Complexity and agent-based modelling

By Richard Taylor and John Forrester

Richard Taylor (biography)

Policy problems are complex and – while sometimes simple solutions can work – complexity tools and complexity thinking have a major part to play in planning effective policy responses. What is ‘complexity’ and what does ‘complexity science’ do? How can agent-based modelling help address the complexity of environment and development policy issues?


At the most obvious level, one can take complexity to mean all systems that are not simple, by which we mean that they can be influenced but not controlled. Complexity can be examined through complexity science and complex system models.

John Forrester (biography)

Complexity science is not a single consistent theory or approach, but spans sets of different tools and techniques from different research disciplines. It is a useful lens to describe real-world policy situations, where formal analytic modelling reaches limitations of tractability – that is, the maths are not workable. Conceptualising the situation as a complex system provides an alternative. The definition of what is ‘complexity science’ can be seen most clearly by looking at some common principles of the models used. However, it should be understood that the underlying phenomena – the real-world complexity – are much more difficult to define than the models would suggest.

Complex system models have characteristics which can make them suitable analogies of complex systems themselves, but these characteristics also make the models difficult to understand fully. The models:

  • have many component parts and therefore many local variables (which give many possible system states).
  • involve interactions among locally connected parts, and the interactions need to be understood just as much as the functions of the individual components. For example, the interactions can contribute to significant non-linearities and emergent properties. An example of a non-linearity is that how one gets from A to B is not the opposite of how one gets from B to A. An emergent property is a property at one level of a system that cannot be defined by its components, for example, water is a liquid, whereas its components (hydrogen and oxygen) are gases. To take a sociological example, ‘resilience’ of a community could be said to be based on the skills and resources of individuals interacting in their landscape.
  • have macro-level properties that are not properties of any components of the system, are difficult to describe formally and – most relevant to policy research – can appear surprising, novel and unpredictable.

Complexity research tends to rely mainly on modelling and interaction theory, but also uses both qualitative as well as quantitative methods. Interaction theory, or network theory, is based on a model which considers the following four aspects: the entities or ‘actors’; the links between them; the attributes of the actors and links; and boundary conditions of the network determining inclusion and exclusion of actors. Qualitative methods – such as narrative inquiry – are needed to understand real-world complexity and its many nuances in diverse social contexts. On the other hand, our models of complexity also need to be supported by research. Qualitative, quantitative, and mixed-method approaches may be employed to help the design and validation of complex systems models.

To take a network theory example – a social network ‘map’ can be generated from a quantitative social survey. This can be validated using a different method, for instance through qualitative stakeholder interviews. This offers the opportunity for ‘ground-truthing’ whether the researcher’s current map or model of the system is broadly correct or as expected, and also for further learning and questioning. (For another example of ground-truthing, see Pete Barbrook-Johnson’s blog post on models as ‘interested amateurs’.)

Agent-based modelling

Agent-based modelling is one type of modelling useful for helping to understand complex systems. It helps in understanding relationships and thus possible causal mechanisms in complex systems, by generating models of them from the bottom up.

For example, agent-based modelling concentrates on describing a social system at the micro-level of the actors within it. This is usually done using a computer model (program). The description for each agent includes a set of instructions or “rules” governing the agent’s behaviour. Agents also have goals and other internal information (knowledge, beliefs, values, etc.) which uniquely shape their actions.

This bundling of data with instructions for agents allows them to be, in practice, coded as autonomous units representing different social entities. The agent descriptions are used as a template to create many copies and thereby populate a model (hence, agent-based models are sometimes also known as multi-agent systems or multi-agent models).

While there is a focus on the micro-behavioural level, models can include many or multiple types of agency at different levels of action, eg., households, firms or local authorities. There is also a focus on interactions with other agents and interaction with the environment: agent-based models have been used quite extensively to understand management and use of environmental resources, as well as adaptation processes under environmental change.

To include greater levels of detail and specificity, of course, brings new difficulties. Where traditional models reduce systems to easy-to-grasp components, agent-based models (and complex system models in general) may be difficult to interpret. It all comes down to how they might be applied, what you want to find out, and whether the difficulties will outweigh the usefulness.

