Keywords

These keywords were added by machine and not by the authors. This process is experimental and the keywords may be updated as the learning algorithm improves.

1 Introduction

Agent-based models (ABMs) can be hard to communicate. They usually include different types of agents and spatial units, each distinguished by a suite of state variables, attributes, behaviours, and environmental processes. Model results can depend strongly on how model entities are initialised, how processes and events are scheduled, the data used to drive the simulated environment, and the details of how submodels represent model processes. As a consequence, many, if not most, descriptions of ABMs are lengthy but nevertheless incomplete.

Communication is, however, part and parcel of using ABMs as a scientific tool. Models are the “Materials and Methods” used by modellers to obtain scientific results. Thus, as with any other method used in science, the very feature that renders an ABM “scientific” is replicability: that the same results can be obtained by peers if they use exactly the same materials and methods. But most published ABMs would be impossible, or at least challenging, to replicate. This situation cannot be tolerated because it undermines the credibility of ABMs as a scientific method. Moreover, because ABMs are increasingly built to support real-world decision making, model structure and assumptions need to be transparent to decision makers; otherwise the models are likely to be (justifiably) ignored, or used inappropriately.

Incomplete and inefficient communication is linked to a second challenge of agent-based modelling: the lack of a framework for designing ABMs. Current practice is that most ABMs are developed from scratch and that the choice of model structure and process representation is more or less ad hoc. Model design and formulation often reflect the disciplinary background, experience, and personal preferences of the modeller more than general principles of model design that would lead to similar model designs for similar problems.

Lack of efficiency and coherence in design and communication is to be expected for an emerging scientific approach. But agent-based modelling is no longer in its infancy: hundreds of ABMs have proven the potential of this approach, both in theory and application, in a wide array of disciplines; many research projects and programs dealing with complex systems include ABMs in their portfolio of approaches; and the issues of standardisation, communication, and transparency have increasingly been addressed over the last 10 years.

There are several examples of the drive for a general framework for ABMs. Railsback (2001) listed design concepts that are important for designing ABMs, in particular when adaptive behaviour is included. Grimm and Railsback (2005) devoted their book on individual-based modelling in ecology to making this approach more “coherent and efficient”. A series of workshops addressed the issue of model replication (e.g., Hales et al. 2003). Standards for model development (Richiardi et al. 2006) and communication were proposed (Grimm et al. 2006, 2010), and a community has been established with the mission to “foster the agent-based modelling development, communication, and dissemination for research, practice and education” (Janssen et al. 2008): the Open Agent Based Modelling Consortium (www.openabm.org).

Standardisation is an indicator for a maturing approach: the approach has been used enough to understand its strengths and weaknesses, what decisions have to be made when using it, and how standards can make its use more efficient and coherent. Standards must of course avoid restricting the approach’s flexibility and usefulness.

Here we present the ODD protocol, a standard for communicating and formulating ABMs; and pattern-oriented modelling (POM), a general strategy for tying model structure to multiple observed patterns to make model design and analysis less ad hoc. Both approaches were formulated in ecology but are relevant in any discipline using ABMs. Since both ODD and POM have been described in detail elsewhere, we focus here on what these approaches are, how they can be used, and what benefits they provide to the agent-based modeller.

2 The ODD Protocol

The ODD protocol was developed as a standard format for describing individual- and agent-based models (Grimm et al. 2006; Grimm and Railsback 2005). The acronym stands for the three blocks by which ODD’s elements are grouped: Overview, Design concepts, Details. ODD is designed to be used for all ABMs, independent of their domain, purpose, complexity, and computer implementation. The main idea of ODD is to present ABMs in a hierarchical way: an overview of model structure and processes first, with details on processes last. Between overview and details is a checklist that explains how important general concepts for designing ABMs have been taken into account. This checklist ensures that modellers make important design decisions consciously and that readers understand how and why certain design decisions were made.

Using ODD implies referring to this standard explicitly by stating: “The model description follows the ODD (Overview, Design concepts, Details) protocol for describing individual- and agent-based models (Grimm et al. 2006, 2010)” and then presenting the seven elements of ODD (Table 17.1) using exactly the given sequence and element labels. By using ODD always in this way, writers do not waste time figuring out how to organise their model description and readers avoid frustration by knowing exactly where to find information about a model. Moreover, understanding of a model is greatly facilitated by first providing an overview of its structure and processes. For most readers, this information is sufficient to grasp the overall rationale of the model and to understand its results. Readers hooked by the overview can go into more detail and see how certain processes have been represented. Finally, an ODD description should be complete enough to allow the model’s replication.

