by Marcel Mallow (@NetKnow)
In my first blog post, I would like to reflect a bit on the BEHAVE Spring School (online-edition) on Agent-based Modelling and its implementation in NetLogo, which Deborah and I had the pleasure to attend last week. ABMs are used in many different disciplines due to their ability to simulate large-scale dynamics from bottom-up processes (e.g. in Economics, Sociology, Political Science, etc). In this blog post, I am reflecting on three things: Firstly, where the spike for the ABM modelling approach might stem from, second, some key takeaways from the Behave ABM school, and lastly, how I see it fitting in with my research.
Replication of Axelrod’s #ABM on cultural diffusion in @netlogo is nicely shaping up!
Day 4 at #behaveschool21 with @digitaldust_it pic.twitter.com/6prU06Cr9j
— Behave (@BehaveLab_unimi) April 22, 2021
Replicating Axelrod’s model of cultural diffusion on Day 4 of the ABM Spring School “Behave” – @digitaldust_it
What are agent-based models?
First of all, the terminology on what exactly constitutes an agent-based model (ABM) is not clear. Depending on the discipline and focus of the application they are coined differently; however, most definitions share a component of simulating social and/or natural processes on the micro-level in order to understand the outcomes they shape on the macro-level. These macro-outcomes are often not very evident if we solely analyse or model the actions and behaviour of a single actor, or “agent” in ABM-lingo, ). Below picture illustrates such a fascinating macro-outcome, which would not necessarily be expected from the analysis and modelling of only single birds behaviour.

Swarming, an example of a fascinating macro-outcome, which is not easy to explain or compute – Picture by Christiaan Kooyman (ChriKo) from Hengelo, NL [license: CC BY-SA 4.0]
Some finer definitions and the “ABM mindset”
As Jackson et al 2017 summarise, ABMs are used in many different disciplines due to their ability to simulate large-scale dynamics with bottom-up processes (e.g. in Economics, Sociology, Political Science, etc). According to Bianchi & Squazzoni 2015, ABMs are “computer simulations of social interaction between heterogeneous agents, [which are] embedded in social structures”. While the latter might not be a fitting definition for our colleagues from the natural sciences, more generally, agent-based models allow us to model a system as “a collection of autonomous decision-making entities called agents [, where] [e]ach agent individually assesses its situation and makes decisions on the basis of a set of rules” (Bonabeau 2002). Apart from ABMs spiking in popularity over the last two decades, authors such as Eric Bonabeau (2002), already 20 years ago, have been identifying an emerging “ABM mindset” in the way of making, i.e. the recognition of the importance of micro-scale behaviours and thus describing systems more and more from the perspective of its constituent units.
During my two years at the University of Natural Resources and Life Sciences (BOKU) in Vienna, I noticed more and more colleagues discussing the utility of ABMs in natural resource policy and environmental governance. What I was left with was a strong curiosity for the reasons for this spike in popularity of this type of modelling.
Spike in popularity
So what are possible reasons for this observed spike in popularity?
There are many good reasons, but, in my opinion, the most important ones are related to the simultaneous abundance and lack of resources that especially social science researchers are confronted with nowadays. Firstly, regarding abundance: computing power is now readily available even to researchers low on funding and many universities around the globe even make “super”-computing power available to their researchers. Secondly, regarding a lack of resources: It is costly, difficult and, sometimes, ethically questionable, to get ahold of (especially) social data for multiple reasons.
Why (not) simulate your data?
Bearing in mind these two ideas of the abundance of computing power and scarcity of (social) data, the good news is that you do not necessarily need data to use agent-based modelling as you can simulate is, and thus create it yourself. You are in a way producing the data from the ground up from specific dynamics in your model. What this needs is a lot of computing power (depending on how many times you let your model run; the number of simulations and, obviously the complexity of your model). This can therefore be a good strategy for disciplines, where data is usually sparse and possibly unreliable (e.g. archaeology).
A word of caution: Know your modelling purpose
This sounds terrific; so we don’t have to go out into the world and collect our data in a tedious and rigourous fashion? I am afraid to let you down once more. Even though there are both ABMs, which are built without empirical data, and, ABMs, which are built without being “sufficiently” grounded in theory, it is advisable to have at least, either a good theoretical basis, or empirical data to base your model on. All of this depends on your modelling purpose.
You can read up on 7 different purposes for a model in this paper in the journal JASSS (Journal of Artificial Societies and Social Simulation), which is one of the few journals publishing almost exclusively ABM-papers. In their paper Edmonds et al. (2019) explain how the purposes prediction, explanation and description are empirical (i.e. closely relating to “observed evidence”), while the purposes theoretical exploration, illustration, analogy and social learning do not necessarily have to be mirroring what is observed, but are rather about “ways of thinking about things [or] theoretical properties”.
Why I might build an ABM
Verifying a conceptual model
If you have your agent-based model rooted on strong theoretical foundations, or at least, have a comprehensive conceptual model ready, which you might want to verify (e.g. for internal validity), this is a very typical use of an ABM in the research process (probably most relating to the purposes “theoretical exploration” and “description” above). Personally, I think this could be a good way to scrutinise and verify my conceptual model/framework of knowledge-related processes in scientific collaboration networks. By trying to implement a conceptual model as an ABM, one has to be very explicit (no matter which programming language is chosen) as, for example, clear rules have to be defined for our agent-sets. This process of formalising and writing explicit code can be a good way to make progress in your research design. Another option I see for my iCONN-related research, either additionally or alternatively, is calibrating an ABM with empirical data on scientific collaboration networks.
How calibrated models are trying to respond to some direction of critique of ABM
The researcher coding an ABM is the rule-maker of the universe. Much of the critique of ABMs thus goes in the direction of lacking theoretical grounding or an empirical basis for these rules and modelled behaviours. Responding to this critique, by calibrating an ABM one can make it more “close to reality” in the way that past agent behaviour and its evolution (empirical data) is informing the model. The model is thus not fully imagined and explorative any more, but based on empirical data. However, we also have to think about the way our empirical data was collected as, for example, in the process of obtaining social data the measurement is also quite prone to error.
Assumptions vs. rule-making
Some researchers supporting the ABM – approach suggest that in comparison to other modelling techniques and approaches ABMs make hardly any or no assumptions, especially when seeing ABM as a more explorative tool. In one of the next blog posts I would like to go more into depth on this topic of assumptions (comparing e.g. more statistical modelling approaches such as SAOMs, stochastic actor-oriented models to the ABM modelling approach).
Feedback on this post is much appreciated.
So long,
Marcel (ESR14)
Last, but not least: click here to find out how to create your own turtle breed (in NetLogo turtles are your agents and breeds are types [or classes] of agents)
References
Bianchi F and Squazzoni F (2015) Agent-based models in sociology. WIREs Computational Statistics, WIREs Computational Statistics 7(4): 284–306.
Bonabeau E (2002) Agent-based modeling: Methods and techniques for simulating human systems. Proceedings of the National Academy of Sciences, Proceedings of the National Academy of Sciences 99(Supplement 3): 7280–7287. Available from: https://dx.doi.org/10.1073/pnas.082080899
Edmonds B, Le Page C, Bithell M, et al. (2019) Different Modelling Purposes. Journal of Artificial Societies and Social Simulation, Journal of Artificial Societies and Social Simulation 22(3). Available from: https://dx.doi.org/10.18564/jasss.3993.
Jackson JC, Rand D, Lewis K, et al. (2017) Agent-Based Modeling. Social Psychological and Personality Science, Social Psychological and Personality Science 8(4): 387–395.