Potential users or developers of agent-based models often ask the following questions which are addressed in the paper by Taylor and colleagues (2016):

  1. Do I need an agent-based model?
  2. Are there good examples of agent-based modelling in my problem domain?
  3. Is agent-based modelling a stakeholder engagement method?
  4. Can agent-based modelling be used in conjunction with other methods?

What other questions would you ask? What experience do you have in tackling complex policy problems? What methods have you used? What possible uses do you see for agent-based models in helping facilitate policy processes?

To find out more:
Taylor, R., Besa, M. C. and Forrester, J. (2016). Agent-based modelling: A tool for addressing the complexity of environment and development policy issues. Stockholm Environment Institute (SEI) Working Paper 2016-12: Oxford, United Kingdom. Online:

Biography: Richard Taylor develops agent-based social simulation models that can be applied to the study of sustainability-related problems and adaptation, and is interested in participatory approaches to inform and improve relevance of models, and put them to wider use. He has expertise in mixed method research and integrated methodology design in applied research. He is a Senior Researcher at the Stockholm Environment Institute Oxford Centre.

Biography: John Forrester is a social anthropologist who uses maps and models to explore the complex relationships behind stakeholders’ understanding of environmental issues. He does this so that ‘situated knowledge’ may feed into both scientific and policy knowledge co-creation processes. John works with the Stockholm Environment Institute at York, where he has experience in multidisciplinary and transdisciplinary science communication for sustainable development; transport planning; upland ecology; flood risk management; coastal ecosystems; and community resilience.

7 thoughts on “Complexity and agent-based modelling”

    • I don’t know any specific examples of using twitter data but I do think that twitter data can be used to inform/validate ABM. I imagine models of information diffusion /spread of ideas would be one easily testable area…

      Also, social science ideas can be used – through ABM and other distributed computation techniques -to improve how systems like twitter work. A specific example could be peer-to-peer networks. David Hales ( used simple, socially-inspired tag-based systems that proved efficient and robust.

      Would love to hear of any other examples !

  1. This is a nice, succinct description of complexity and ABM.

    I’m most interested in question 3 – “Is ABM a stakeholder engagement method?”

    It requires (1) stakeholders who can recognize when they are dealing with a complex system and (2) stakeholders who work with a researcher who’s sufficiently facile with ABM to offer ABM as a potential explanatory tool.

    • Thank you Tom. Regarding whether ABM is a stakeholder engagement method… I think I’d ask whether ABM can be a stakeholder engagement method? in which case the answer is definitely yes.
      The stakeholders only need to recognise their own systems (yes, of course they are complex) but I’ve never met a human being yet who doesn’t understand a lot of the complexity of the system with which they are involved. Our findings have been – and Richard can and I’m sure will add here – that stakeholders need only a very basic understanding of I/O [input/output] processes to the ABM to be able to engage in a meaningful manner, John.

  2. Thank you – very good thinking and scenario building. The brand of complexity I deal with are wicked problems associated with landscape sustainability; socio-economic ecological systems of the “tragedy of the commons”. My approach was biased toward governance and that became my foundation to enable people to interact and transact. My hypothesis was that all wicked problems are based on three principle sources:
    1- System outputs and outcomes vary in scope and scale
    2- System stakeholders measure and value the outputs and outcomes in varied manners and degrees
    3- Stakeholder organizations develop strategies and values based on varied governance frameworks

    Most entities not only start in the wrong source category (1 or 2) but usually ignore or are unaware of the significant force of #3.

    • Tim, Ehh, yes. I agree totally. And the focus upon Participatory ABMs (as opposed to any other sort of individual based modelling) hopefully starts to address this lack? Happy to discuss, John.

      • Yes, I agree that ABM is key as individuals are so easily interconnected. In my work (see eg my blog post at I described these agents as actors described in a matrix of public or private sectors and as a policy-maker or practitioner. This addressed the issue of organization boundaries being less relevant, as belonging to any organization would not define one specifically as a public policy-maker, private policy-maker, public practitioner or private practitioner, although there is tendencies. But beyond the agent status is also the governance style they are beholden to whether that is hierarchy, market or network styles. It is this complex mix that I am able to describe from intra- and inter-organizational perspectives. So I think ABM could incorporate the actor sector type and the governance footprint (ratio of styles) and when actors and footprints are accounted for describes the governance framework.


Leave a Reply to ChrisCancel reply

This site uses Akismet to reduce spam. Learn how your comment data is processed.

%d bloggers like this: