By Jen Badham
The most familiar models are predictive, such as those used to forecast the weather or plan the economy. However, models have many different uses and different modelling techniques are more or less suitable for specific purposes.
Here I present an example of how a game and a computerised agent-based model have been used for knowledge synthesis and decision support.
The game and model were developed by a team from the Centre de Coopération Internationale en Recherche Agronomique pour le Développement (CIRAD), a French agricultural research organisation with an international development focus. The issue of interest was land use conflict between crop and cattle farming in the Gnith community in Senegal (D’Aquino et al. 2003).
Agent-based modelling is particularly effective where understanding is more important than prediction. This is because agent-based models can represent the real world in a very natural way, making them more accessible than some other types of models.
Developing the game and model
Working with a local partner, CIRAD ran a series of workshops with Gnith villagers and other land use stakeholders. Together, the group developed a game that highlighted the key features of the conflict. CIRAD then created a computerised version of that game, an agent-based model.
In the first step, workshop participants decided that the important decision makers are (crop) farmers and cattle herders, and they became the two types of players in the game. The key influences on their land use decisions were identified as soil type, distance to water and time of year.
These factors were used to establish the game rules. A grid was overlaid on a map of the local area to create a playing board. Each grid location contained either water or one of six soil types and was scored by workshop participants for:
- suitability for two types of agriculture (rice and vegetables), based on soil type; and,
- suitability for cattle at different times of the year, based on access to water and fodder availability.
Gameplay consisted of each player selecting a location in turn, with a new selection each ‘month’ for twelve rounds (one year). Players were not permitted to play the same role as their real-life occupation. The objective of the game was to acquire enough points throughout the year to support a family by occupying high scoring locations.
In many ways, this game is already an agent-based model, as it has the following characteristics:
- The game represents a process, that of land use decisions.
- Actions are autonomous; each player makes their own decision without any central control.
- Differences are meaningful; plots of land change their characteristics over time, farmer and herder players make different decisions, and the same player will make different decisions at different times.
- There is interaction between players and between players and the environment; a player gains points by choosing a plot of land, that choice makes the land unavailable to other players and also changes the land’s characteristics, for example, when a farmer harvests a crop.
The capacity to simulate individuals with autonomy, heterogeneity and interaction is the defining characteristic of agent-based modelling, and what makes it the appropriate choice if those features are important drivers of the system being modelled. All of these features are essential in appropriately representing the decisions that Gnith land users make. The conflict is a feature of the system that is generated by the individual land use decisions.
At the end of the game, players explained the motives for their actions and the problems they faced during play. The researchers then developed a computer simulation to replicate the game, including the point system and decision processes of players. This programmed implementation of the game was an agent-based model.
Using the game and model for knowledge synthesis
The game and model captured a shared understanding of the most important features of the conflict situation and the preferences among the participants. This synthesis of diverse of knowledge held by different stakeholders is facilitated by the modelling process and is not unique to agent-based models.
The figure shows the common phases in projects that develop models with stakeholders (participatory modelling) and the relevant steps within those phases (Badham et al. 2019); knowledge synthesis occurs primarily in the planning and development phases.
The steps of problem definition (focus on land use conflict) and stakeholder planning contributed to the shared understanding by ensuring that the workshop participants covered the range of relevant knowledge. This knowledge was elicited through a process where stakeholders themselves decided what was essential to represent in the model. In establishing the game rules, required data about such factors as the soil quality were automatically collected. The game then became a very detailed conceptual model for the agent-based model.
While all participatory modelling processes support knowledge synthesis, the rules in agent-based models are relatively natural and accessible compared to some other modelling methods. The computer code takes the perspective of individuals taking actions based on their own characteristics and situation. This perspective facilitates engagement and participation, as the stakeholders are providing their own experience for translation into model rules.
