innovation

The key objective of our study of the horticulture sector is to elucidate the effects social institutions have on innovation practices in Westland, a region that is home to about 70% of all greenhouse acreage in the Netherlands.

The horticulture sector in the Netherlands at large is facing economic difficulties, which have become more severe since the crisis begun in 2008 (Schrauwen 2012). The dominant presence of innovation strategies that target cost-reduction and volume-increase brings down the cost of products. They fail to bring the growers sustained benefits however, which causes serious problems in the sector. Due to mechanisms in the market, the growers only benefit financially from their innovations for a relatively short period. When their innovations spread in the sector, the market price of their products drops rapidly, because it is subject to fierce price competition, a characteristic of 'cost leadership' market segments. Few growers attempt to increase the value of their products by developing niche product-market combinations, or expand their activities in the value-chain by developing new channels to the market to capture a greater share of the value created between growers and consumers. Such innovation strategies beyond process innovation for unit cost-price reduction are less popular in the sector, despite their potential to counteract the effect of downward spiralling prices in competitive markets.

The goal of this study is to investigate the innovation practices in the Westland horticulture sector to obtain an understanding on how this observed pattern of innovation has emerged and how the underlying behaviour of growers is shaped and maintained. We use grounded theory as our methodology to perform ethnographic field work. Besides using MAIA for data collection and model development, we perform a theoretical analysis using the Bathtub model of Coleman (1986) and several other theories (see Schrauwen 2012). The rationale for adopting a fieldwork approach (rooted in cultural anthropology) is that the organizations and innovation practices are socially embedded, and can be studied as such. Furthermore, the Westland is said to be home to Westlanders who share a common identity with respect to social and business culture, which is shaped by and has shaped their core business for centuries (Kasmire et al. 2013).

* The Modelling Process

The purpose of our methodological practice is to guide the collection of data for building an agent-based model using an ethnographic approach. This process is divided into two parts. The first part uses MAIA as a template for information collection, which includes field observation, interviews and the study of formal documents. For each of these methods, we make use of the MAIA framework to semi-structure the data collection process. The second part uses the collected information to build a MAIA model.

Collecting data using MAIA

Structuring interviews with MAIA

In inductive ethnographic research, interviews are normally semi-structured. Therefore, it is common practice, to develop a general structure or guideline for the interviews, to ascertain that at least all relevant aspects are addressed. We use MAIA as the general structure for the interviews in order to cover all the information required to build an agent-based model. At the same time, we leave the questions open-ended, so that the interviewees feel free to talk about what may seem relevant to them.

The interviews were conducted with various stakeholders in the Westland horticulture sector (Schrauwen 2012):
  • Experts: Experts were interviewed to gain better insight into the sector as a whole and also to evaluate the assumptions that were being made during the analysis and modelling phase.
  • Growers: Fifteen growers were visited at their organization. Each interview took between two to five hours. The growers were either contacted directly or introduced by other respondents.
  • Organizations: The bank, churches, educational institutes, municipality, LTO GlasKracht and supermarket were the other actors interviewed in order to find out their influence on the social network of growers, their individual capital and investment, and their knowledge and background.

The concepts that were used to structure the interviews and direct the questions are:

     — Collective Structure

  • Agent Decisions: What decisions do the growers make regarding their innovation practices? The growers are allowed to talk about their decisions freely without being forced to explain how they make those decisions[6].
  • Agent personal value: The growers are asked about what they care about most when they are making those decisions.
  • Related Agents: During the interviews, the growers are asked about other social entities they may be interacting with. These can be individual actors, such as other growers, or composite actors (i.e., organizational type) such as the bank, or the municipality.
     — Operational Structure
  • Actions and Plans: The growers are asked about what their general activities are and how often they perform these activities. In this case study, they were asked about their daily, monthly and yearly activities. If each of these practices constitutes a process, they were also asked about the events that take place in that process. For example, if a grower decides to apply for a subsidy, what actions does he have to perform during the application process?
     — Constitutional Structure
  • Roles: The growers are implicitly asked about the different roles they take in their activities. This is not a straightforward question, but one that would rather need to be extracted from the explanations the growers provide. For example, a grower explains that he has to be a client of the bank to apply for a particular subsidy or he emphasizes that he would only expand his greenhouse, if he has a child who is willing to take over. From these remarks we can identify 'bank client' and 'being a father' as two of the roles, the growers may assume under certain condition.
  • Formal Institutions: While asking about the operational activities and decisions, the subjects are also asked about the formal procedures, rules and regulations they need to go through. This is later used to collect relevant institutional documents.
     — Physical Structure
  • Physical Components: During the interviews, the subjects are asked about the physical entities they use in their activities, the ones they own or the ones that influence their actions. It is important to ask about this aspect; while the interviewee is talking about the activities he performs in order to limit the information to what is relevant.
The interviews are recorded and coded in Atlas.ti for later analysis.
Using MAIA for field observation

