Frequently asked questions about agent-based models
While agent-based models are not a new methodology, they are not currently widely used in population health. We’ve put together a list of frequently asked questions about agent-based models, to make this methodology more accessible and to start discussions around how agent-based models could be used to address population health challenges.
Agent-based models are computer simulations of the behaviour and interactions of autonomous ‘agents’ such as individuals, groups or organisations. The goals of using an agent-based model in a health context include: discovering how the individual-level actions of agents and their interactions with each other and with their wider environment lead to population-level outcomes and patterns of health-related behaviour; revealing which of these actions and interactions we need to research more; and designing or testing interventions virtually before committing resources in the real world.
How is it different from the usual kinds of evidence or models that health policy/practice uses?
More traditional methods, such as statistical modelling and machine learning, use data we already have about a system or process and use this to predict future behaviour. In contrast, agent-based models are built on theories about the mechanisms underlying a system or process, which are used to construct decision-making agents and virtual environments. Agent-based models are particularly useful when we are interested in exploring and explaining the behaviour of a system, and not just predicting outcomes.
How is it different from other kinds of simulations?
Agent-based modelling is often compared with microsimulation in which individuals in the model move through their life entering and leaving different states based on probabilities. In contrast, in agent-based modelling, agents are granted the capacity to decide when, or if, they should change state or behaviour: they have built-in decision rules that govern how they will react when environmental or social factors change around them. The consequence is that agents in agent-based models are far more complex than an individual in a microsimulation.
Are agent-based models new?
That depends on your definition of ‘new’! Many credit Thomas Schelling as the creator of the first agent-based model – he developed a simple and elegant model of residential housing segregation in 1971. So, compared to statistical methods, which have a long and storied history stretching across centuries, agent-based models are relatively new – however, they have been around for nearly 50 years. While agent-based models have been successfully applied across wide range of disciplines, they have not yet been widely used in population health.
We know that population health is a complex ‘system’ driven by multiple influences that interact with other. These systems are very difficult to model statistically due to the many different relationships between individuals and the environment which are often non-linear and bi-directional. Agent-based models are ideal for these situations and sometimes the development of an agent-based model can be the only way to improve our understanding of the system’s behaviour.
Can agent-based models replace other modelling techniques?
Agent-based models are often a more suitable tool than other modelling techniques when we deal with complex systems, and they are necessary to answer certain questions about population-level behaviour. However, agent-based models have both strengths and weaknesses and shouldn't be viewed as a replacement for other modelling techniques, but as an additional tool in the researchers’ tool-box.
What can agent-based model do that conventional approaches to providing evidence about population health cannot?
For public health outcomes that are driven by individual-level actions and the interactions of individuals with their social and physical environments (e.g. obesity), building an agent-based model can provide insight into the mechanisms underlying these outcomes. Agent-based modelling can help us to understand and predict a system’s behaviour by developing a theory of the individual behaviours and interactions between agents and their environment that are relevant to the chosen system. Statistical methods don’t model these behaviours and environmental interactions explicitly, so in these cases they’re less suitable than an agent-based model.
What can’t they do that more conventional approaches to providing evidence can?
When outcomes are characterised by relatively simple and causal relationships, the relationships between variables can be effectively and precisely analysed by statistical modelling tools. Agent-based modelling would not add much to our knowledge, and the time taken to build and tune an agent-based model is likely to be misplaced to answer these types of questions.
What kind of questions are agent-based models most suited to answer?
In general, we suggest agent-based modelling for when you want to investigate, understand, and probe the complex interactions and relationships underlying population health challenges. Agent-based models are most useful when we're interested in explaining and exploring the behaviour of a complex system – such as population health questions that may involve interactions between individual health behaviours, social and economic policy, and environmental factors. We can also use agent-based modelling to develop and evaluate complex interventions, such as policy solutions aimed at population behaviour change.
In agent-based models agent decision-making is based on decision rules that lead them to change their state or behaviour in response to other agents or environmental changes. While technically speaking, any computer program will only do ‘what we’ve programmed it to do’, in practice the consequences of our programmed decision rules are often far from obvious.
Agent-based models have been applied to a range of different population health challenges including physical activity, active travel, alcohol, diet, obesity, smoking and tobacco control, and social care. We have provided some examples of how agent-based models have been applied to help understand population health questions and evaluate the potential impacts of policy implementation in our full-length FAQs.
There are no hard and fast rules on how to create an agent-based model. In general, the key steps in developing an agent-based model are:
- Define the topic, or system of interest, for your model
- Identify the key dynamics of the system. i.e. agents, environment, actions, interactions
- Develop a modelling plan
- Build, refine and validate your model
- Close the modelling project
Our full-length FAQs provide additional guidance on each of these steps.
How long does it take?
The length of time to create an agent-based model depends entirely on the characteristics and complexity of the system of interest. A basic model could be developed by an individual in a few hours, while a more complex agent-based model relating to a specific real-world problem might involve a team of people working for several years.