Using the game and model for decision support
Players were able to try out different approaches and see the consequences of their actions with the game. However, game playing is time consuming, which limits the number of approaches that can be tried. The advantage of the computerised version is that it is much quicker and can therefore be used to extrapolate from many different ‘what if?’ proposals.
Furthermore, because participants played different roles than their own real-world role, they gained an understanding of the problems faced by others and were more willing to compromise and find proposals that worked for everybody. For example, farmers had previously been unwilling to discuss herder access to post harvest fodder but the game and model allowed this idea to explored.
The participants experimented with several proposals, including regulations about access to water and designating certain areas of land for livestock or to forbid livestock. The agent-based model was used to explore the likely outcomes of these proposals and then the participants were able to discuss these outcomes to form a consensus. The game and model facilitated productive discussions that led to much greater commitment to the agreed policy.
Much of the knowledge synthesis occurred during the planning and development phases where stakeholders made their decision making processes explicit. Using the model during the application phase allowed potential solutions to be explored and facilitated a community decision. The CIRAD approach of using a game to design the agent-based model promotes engagement as the game is very accessible. However, similar knowledge synthesis and decision support outcomes can be achieved without the intermediate step of the game, as described in a different land use agent-based model presented in the blog post Models as ‘interested amateurs’ by Peter Barbrook-Johnson.
For non-modellers, the unique perspective of agent-based modelling can be difficult to understand. If you are a reader who is not a modeller, does the intermediate game help to understand why agent-based modelling is effective when the system being modelled has heterogeneity and interaction?
I tend to use diagrams as the conceptual model when designing an agent-based model. For readers who are modellers, what do you use and what advantages have you found?
Badham, J., Elsawah, S., Guillaume, J. H. A., Hamilton, S. H., Hunt, R. J., Jakeman, A. J., Pierce, S. A, Snow, V.O., Babbar-Sebens, M., Fu, B, Gober, P., Hill, M. C., Iwanaga, T., Loucks, D. P., Merritt, W. S., Peckham, S. D., Richmond, A. K., Zare, F., Ames, D. and Bammer, G. (2019). Effective Modeling for Integrated Water Resource Management: A Guide to Contextual Practices by Phases and Steps and Future Opportunities. Environmental Modelling and Software, 116: 40-56. (Online) (DOI): https://doi.org/10.1016/j.envsoft.2019.02.013
D’Aquino, P., Le Page, C., Bousquet, F. and Bah, A. (2003). Using Self-designed Role-playing Games and a Multi-agent System to Empower a Local Decision-making Process for Land Use Management: The SelfCormas Experiment in Senegal. Journal of Artificial Societies and Social Simulation, 6, 3: Article 5. (Online – Open Access): http://jasss.soc.surrey.ac.uk/6/3/5.html
Badham, J. M., Chattoe-Brown, E., Gilbert, N., Chalabi, Z., Kee, F. and Hunter, R. F. (2018). Developing Agent-based Models of Complex Health Behaviour. Health and Place, 54: 170-177. (Online) (DOI): https://doi.org/10.1016/j.healthplace.2018.08.022
Biography: Jen Badham PhD is based at Queen’s University Belfast, Northern Ireland. She has worked as both a modeller and policy advisor and is currently interested in the way in which social networks influence changes in behaviour.
20 thoughts on “Agent-based modelling for knowledge synthesis and decision support”
Jen, In your section(s) on ‘Developing the Game and Model’, I finally see where the assumptions in an agent-based model can come from. Up to this point I’ve only seen the structured programing statements that make up the complicated details of the model, but your article shows how the development can start with what system experts and their friends know about the reality that is bothering them. Thanks!
hi George, good to hear from you. It sounds like you have seen the UML (unified modelling language) approach, which I have never used. I guess the system dynamics equivalent would be to try and design using difference equations rather than the normal approach of causal loop diagrams.