During field observation, it is important to identify the relevant properties of the entities (i.e., agents and physical components) that are addressed during the interviews. The composition of the physical entities and their connections may be observed in the field and defined as physical components in the physical structure of MAIA. Thus, in a fashion similar to setting up the general structure for the semi-structured interviews, the MAIA structures can be used as a template for collecting data during field observation.
Using MAIA for studying formal documents

The formal documents are collected according to the information provided by the subjects. To collect the right information for modelling institutions, the ADICO structure (see Section Background) is used as the template.

Building a MAIA model

Upon completion of the previous steps, the collected data is used to build an agent-based model. This process is conducted by extracting relevant information from the data by using the MAIA framework. Again, we look at the structures one-by-one to clarify the process [7].
Collective Structure

The interviewed subjects can be defined as agent-types. Each subject can be defined as one separate agent-type if the simulation is limited to the people interviewed; alternatively, one may group the agents according to some criterion and use each category to define a separate agent-type. In the greenhouse case, the 15 growers that were interviewed were divided into five categories distinguished by their stated priorities, their physical assets and characteristics. The first category is the niche growers whose greenhouse is relatively small in size and whose innovation activities are mainly marketing- and product-oriented. The other four categories are large bulk growers, the innovative bulk growers, moderate bulk growers and shop growers (see Schrauwen 2012).

Agents in the simulation are not limited to the interviewees; there may also be social entities that were addressed during the interviews. For example, the European Union was a social entity addressed by the growers, who influences their innovation strategies. This entity is, therefore, also defined as an agent in the simulation.

From the qualitative data, whether in the form of field observation or interview, the properties, personal values, intrinsic behaviours and decision-making of the actors are extracted to build the agents in the model.
Constitutional Structure

The main aspect of the constitutional structure is the institutions. These can be formal institutions extracted from legal documents, or informal institutions, namely, norms of behaviour and shared strategies extracted from the interviews or field observations. The patterns of behaviour observed from interviews can be the result of rules imposed by the society. These are defined as norms or shared strategies. If the rule of behaviour contains an obligation or prohibition by definition, the rule is considered to be a norm. If the actors perform the same routine without any obligation from the system, that routine can be considered as a shared strategy. All the formal and informal institutions are modelled as ADICO statements as defined in Section Background. Table 1 shows some of the institutions extracted from the interviews and legal documents.

Table 1: Some of the identified institutions in greenhouse case study.

# A D I C O      Type
1.a Growers may get a subsidy from the EU for max 50% of an investment if they invest in an accepted innovation and follow the rules 1.b 1.c rule
1.b Growers must join one of the 6 sales cooperations if they want to get the GMO subsidy from the EU 2 rule
1.c Growers may not market under their own brand if they want to get the GMO subsidy from the EU 2 rule
2 EU may fine the grower if growers don't follow the rules attached to the subsidy rule
3 Rabobank may not increase the interest on loans when growers are a less optimal financial situation rule
4 Growers copy the successful innovations of their colleagues if the colleague is more successful norm
5 Growers cooperate together with other growers If performing similar practice shared strategy
6 Growers adopt an innovation when it has shown to be working at other greenhouses /test centres shared strategy
7 Growers invest and modernize in the organization if there is a successor shared strategy

Physical Structure

Similar to building agents, the physical entities that are addressed by the interviewees are extracted from the text and defined as physical components in the MAIA model. These include energy, greenhouse and machinery (i.e., the innovative technology they adopt). The properties of these components are identified through field observation in addition to interviews. For example, during field work, it became clear that two properties, namely, the size of the greenhouses and their type of crops, mainly distinguish growers from each other.
Operational Structure