Table 17.1 The seven elements of the ODD protocol (See also the ODD webpages at www.ufz.de/oesatools/odd)
Full size table

For complex models, the Details and possibly also the Design concepts parts may be too long, so they can be moved to an electronic appendix, or the full ODD model description could be published in a separate report.

2.1 Design Concepts

The design concepts of ODD were first summarised by Railsback (2001), then discussed in detail in Grimm and Railsback (2005), and most recently updated by Grimm et al. (2010). The ten concepts are not actually needed to replicate a model, but they are important for communicating the essence of how (and why) an ABM has been designed—because traditional formats such as equations and diagrams cannot capture the essence of most ABMs. For example, the first design concept is “emergence”. Its discussion should explain whether key model results are imposed, or whether instead they emerge from agent behaviours and other processes – information that is the key to understanding the rationale of the model. Here “imposed” means that model rules and equations force the model to behave in a predictable way. For example, if a model is designed so that an event like a traffic jam occurs with a constant probability of 5%, then it is no surprise that on average the model produces traffic jams 5% of the time. If, however, traffic jams emerge from the behaviour and decisions of the agents and the roads they follow, there is no simple way to predict how often traffic jams occur: we have to run the model and see what emerges.

In many descriptions of ABMs, it is not entirely clear what the authors wanted to emerge and what they imposed. Similar “ad hoceries” often occur with the other design concepts. The following description of the concepts is adopted from Grimm et al. (2010):

  • Emergence. What key outputs of the model are modelled as emerging from the adaptive behaviour of its agents? Are there other outputs that are more tightly imposed by model rules and hence less dependent on what individuals decide to do?

  • Adaptation. What rules do agents have for changing behaviour in response to changes in themselves or their environment? Do these traits explicitly aim at increasing some measure of individual objectives or success? Or do they instead cause individuals to reproduce observed behaviours that are implicitly assumed to convey success?

  • Objectives. If adaptive behaviour is represented as explicitly seeking some objective, what is the objective and how is it measured? Examples of “objectives” are “fitness” for organisms, “utility” for economic reward in social models, or simply “success”.

  • Learning. Do individuals change their adaptive behaviour over time as a consequence of their experience? How?

  • Prediction. To make decisions, model agents often need to predict future consequences of their alternatives. What internal models are used by the agents to estimate future conditions or consequences of their decisions? What “tacit” or hidden predictions are implied in these internal models?

  • Sensing. What information (state variables of other model entities and themselves) can agents sense and consider in their adaptive decisions? Are the mechanisms by which agents obtain information modelled explicitly, or are agents simply assumed to “know” these variables?

  • Interaction. What kinds of interactions among agents are in the model? Are there direct interactions, or are the interactions indirect, e.g. via competition for a mediating resource? How do agents interact with their environment?

  • Stochasticity. What processes are modelled by assuming that they are random or partly random? Why is stochasticity used – to represent variability in a simple way, or to cause events or behaviours to occur with a specified frequency?

  • Collectives. Are there aggregations of agents that affect, and are affected by, the agents? Examples include social groups, fish schools and bird flocks, human networks and organisations, or cells constituting an organ. Are collectives represented as emergent properties of the agents or as a separate kind of entity with its own state variables and traits?

  • Observation. What data and patterns must be observed from the ABM for testing, understanding, and analyzing it, and how are they collected?

ODD model descriptions should describe how these design concepts were taken into account. It is possible to leave out some of the concepts if they are not included in the ABMs design at all (e.g., if adaptive behaviour is simple, then Objectives, Learning, and Prediction may be irrelevant). Four concepts are relevant to almost every ABM: emergence, interaction, stochasticity, and observation.

2.2 Examples

As a simple example, we apply the ODD protocol to the “Segregation” model published as part of NetLogo (Wilensky 1997), which was inspired by the work of T. Schelling (e.g., Schelling 1971).

  • Purpose. The model addresses segregation of households in cities: why members of two different groups tend to separate into different neighbourhoods. The model explores the relationship between how tolerant individuals are of the opposite group and how segregated neighbourhoods are, when individuals express intolerance by moving.