I love your “game” approach, and your analogy. Prompts me to say good system dynamics practice, as Forrester preached for about sixty years, is never to use “causal-loop” diagrams for conceptualization, but rather always to begin with dynamics and stocks. I have been urging CLDs with explicit stocks for decades. But CLDs have acquired a life of their own, which is proving hard to counter.
Unfortunately, it’s not actually my approach. I love the game idea and I find it much easier to explain ABM by going through a game, but I’ve never had a project where I could use it. I do half/one day workshops to design then other workshops every few months to test prototypes etc. That’s pretty true whatever modelling method I use. But the type of questions I ask are definitely flavoured by the method so individual choice for ABM and stocks/flows for system dynamics etc. You may be interested in participatory systems mapping (https://www.cecan.ac.uk/sites/default/files/2019-03/PSM%20Workshop%20method.pdf), one of the many slightly different flavours of more formal conceptualisations where the conceptual model is analysed directly rather than creating a quantitative model.
Hi Jen – really nice story in that post (and in the comments!). I really enjoyed being walked through the various stages and approaches at each. Thanks for sharing. Val
Thanks Val, glad you found it interesting
long time no see…;-)
Self-cormas was the ‘founding’ project of what has become known as a the Companion Modelling (ComMod) approach and research collective I once called home (my career path took me through other avenues since that time…). A couple of thoughts I’d like to share:
1) Self-cormas has probably been a once-off experience in its complexity, richness and success (a co-management charter was signed between parties at the end of the process). Several years later, D’aquino, Dray and I were involved in the Atollgame project/experiment (https://ro.uow.edu.au/cgi/viewcontent.cgi?referer=https://scholar.google.com.au/&httpsredir=1&article=1024&context=smartpapers ) that used the same approach to solve conflict between landowners and public works in Tarawa. The modelling and gaming process worked beyond expectation…unfortunately, the charter was never to be implemented due to (1) vested interests and (2) inability of the approach to ‘enforce’ responsibility principles upon participants after the game (see: https://journals.sagepub.com/doi/pdf/10.1177/1046878107300673).
2) in situations of open conflict, it is important to the model to be perceived by all parties as ‘neutral and transparent’ as possible. In the Atollgame experiment, we used knowledge engineering techniques to create mind maps of various actors and extract two sets of beliefs out of them: (1) common or non conflicting beliefs and (2) conflicting beliefs. The first category was used to build the agent-based model, used as computing engine for various dynamics and interactions; the second category fed a set of rules in the game itself, triggered by the Game Master. We ordered the latter in order to create a growing tension in the game, starting from an utopian setting (‘paradise island’) and moving progressively into a dystopian reality. This strategy proved to be a winning one (see first link above).
Hi Pascal, yes this hemisphere switching does tend to make it hard to bump into each other 🙂 Thanks for the additional references and comment. I think it’s a particularly important observation that the selection and commitment of participants is critical in later implementation.
What are “cormas” and “self-cormas”?
Hi George – Thanks for raising this! As far as I can ascertain, Cormas (for Common Pool Resources and Multi-Agent Systems) is an Agent-Based Modeling platform – see http://agritrop.cirad.fr/589018/.
Regarding SelfCormas: in the article cited by Jen, the authors say: “Before the SelfCormas experiment, the Cormas platform proposed a very schematic display of the agents located on the spatial grid of the model, using simple polygons. The user had only the possibility of selecting the number of sides to make a triangle, a diamond, or a hexagon. For the SelfCormas experiment, in order to make a direct mapping of the sticks used during the role-playing game into agents moving on the computer screen, we have introduced the possibility of editing images with more representative shapes. This feature has been generically set into the simulation platform. Other examples of features specially designed for the SelfCormas experiment that were later integrated into the kernel of the simulation platform are (i) the possibility of encoding access rules defined for a given activity during an interval of time (initial and final time-steps) by mouse-clicking in any cell of the spatial grid, (ii) the possibility of adding or removing objects located in any cell of the spatial grid by mouse-clicking. In the SelfCormas experiment, the first one has been used to assign restricted access during the wet or dry season, and the second one has been used to add new water points or to design channels.”