The events that were described by the interviewees are defined as actions in MAIA. The condition for performing those actions and the outcomes of the actions should be extracted from the descriptions the subjects provide. The described sequence of actions helps to define agent plans in MAIA. Finally, the modeller has to make a decision about the time loop and the actions that take place per tick. For this study, we decided that in each tick, seven action situations take place according to the following sequence:
  • Daily life: In this action situation, the intrinsic capabilities of actors take place: being born, die, have a child, learn and start relationships.
  • Cooperating: Within the action situation of cooperating, growers can group together and make a joint decision on investments in innovations. Also, knowledge, norms and values are shared amongst growers that are cooperating, adding up to the social capital of the growers.
  • GMO: In this action situation, growers request GMO (Gezamenlijke Markt Ordening - collective market structuration) subsidy where they may recover half of the investments. GMO applications can either be accepted or rejected. Previous subsidy receivers may also be punished in this action situation, based on their previous actions.
  • Loan: In this action situation, the grower can apply for a loan. He has to pay back his loan and report his money level to the bank, who may take over, when the grower is in trouble.
  • Innovating: In the innovation situation, the decisions are made by the growers to invest in one of the categories of innovations. They invest their money in that innovation, while adopting a new physical component (i.e., technology) in their greenhouse with specific characteristics.
  • Cultivation: In the cultivation situation, all horticulture-related activities are performed such as cultivation, employing technologies, and increasing efficiency. The investments of the previous round of innovations affect the cultivation process and produce outcomes, in terms of products, efficiency, use of inputs, etcetera. Also, the money level is checked and reported to the bank (if the grower is a member).
  • Selling: In the selling situation, growers calculate the costs and value of their products and calculate a market price. They sell their products to the merchandisers. Products are exchanged with money.
Evaluative Structure

To build the evaluative structure of MAIA, not only the data collected was used, but also the anthropological analysis. We defined a set of variables that can be used to measure and study the possible emergent system elements from the simulation according to this analysis.

The theoretical analysis showed that a phenomenon called 'isomorphism' steers companies towards the same characteristics which gives rise to similar innovation practices that are not effective in the long run and may even harm the sector. To explore this phenomenon in the simulation, we defined the variable 'homogenization' to calculate the variation in innovation types. This value would be measured through time. The correlation between subsidies and this variable is also identified as a parameter of interest according to the ethnographic analysis.

One other issue in the analysis was 'decreasing product value'. Many products, especially bulk products, are sold with little margin. This means that the income flowing back to the grower is at risk of being less than cost, which decreases their capital. With just one innovation not giving good returns, this may put them in danger. This may even cause bankruptcy. Therefore, another variable to keep track of in the simulation is the developments of product value (i.e., product price) in relation to time and different innovation types.

The sector's sustainability is another point of interest in the study. This issue stands on three different pillars, namely, economical, ecological and social. To experiment with these pillars in the simulation, for the economical part, the ratio between product value and bankruptcy is calculated in relation to subsidies, loans and time. For the ecological aspect, the relation between water, energy and nutrient, and amount and value of products is defined as a metric. Finally, to track the social influence, we define two variables: social capital and bankruptcy.

In this section, we presented an overview of the process of ethnographic data collection and analysis used for conceptualizing an agent-based model of the horticulture sector. We explained how MAIA concepts can be used to inform data collection, and to build an agent-based model. In the next section, we will generalize this methodological procedure, to make it applicable to other social studies.

Generalizing the Process

Figure 1
Figure 1. The MAIA framework can be used to semi-structure ethnographic data collection as well as shaping the collected data into an agent-based model.

Figure 1 shows the general process of using ethnographic data to build an agent-based model using MAIA. Some concepts in the MAIA structures, as illustrated on the left side of the figure, are primarily used to semi-structure the data collection process. The collected data is then decomposed into an agent-based model, again, using the MAIA structures.

As Figure 1 shows, there is a cycle between the ethnographic research and the building of a MAIA model. Although semi-structuring data collection minimizes the need to redo interviews, it may still be required to collect further information for the model. This would especially hold for field observations and document collection.

Besides building the conceptual model, the ethnographic data is also used to perform theoretical analysis. Not only can this analysis be used to further enrich the model, specifically in the evaluative structure (see previous section), it is also used to draw conclusions. These conclusions can be used independently or in combination with the simulation results. Some sort of triangulation can thus be completed, comparing the social analysis with the dynamics generated by running the model. What may be an issue here, however, is that the same input data is used for both methods, so they are not completely independent.

* Discussion

Building an agent-based model requires both quantitative and qualitative data. Although much of the information can be represented in the form of numeric values, the actual context of the model which shows the order of the events, and how agents make decisions and interact, requires qualitative information. Ethnography can provide rich data for building agent-based models both at micro and macro levels. However, it needs structure and interpretation to be actually applicable to this simulation approach (Yang & Gilbert 2008). In this paper we presented MAIA as a tool to collect and structure ethnographic data for ABMS. The process of building an agent-based model for the horticulture sector helped us to identify several benefits of using this tool.