  • Entities, state variables, and scales. The model entities include mobile agents that represent households, and square grid cells that represent houses. Households are characterised by their location (which house they occupy) and their colour, which is either blue or red. Households also have a state variable “happy?”, a boolean variable set to false if the household is intolerant of its neighbours. The grid cells make up a square 51  ×  51 cells in extent, with no depiction of roads or other spaces between them. The space is “toroidal”: if an agent moves off one edge of the space, it reappears on the opposite edge; and the neighbours of a household on one edge of the space include those on the opposite edge.

The length of a time step is unspecified but represents the time in which a household would decide whether to move. The number of time steps in a model run is an emergent outcome: the model runs until no households are intolerant of their neighbours and want to move (see submodel move).

  • Process overview and scheduling. The following actions are executed, in this order, once per time step.

    • If no households are intolerant of their neighbours (“happy?” is true for all households), then the model stops.

    • The households who are intolerant of their neighbours (“happy?” is false) execute the submodel move. The order in which these households execute is randomly shuffled each time step.

    • All households update their “happy?” variable (see submodel update).

    • Outputs for system-level results are updated.

  • Design concepts. The key outcomes of the model are segregation patterns, especially how strongly segregated the entire system is; these outcomes emerge from how tolerant households are to unlike neighbours. The households’ adaptive behaviour is to move when their objective – to live in a neighbourhood with the fraction of unlike neighbours below their tolerance threshold – is not met. The behaviour does not involve learning, or prediction other than the implicit prediction that moving will lead to a neighbourhood where the tolerance objective is met. Households sense the colour of other households on the eight surrounding grid cells. Stochasticity is used in only two ways: to initialise the model so that it starts unsegregated; and to determine the new location of households when they move, because modelling the details of the movement is beyond the scope of this model. Observations include a visual display of which colour household is on each grid cell, and two numerical results: the mean (over all households) percent of neighbours of similar colour, and the percent of households with “happy?” false.

  • Initialisation. A user-chosen number (typically, 2000) of households are initialised. They are each placed on a random empty grid cell and given a colour randomly, with equal probability of red and blue.

  • Input data. The model does not use input from external models or data files.

  • Submodels. The submodel move is performed by individual households if they are not tolerant of their neighbours. The household chooses a direction randomly from a uniform real distribution between 0° and 360°, then moves forward a distance drawn randomly from a uniform real distribution of 0–10 grid cell widths. If there is already a household on the grid cell at this new location, the household repeats the move. If the new grid cell is empty, the household moves to its centre.

The submodel update is conducted by all households, to determine whether they tolerate their neighbourhood. The tolerance of households is determined by a parameter “%-similar-wanted”, the value of which ranges from 0 to 100 and applies to all households. A household’s “neighbours” are all households on the eight surrounding patches. The household’s variable “happy?” is set to false unless the number of neighbours with the household’s colour is greater than or equal to “%-similar-wanted” divided by 100 and multiplied by the number of neighbours.

Using ODD for such simple models may look overdone. However, the ODD model description has a clear hierarchical structure, different elements of the model are easy to find, and all information required for implementation is provided. Moreover, the design concepts provide some explanation of why the model was designed as it was.

An important benefit of ODD is that any ABM can be described in exactly the same format as the segregation model. ODD thus provides a unifying format. Consequently, using ODD to rewrite model descriptions made it much easier to compare three different models of land use/land cover changes (Polhill et al. 2008a): SLUDGE (Parker 1999; Parker and Meretsky 2004), SOME (Brown et al. 2004, 2005), and FEARLUS (Polhill et al. 2001; Gotts et al. 2003) with the ELMM extension (Polhill et al. 2005, 2008b).

Just identifying the entities and state variables of the three models made it easy to see the conceptual similarities of these models, but also their differences in complexity, resolution, and scope (Table 17.2). Obviously, SLUDGE and SOME are conceptually very simple and similar, whereas FEARLUS  +  ELMM can represent quite complex landscapes and scenarios. FEARLUS  +  ELMM could, however, probably be mapped to the two simpler models by choosing simplified parameterisations and initialisations. Model-to-model comparison, and thus transfer of questions, model designs, and results, is greatly facilitated by using ODD. This applies also to comparing process overviews and schedules, and design concepts (Polhill et al. 2008a).

