Book Review: Agent Zero by Joshua Epstein
Steve Senior is a specialty registrar in public health in Greater Manchester. In previous lives he worked as a policy adviser for the UK Government, and completed a doctorate in neuroscience. In this review he assesses the applicability of the insights in Joshua Epstein's Agent Zero to systems approaches.
In this book, Joshua Epstein sets out a simple but compelling framework that shows how simulated people (agents) following some basic rules can generate surprising social phenomena that are relevant to public health.
Agent_Zero is published by Princeton University Press. The preface and first chapter are available as a free download from the Princeton University Press website. The website also features all the code and models used in the book, so you can play with the model settings for yourself.
Epstein has argued that agent-based modelling represents a new model of ‘generative’ social science [1], in which theories about behaviour are embodied in agents, representing people (or sometimes organisations). These agents are allowed to interact freely in a computer simulation, following some basic rules. If the outcome of the system as a whole resembles the social phenomenon that you are interested in, then you can say that the rules are sufficient to explain the social phenomena. This contrasts with the usual epidemiological approach of studying patterns in the real world and using regression models to tease out the causal variables.
An early and well-known example of agent-based modelling is Thomas Schelling’s model of racial segregation [2]. Shelling’s model has agents of two colours that ‘live’ on a two-dimensional grid like a chess board that might represent a city. Schelling’s agents can move to empty squares if they are not ‘satisfied’. Agents are satisfied when a certain proportion of neighbouring agents are the same colour. Schelling’s model showed that even when agents are happy to be in the minority, the model can still generate segregation resembling the racial segregation in some cities.
Schelling’s model was designed to study a specific social phenomena. In Agent_Zero, Joshua Epstein offers a general framework for agent based modelling that is based on evidence from psychology and neuroscience. Epstein’s agents have three components: a cognitive component; an emotional component; and a social component. These components together are referred to as an agent’s ‘disposition’, and determine whether an agent carries out some action. Epstein bases his emotional component on psychological studies of fear conditioning. His cognitive component, although simple, reflects some of the better known cognitive biases [3]. These two added together make up an agent’s ‘solo disposition’ towards some action. To this, Epstein adds a weighted sum of the agent’s contact’s solo dispositions, reflecting a form of emotional contagion. Epstein’s agents move randomly around a two-dimensional grid. These squares change at random from benign to aversive, reflecting some adverse experience, and this informs the agents’ emotional and cognitive values.
Epstein’s central interpretation is about violence - his agents might represent peace-keeping soldiers moving around a terrain occupied by locals that may or may not be hostile. But there are numerous other interpretations throughout the book. One example that is relevant to public health is about vaccine refusal. In this interpretation, the yellow squares represent benign experiences with vaccines, and orange squares may represent adverse experiences (perhaps some diagnosis that coincides with vaccination).
One of Epstein’s early results is to show that an agent with no personal adverse experience can act (in our example, this would be someone with no personal reason to be concerned about a vaccine but who nevertheless refuses a vaccine). More surprising still, an agent that is more susceptible to emotional contagion than others (that is, puts more weight on its neighbours dispositions), but which has no adverse experience may even act before those with direct adverse experiences. Epstein suggests that history’s great leaders might just be those people who are most sensitive to others’ dispositions, rather than those with a distinct vision of the world.
Although very simple - most of the examples feature only three agents - Epstein shows that simple agents, if connected in a social network, can generate surprisingly social behaviour. His book makes a convincing case for the importance of psychology and behavioural science in public health, even in understanding apparently complex phenomena.
Overall, Epstein’s model provides a versatile framework that captures some important behavioural phenomena. Some of the specifics, such as the use of learned fear might not be as useful in a public health context. For example, we generally don’t have to learn much to like high calorie foods. Missing features from a public health perspective also include discounting - the way that we tend to undervalue costs and benefits in the future, leading to behaviours like smoking despite overwhelming negative consequences later. But the general approach of modelling competing cognitive components could be adapted to include these features.
A further criticism might be that some of Epstein’s results depend on specific values, such as the weights applied to other agents’ dispositions. These values don’t appear to have any particular basis in evidence. But these results suggest that the phenomena can at least in principle be generated by agents acting in a network.
For some, Epstein’s approach might feel too reductive. There is no room for culture and little for meaning in Epstein’s agents’ simplified world. But I think the point is that we should be cautious about appealing to higher-level concepts like culture to explain things like health inequalities if we can generate something that looks very similar using individual-level effects drawn from the psychological literature.
Personally, I found Agent Zero deeply satisfying. The approach of cognitively plausible agents acting in social networks seems to have a lot of potential for understanding complex problems in public health. Agent based models are under-used for non-communicable diseases [4]. I can imagine an approach drawing on the types of findings presented in Mullainathan and Sharfir’s ‘Scarcity’ [5], or some of the models by Daniel Nettle (who has argued that people who face many risks that they can’t control shouldn’t spend as much time or effort controlling the ones that they can [6]), to develop agent-based models of health inequalities.
Agent_Zero is a fairly technical book: Epstein uses mathematical formulae and formal logic in places. However it is not necessary to follow all the mathematical logic to get a lot from the book. The basic model is simple enough that it should be accessible to readers who are comfortable with regression models (and is considerably less complex than some health economic models).
References
2. Schelling TC. Dynamic models of segregation†. J Math Sociol. 1971;1: 143–186.