First, the MAIA framework ensures consistency and coherence between the features extracted from the ethnographic process. Since MAIA is constructed as software meta-model, its soundness, completeness and parsimony have been verified (Ghorbani 2013). Therefore, the modeller can be confident that the collected and structured data is by default consistent in the model.

Second, as Dey (2003) indicates, analysing qualitative data also involves an abstraction process which may not be a straightforward task given the immense amount of details provided by ethnography which mostly concerns individuals. Since MAIA is an abstract template or 'ontology' for a set of concepts, it proved to be highly instrumental for facilitating and documenting this abstraction process.

Third, another contribution of MAIA in making use of ethnographic data is that it helps to identify the normative aspects of the system. The insights people provide about their view of the world through interviews are not based on external reality but are culturally generated and emergent. With the ADICO statements in MAIA, the modeller can extract the norms and shared strategies from the interviews in order to add a cultural/institutional dimension to the simulation.

Fourth, an important contribution of using MAIA is that not only the collected ethnographic data can be used to build an agent-based model; the theoretical analysis performed on the data is also put to use. The theoretical ethnographic analysis helps define the variables that measure the outcomes of the simulation. These variables are covered in the evaluative structure of MAIA. Therefore, besides informing agent behaviour, the methodological process introduced in this paper can help measure the possible outcomes of interest, i.e., macro-level patterns for the simulation.

Fifth, when an ethnographic researcher uses MAIA, her activities become more structured and tractable. We anticipate this will facilitate the interpretation and discussion of field research, and lead to a growing body of empirically grounded information that can be re-used for modelling and research studies.

Finally, linking the body-of-knowledge of anthropology and agent-based modelling of social systems may be mutually beneficial. We believe, the proposed method supports non-computing anthropologists in building agent-based models in order to complement their research methods. To explore the feasibility of this claim, an anthropologist performed the whole process starting from the ethnographic fieldwork to the development of the conceptual model. We observed that MAIA can indeed bring ABMS within the reach of anthropologists who even have no familiarity with modelling.

Indeed, to build agent-based models from such data, a major difficulty is the step from a limited number of individuals interviewed to the creation of a whole society. The stories and decision-making are usually personal and related to personal incidents; it is hard to draw certain 'types' of agents from that, because those coincidental incidents in life have a large influence. While estimating the percentages of the type of people forming the society is hard, in the eventual ABM, these can become parameters for variation.

Finally, it is important to emphasize that the structuring of collected data although highly facilitated with MAIA, still depends on the creativity of the modeller. There are many choices and interpretations that the modeller has to make to transform qualitative data into an agent-based model. When MAIA is used, however, there will be both a unambiguous language to communicate about the decision taken, and a traceable track record of how the researcher arrived from empirical data to interpreted model results and model.

* Conclusion

Managing and structuring data, especially qualitative, is a major challenge for agent-based modelling. This research presented a method to effectively use ethnographic data for building agent-based models.

We used the MAIA framework to semi-structure the data collection procedure and later on used the same framework to decompose the information and build a conceptual agent-based model. The conceptual model is then used to produce running simulations.

Although MAIA facilitated the structuring of qualitative information, another phase of data collection is required, namely one to complete the quantitative aspects of the simulation. This phase is not yet supported by the methodological process presented here. Therefore, the next step of this research is to extend the MAIA framework to support the quantitative data collection process.

* Acknowledgements

This work was supported by the European Regional Development Fund, Duurzame Greenport Westland-Oostland Task Force (http://greenportduurzaam.nl).

* Notes

1 In ethnography, coding is the process of organizing the collected data for analysis. 2 See http://www.helium.com/items/948634-explaining-the-inductive-approach- cultural-anthropology 3 Unstructured interviews are conversations that can take place anywhere and anytime; structured interviews are completed while strictly adhering to the predefined interview protocol. 4 http://www.atlasti.com/ 5 More about MAIA and the modelling environment can be found at maia.tudelft.nl. 6 To setup interviews and questions using the MAIA framework, it may also be helpful to start conversations with a life history/narrative of the interviewee before diving into specific questions. While such conversations may take hours, such 'off the record' conversations can be very helpful. MAIA may help to strike the right balance between such talks and more to-the-point conversation and interview. 7 The ethnographic field work, the full MAIA model and the ethnographic analysis can be found in (Schrauwen, 2012).