Good choice of quote Gabrielle!
In a nutshell:
Cormas is an ABM platform that is meant for multi-disciplinary teams to collaborate around a model of a given socio-Environmental system.
SelfCormas was the name of a project developed in Senegal, using Cornas in a fully Participatory process: local participants defined the problem, identified the dynamics and future rules of the game, as well as validating the outcomes of interactive simulation. Henceforth the name: SELF Cormas.
Thanks! But I have a long way to go here before I have real understanding of the schematic display, “sticks”, shapes, the kernal of the simulation platform, access rules, and so on.
Great stuff Jen, useful overview of well-tested and reliable approaches.
For me, as a modeller, the most interesting things to think about now, are how different participatory modelling approaches fit together. Voinov et al (2018) largely help us pick one method at a time and think a bit about workflows, but I think things get really useful and fun when we think about using methods in combination. I know you, Corinna Elsenbroich and Barbara Befani are thinking about using ABM (agent-based modelling) with Bayesian Updating (though not necessarily in participatory mode), myself, Brian Castellani and Corey Schimpf have been thinking about ABM and QCA (qualitative comparative analysis; Castellani et al 2019), and Alex Penn and myself are keen to develop our Participatory Systems Mapping method to include some dynamical modelling of key sections of a map. The emerging web of complexity-appropriate modelling methods could be very compelling and valuable. Probably it broadens out beyond participatory use alone, but personally I think all modelling of complex social issues should always have some participatory element.
Voinov et al 2018 – Tools and Methods in Participatory Modeling: Selecting the Right
Tool for the Job – https://pdxscholar.library.pdx.edu/cgi/viewcontent.cgi?article=1162&context=etm_fac
Castellani et al 2019 – Case-based methods and agent-based modelling: bridging the divide to leverage their combined strengths – https://www.tandfonline.com/doi/abs/10.1080/13645579.2018.1563972
One of the reasons that I like this case study is because the combination of methods delivers good participation. But there is also a pedagogical aspect, I feel that explaining ABM is easier via the game. As well as the combinations you mentioned, system dynamics (SD) modellers routinely use causal loop diagrams as the path into the more mathematical model, recognising the diagrams as a separate method but very much part of the SD process.
Very interesting example of an agent-based model. We use these commonly in modelling infectious diseases, but we rarely need to consider competition or motives of the agents. It seems that there could be a game-theory component to analyzing this system – agents either selfish or acting for the common good. Or is this entirely captured by regulation and the model assumes all agents act in their own interests?
There is a substantial body of work that uses game theory to define agent behaviour. Another common rational choice approach is to construct some sort of utility function and the agent chooses their action so as to maximise the utility. However, one of the characteristics of ABM is the flexibility of agent decision rules, you can pretty much implement any agent decision rules you want.This is great if you happen to know how the entities in the system being modelled make decisions (which is why the game is so useful in this example). But for much social science, the behaviour mechanisms are unclear. That’s one of the three challenges I discuss in the ‘additional reading’ paper.
In epidemiological models, the infection may be relatively straightforward to model (eg if an infectious and susceptible pair meet, there is some probability of transmission). But the behavioural responses may be more difficult (eg people reduce their social contact if there are many new infections in their neighbourhood). Do you include that feedback process in your infectious disease models?
We certainly can model behaviour, and it’s likely important (see: https://www.ncbi.nlm.nih.gov/pubmed/17916550) but it is difficult to measure, and likely influenced by media attention… we only know that there is a lot of disease around if there is lots of news about it! So, a little tricky to model well.
True – I’m not sure there’s been any real progress since the approaches were summarised by Funk et al in 2010 (https://doi.org/10.1098/rsif.2010.0142)
Reblogged this on Systems Community of Inquiry.
thanks for the reblog, much appreciated