Table 17.2 Comparison of the entities (bold) and their state variables (itemized lists) of three land use model (see text for references). For a full ODD description, in addition the precise number and type of state variables needs to be specified, for example for “biophysical characteristics” or “aestethic quality” (See also Polhill et al. 2008a)
Full size table

2.3 ODD as Design Patterns for Formulating ABMs

Since 2006, ODD has been used for over 70 individual- and agent-based models. ODD has mainly been used by ecologists, but recently applications are appearing in other domains as well, e.g., microbiology (Hellweger et al. 2008), socio-ecology (Guzy et al. 2008), biomedical research (Galvão and Miranda 2009), and anthropology (Premo and Hublin 2009). After reviewing most of these applications of ODD and summarising the questions and comments from many ODD users, Grimm et al. (2010) revised ODD slightly but completely rewrote its description and explanation to make its use easier.

Reviewing these applications of ODD revealed unexpected benefits, including that ODD provides a general format for thinking about and designing ABMs. ODD makes us ask questions about model structure and design in a hierarchical way: structure and overview first, processes and details later. This facilitates translating the often confusing and vague conceptual models which are the starting point of any modelling process (Grimm and Railsback 2005) into model formulations. Scientists with no background in agent-based modelling found ODD helpful for understanding what would be involved to implement an ABM of their system and problem. Likewise, communication between students who develop ABMs and their supervisors is made much more efficient by using ODD. ODD also facilitates communication between different domains, which usually have their own styles of modelling and communicating models.

Last but not least, ODD corresponds to “design pattern” in software engineering (Gamma et al. 1994), which describe “recurring solutions to common problems in software design” (Wikipedia 2009). In agent-based modelling, ODD is a design pattern for formulating new models, and for re-formulating existing models.

3 Patterns and Stylized Facts

While ODD has proven very important and useful, it provides only a general structure for formulating ABMs. It cannot, by itself, prevent “ad hocery” in the choice of entities and state variables. ABMs can still be too simple or too complex to be useful for inferences about how their real counterparts are working. We need a strategy for increasing the chance that our models explain the internal organisation of real complex systems by guiding our choices of what to include in a model and in what detail.

Such a strategy emerged in ecology over the last 15 years: pattern-oriented modelling (POM). It is based on the notion that complex systems provide indicators of their internal organisation: patterns, signals, or regularities that cannot be explained by random processes and therefore call for an explanation (Heine et al. 2007). For example, if vegetation in semi-arid regions shows a characteristic banded structure (Fig. 17.1), we can take this pattern as an indicator of how vegetation is affected by, and affects, the distribution of water. Our task thus is to “decode” the patterns observed in the real systems. We do this inversely by trying to find models that produce the same patterns.

Fig. 17.1
figure 1_17

Banded vegetation pattern in semi-arid regions produced by a simple grid-based model (Thiery et al. 1995; model reimplementation courtesy of U. Berger). The 100  ×  100 cells each have the size of a tree. Dark cells are covered by woody vegetation, white cells are not. Implicitly, the landscape is assumed to have a gentle slope so that runoff from rain flows from top to bottom

Full size image

This approach is certainly not new but just the basic idea of science, or inference (Platt 1964). Patterns were key to revealing the internal organisation of atoms and the universe (spectral patterns), identifying quasars (periodic signals), and proposing asteroids as the cause of mass extinctions (unusual Iridium concentrations in geologic strata). However, one important and often neglected point in the science of agent-based complex systems is that a single pattern is rarely enough to decode the internal organisation. Many different models are, for example, able to reproduce cycles in the population dynamics of small mammals (Czárán 1998). How can we select the right model or falsify models that produce cycles for the wrong reasons?

The basic and simple idea of POM is to use multiple patterns, observed at different scales and hierarchical levels, as multiple criteria for selecting among alternative model structures, submodel formulations, and parameter sets. One single pattern might be relatively “weak” and contain little information so that still, say, half of the models considered would reproduce it. But by adding a second pattern, we possibly reduce the degrees of freedom in model structure. Thus, every single pattern serves as a filter. Again, this method of inference is old and regularly used by detectives who use motives, alibis, witness statements, and evidence to filter possible suspects. Or, in the famous story of the discovery of the structure of DNA (Watson 1968), X-ray diffraction patterns indicated a spiral structure, but many such structures were compatible with this patterns. Only after two additional patterns, Chargaff’s rule and the geometry of purine and pyridimine bases, were taken into account was the real structure identified.

In social sciences and in particular in economics, the notion of “stylised facts” (after Kaldor 1961) corresponds to what we here mean by patterns. In the modelling literature of these domains, stylised facts often seem to be disregarded as too vague to use for rigorous inferences. The point of POM is, however, to increase inferential power by combining multiple vague, weak, or qualitative patterns or stylised facts (see also the discussion of ‘middle range models’ by Gilbert (2008, pp. 41–44)).