* Appendix

Figure 2
Figure 2. The UML class diagram for the MAIA meta-model (Ghorbani et al. 2013)

* References

BHARWANI, S. (2004). Adaptive knowledge dynamics and emergent artificial societies: Ethnographically based multi-agent simulations of behavioral adaptation in agro-climatic systems. University of Kent, UK.

COLEMAN, J. (1986). Social theory, social research, and a theory of action. American journal of Sociology, pages 1309–1335. [doi:10.1086/228423]

CORBIN, J. & Strauss, A. (2008). Basics of qualitative research: Techniques and procedures for developing grounded theory. Sage Publications, Inc.

CRAWFORD, S. & Ostrom, E. (1995). A grammar of institutions. American Political Science Review, pages 582–600. [doi:10.2307/2082975]

DEY, I. (2003). Grounded theory, Qualitative Research Practice, pages 80–93.

DIA, H. (2002). An agent-based approach to modelling driver route choice behaviour under the influence of real-time information. Transportation Research Part C: Emerging Technologies, 10(5), 331–349. [doi:10.1016/S0968-090X(02)00025-6]

GELLER, A. & Moss, S. (2008). Growing qawm: An evidence-driven declarative model of afghan power structures. Advances in Complex Systems, 11(02), 321–335. [doi:10.1142/S0219525908001659]

GHORBANI, A. (2013). Structuring Socio-technical Complexity - Modelling Agent Systems Using Institutional Analysis. PhD thesis, Delft University of Technology, The Netherlands.

GHORBANI, A., Bots, P., Dignum, V., & Dijkema, G. (2013). Maia: A framework for developing agent-based social simulations. Journal of Artificial Societies and Social Simulation, 16(2).

GIDDENS, A. (1984). The constitution of society: Outline of the theory of structuration. University of California, US.

HEDSTRÖM, P. & Swedberg, R. (1996). Social mechanisms. Acta Sociologica,39(3), 281–308. [doi:10.1177/000169939603900302]

JANSSEN, M. & Ostrom, E. (2006). Empirically based agent-based models. Ecology and Society, 11(2), 37.

KASMIRE, J., Nikolic, I., & Dijkema, G. (2013). Evolving greenhouses: An agent-based model of universal darwinism in greenhouse horticulture. Journal of Artificial Societies and Social Simulation, 16(4), 7.

MOORE, D., Dray, A., Green, R., Hudson, S. L., Jenkinson, R., Siokou, C., Perez, P., Bammer, G., Maher, L., & Dietze, P. (2009). Extending drug ethno-epidemiology using agent-based modelling. Addiction, 104(12), 1991–1997. [doi:10.1111/j.1360-0443.2009.02709.x]

NEUMANN, M. (2014). Grounded simulation. Advances in Social Simulation (pp. 351–359). Springer. [doi:10.1007/978-3-642-39829-2_31]

OSTROM, E. (1991). Crafting institutions for self-governing irrigation systems. Sai Francisco, Institute of Contemporary Studies.

OSTROM, E. (2009). Understanding institutional diversity. Princeton University Press.

ROBINSON, D. T., Brown, D. G., Parker, D. C., Schreinemachers, P., Janssen, M. A., Huigen, M., Wittmer, H., Gotts, N., Promburom, P., Irwin, E., Berger, T., Gatzweiler, F., & Barnaud, C. (2007). Comparison of empirical methods for building agent-based models in land use science. Journal of Land Use Science, 2(1), 31–55. [doi:10.1080/17474230701201349]

SANCHEZ, S. M. & Lucas, T. W. (2002). Exploring the world of agent-based simulations: simple models, complex analyses: exploring the world of agent-based simulations: simple models, complex analyses. In Proceedings of the 34th Winter simulation conference: exploring new frontiers, pp. 116–126.

SCHARPF, F. (1997). Games real actors play: Actor-centered institutionalism in policy research. Westview Press.

SCHRAUWEN, N. (2012). Innovativeness of innovation. Master's thesis, TU Delft.

YANG, L. & Gilbert, N. (2008). Getting away from numbers: Using qualitative observation for agent-based modeling. Advances in Complex Systems, 11(02), 175–185. [doi:10.1142/S0219525908001556]

YBEMA, S., Yanow, D., Wels, H., & Kamsteeg, F. (2010). Ethnography. In Mills, A., Durepos, G., & Wiebe, E. (Eds), Encyclopedia of case study research (pp. 348–352). Thousand Oaks, CA: SAGE Publications, Inc. [doi:10.4135/9781412957397.n129]