Not necessarily, however some experience with computer programming or scripting languages is extremely valuable. There is a range of accessible agent-based modelling software, including NetLogo, AnyLogic, GAMA Platform, and Repast Simphony, some of which require programming experience, while others have a visual interface. Getting into agent-based modelling does require more programming knowledge than some statistical methods, but probably not as much as you might think, and there are lots of resources and examples available to help you get started.
In order to determine if the agent-based model is right it needs to be validated as an accurate representation of the real-world system that the model was based on. Validation methods often involve comparison of the model output with real-world data to test for similarity in the properties or patterns of population level outcomes. Cross-validation against another model that has already been validated is another option.
By building an agent-based model we try to provide an explanation for the outcomes of a real-world system by including the process and the mechanism that generates the system and outcome of interest. The agent-based model simulation generates data, and once a model is validated the data it generates are considered as crediable, and provide insights into certain aspects of the system.
Join our mailing list to be the first to hear about upcoming events and training, new resources and funding opportunities. If you’d like to be involved in planning PHASE events and activities or if you’re interested in joining one of our working groups, please email email@example.com.
Using machine learning as a surrogate model for agent-based simulations
C. Angione, E. Silverman, E. Yaneske
PLoS One, 2022
Network structure influence on simulated network interventions for behaviour change
J. Badham, F. Kee, R.F. Hunter
Social Networks, 2021
Situating agent-based modelling in population health research
E. Silverman, U. Gostoli, S. Picascia, J. Almagor, M. McCann, R. Shaw, C. Angione
Emerging Themes in Epidemiology, 2021
PHASE: Facilitating agent-based modelling in population health
E. Silverman, U. Gostoli
Proceedings of the 2020 Winter Simulation Conference
Social and Child Care Provision in Kinship Networks: an Agent-Based Model
U. Gostoli, E. Silverman
PloS One, 2020
Sound behavioural theories, not data, is what makes computational models useful
U. Gostoli, E. Silverman
Review of Artificial Societies and Social Simulation, 2020
Modelling social care provision in an agent-based framework with kinship networks
U. Gostoli, E. Silverman
Royal Society Open Science, 2019
Modelling Protocols and Standards
ODD (Overview, Design concepts, and Details) Protocol
The ODD protocol is a standard format for describing agent-based models. Since the protocol was introduced in 2006, several updates have been provided that aim to make it easier to use ODD and make model descriptions more useful and coherent.
A standard protocol for describing individual-based and agent-based models
V. Grimm, U. Berger, F. Bastiansen, et al.
Ecological Modelling 2006
The ODD protocol: A review and first update
V. Grimm, U. Berger, D.L. De Angelis, J.G. Polhill, S.F. Railsback
Ecological Modelling, 2010
Describing human decisions in agent-based models – ODD+D, an extension of the ODD protocol
B. Müller, F. Bohn, G. Dreßler, et al.
Environmental Modelling & Software, 2013
The ODD Protocol for Describing Agent-Based and Other Simulation Models: A Second Update to Improve Clarity, Replication, and Structural Realism
V. Grimm , S.F. Railsback , C.E. Vincenot, et al.
Journal of Artificial Societies and Social Simulation, 2020
TRACE (TRAnsparent and Comprehensive model Evaludation) Framework
Framework to support modellers to “trace” the iterative modelling process and document model development, testing and analysis.
Towards better modelling and decision support: documenting model development, testing, and analysis using TRACE
V. Grimm, J. Augusiak, A. Focks, et al.
Ecological modelling, 2014
Keeping modelling notebooks with TRACE: good for you and good for environmental research and management support
D. Ayllón, S.F.Railsback, C. Gallagher, et al.
Environmental modelling & Software, 2021
Community Standards for Modelling Science
Community standards coordinated through the Open Modelling Foundation that promote the creation and use of more reusable, replicable, interoperable, and reliable models. Standards cover model discoverability and accessibility; documentation; reusability and replicability; and interoperability.
CoMSES Net host community curated information on agent based modeling software frameworks, documentation standards, educational materials, and guides to good practice for developing and documenting computational models for reuse and reproducibility.
NetLogo is a free software tool that provides a multi-agent programmable modelling environment.
Complexity Explorer delivers a range of online courses, tutorials, and resources essential to the study of complex systems, including tutorials on NetLogo and agent-based simulation. Complexity Explorer is an education project of the Santa Fe Institute.
CECAN is working on methods and tools to improve the design and evaluation of policies related to the food, energy, water and environment ‘nexus’, and within areas where these issues interconnect in complex ways. They provide a range of resources around modelling of complex systems and policies, including some specific to agent-based modelling.
On-Line Guide for Newcomers to Agent-Based Modeling in the Social Sciences
Robert Axelrod and Leigh Tesfatsion