4 Pattern-Oriented Modelling

In POM, multiple patterns are used for three elements: designing model structure (what entities, state variables, and processes to include); selecting among alternative submodels; and determining parameter values.

4.1 Patterns for Model Structure

The key question of POM’s first element is: what do experts who know the system well consider to be characteristic features, or essentials, of the system? For example, in growing cohorts of trees, competition among neighbouring trees increases as the trees grow. Initial differences in size and the trees’ spatial distribution cause some trees to be more suppressed by competition so they die, which is called “self-thinning” in ecology. The resources used by the trees that die (space, light, nutrients) become available to survivors, who can grow further until competition is again strong enough that more trees die.

A characteristic pattern of this process is that the relationship between average tree biomass and tree density follows a power law: a straight line on a log-log graph. Moreover, the slope of this line is similar for virtually all plant species and environments. Until about 2000 it was believed that the slope has to be, for theoretical reasons, –3/2 but a new theory predicts a slope of −4/3 (Enquist et al. 1998).

The consequence of this pattern for modelling self-thinning was that models focussed on tree density and average biomass as state variables. Different models, however, can each produce a power law with the right slope. So, what other patterns characterise self-thinning? This question is not fully answered yet, but it seems that the spatial distribution of trees is quite regular during the entire process, which indicates that neighbour competition, depending on local tree density, is important.

To potentially explain this regular spatial distribution, a model should be individual-based so that space and local interactions can be represented. Trees might be represented just by their biomass, which corresponds to a certain “zone of influence” over which they interact with neighbours (Berger and Hildenbrandt 2000). No further patterns seem to be known yet that would suggest including further state variables of the trees, e.g. height, crown shape, etc. Interestingly, power laws are also found in the size distribution of cities, where probably similar mechanisms of neighbour competition are the underlying mechanism (Batty 2005).

To use POM to design a model structure, we identify multiple patterns, typically two to five, which seem to characterise the system and problem being modelled. Making it possible for the model to reproduce each pattern requires adding some state variables and processes. These make the model more complex but also rich enough in structure to be tested against not only one, but multiple patterns. The models are also “mechanistically rich” (DeAngelis and Mooij 2003) so that they can be checked for patterns that were not used for model formulation and calibration.

An example of a mechanistically rich model is the natural beech forest ABM of Rademacher et al. (2004), who found characteristic patterns in the age structure and spatial distribution of canopy trees in model results. These patterns were not at all considered during model formulation and verification, which were driven by other characteristic patterns, but agreed very well with observations (Rademacher et al. 2001). This discovery indicated that the model captures essentials of the system’s internal organisation instead of just being fit to observations. One generic feature of structurally realistic models is that they reproduce observed patterns at different hierarchical levels simultaneously. They capture essential behaviours of both the system and its agents.

Quite a few ABMs have been oriented towards reproducing sets of observed patterns, but usually this is not stated explicitly. POM means to use multiple patterns systematically and explicitly.

4.2 Patterns for Selecting Submodels

ABMs usually comprise a set of submodels that describe the model entities’ behaviour. The segregation model described above, for example, has two submodels, move and update. Formulating submodels includes many degrees of freedom. The decision of whether and where to move could, instead of being very simple and partly random, have been a complex algorithm using knowledge about distant locations and prediction of future changes in neighbourhoods. Submodels, like entire models, must compromise between being simple and unrealistic vs. more realistic and harder to understand. How much detail, and what detail, must we have to solve a particular problem?

The idea of POM for selecting submodels is to use the entire ABM as a virtual laboratory in which we separately test alternative submodels: which submodel, or theory (Grimm and Railsback 2005), is able to reproduce multiple observed patterns simultaneously? If more than one submodel can do so, which is the most simple one? For example, Railsback and Harvey (2002) distilled from the literature six broad patterns in the habitat selection behaviour of trout; their ABM was designed so these patterns could emerge but were not forced to emerge. Three alternative habitat selection submodels were tested, and only one could reproduce all six patterns.

Contrasting alternative submodels or theories by their ability to reproduce sets of patterns corresponds, in principle, to more formal model selection algorithms from information theory, which are used to select among alternative simple descriptive models (Piou et al. 2009).

One important advantage of contrasting alternative submodels is also that it helps communicate that “the model can be wrong” (J. Goss-Custard personal communication) and thus counter the widespread notion that complex ABMs cannot be wrong because you can fit them to any result you want. The segregation model, for example, does not reproduce segregation if the decision to move is entirely random. Likewise, Janssen et al. (2009) found that two ‘naïve’ models of behaviour in a common dilemma game, random walk and greedy agents, did not reproduce all four patterns identified in controlled laboratory experiments.

Of course, this kind of high-level, pattern-oriented submodel selection cannot be applied to all submodels of an ABM but only to those that represent key behaviours; examples from ecology are habitat selection (Railsback and Harvey 2002) and foraging strategy (Goss-Custard et al. 2006).

4.3 Patterns for Parameterisation

Pattern-oriented models are usually of moderate complexity with typically 10–20 parameters. Ideally, all parameter values could be determined directly from information about the agents being modelled, but this is virtually never the case. Estimating parameter values from expert knowledge is often sufficient to get the qualitative behaviour of the model right, because experts often know much more than can be expressed by hard numbers. One important benefit of POM is that this qualitative knowledge can be included via empirical if-then rules and tested against multiple patterns.

Nevertheless, if models are to be used to make inferences about the real world and support decision-making, “guestimated” parameterisations can be too uncertain. For example, Wiegand et al. (2004) developed a spatially explicit ABM of brown bears re-invading the Alps from Slovenia. Uncertainty in the predicted population growth rate in a certain area was very high (Fig. 17.2). Wiegand et al. therefore used multiple patterns in spatial distributions and census time series to reduce uncertainty in parameter values.

Fig. 17.2
figure 2_17

Population growth rate (red bars) and its coefficient of variation (black bars) of a brown bear population as predicted by the model of Wiegand et al. (2004). Each pair of bars represents one set of model parameterisations, as (left to right) more patterns are used as “filters”: parameterisations that do not reproduce the pattern are no longer used. Uncertainty decreases greatly as more patterns are used

Full size image

They created a large number of parameter sets by sampling values of the uncertain parameters from their assumed ranges. Then, each observed pattern was used as a filter: parameter sets leading to model output which did not reproduce the pattern, according to quantitative criteria, were discarded. Some patterns could be reproduced by many parameter sets, others contained more information and reduced the set of possible parameterisation more. In combination, however, only a few (typically 10–30) parameter sets fulfilled all patterns. Using only these “filtered” parameterisations strongly reduced uncertainty in the model output (Fig. 17.2).

Again, this technique of “Monte Carlo filtering” is not new and is used in other domains under the name “inverse modelling”, but Wiegand et al. (2004) first demonstrated its power for parameterising ABMs and linked it to the other two elements of POM. Inverse parameterisation, like the entire POM strategy, is independent of its origin in ecology (Grimm et al. 2005) and applicable to ABMs in general (e.g. Janssen et al. 2009).

5 Discussion

ABMs are fundamentally different from the traditional models of many fields, especially from models based on system-level equations. We now have enough experience to know that the processes of designing, building, and doing science with ABMs are different from traditional approaches. Traditional system-level models are designed and described using established “languages” such as differential equations and flow diagrams, and have their complexity determined mainly by mathematical tractability. With ABMs, we need new standards and strategies. This chapter describes two such standards and strategies that have already proven very useful.

The ODD standard addresses the need for a way to describe both the essential characteristics and the full details of an ABM. The value and success of ODD is illustrated by how widely and rapidly it has been adopted. Benefits of ODD include allowing readers to quickly get an overview of a model’s most important characteristics, providing sufficient detail for replication, and providing the modeller with a framework for thinking about and designing an ABM from the start.

The POM strategy addresses three needs in the process of designing and using an ABM. First, it helps with the fundamental model design problem: determining what entities, variables, and processes must be in a model, so it is complex enough but not unnecessarily complex. Second, it provides a way to develop and test theory: rules for how individual agents behave that are shown, via pattern-oriented analysis, to reproduce essential dynamics of the system. Finally, POM can be used to filter potential parameter values in an efficient and rigorous way.

Both ODD and POM seem “messier” than the standard techniques of traditional modelling, but that is simply a reflection that ABMs are messier than simpler models: if we want to model complex systems, we must use more complex models. But the messiness of ABMs is exactly why we need techniques like ODD and POM: to help us cope with the complexity. We can view these techniques as evidence of the establishment and acceptance of ABMs as a way of doing science.