Tuesday, October 2, 2012

Domain of agent-based modeling methods

Agent-based modeling is an intriguing new set of tools for computational social science. The techniques permit us to project forward the system-level effects of a set of assumptions about agent behavior and a given environment. What kinds of real social phenomena are amenable to treatment by the techniques of agent-based modeling? David O'Sullivan and his co-authors offer an assessment of this question in their contribution to a valuable recent handbook, Heppenstall et al, Agent-Based Models of Geographical Systems. (Andrew Crooks and Alison Heppenstall provide a valuable and clear introduction to ABM methodology in their contribution to the volume.)

O'Sullivan and colleagues offer a basic taxonomy of different applications of ABM research.
  • simple abstract models where the focus is on exploring the collective implications of individual-level decision making. 
  • more detailed [accounts that] locate virtual model agents in a representation of the real world setting of interest. Typically, such models operate at a regional or landscape scale
  • some of the most ambitious models aim at detailed ... representations of both the geographical setting and the processes unfolding in that setting (111-112)
This taxonomy depends on the degree of abstraction and realism that the model aspires to.

Here are a handful of research projects that are amenable to these techniques, most of which are illustrated in the Heppenstall volume.
  • Land use patterns in peasant agriculture 
  • Residential patterns -- urban and rural
  • Patterns of burglaries
  • Occurrence of interpersonal violence in civil war
  • Traffic patterns -- pedestrian and vehicular
What do the clear examples have in common? They are situations where a number of independent individuals react to a social and natural environment with a set of goals; and they are usually situations where individuals influence each other through their actions. These are situations of dynamic interactive choices. O'Sullivan and colleagues put these points this way:
We consider the most fundamental characteristics of agents in spatial models to be goal-direction and autonomy.... However, more specific definitions of the concept may add any of flexibility, ‘intelligence', communication, learning, adaptation or a host of other features to these two. (115)
(Crooks and Heppenstall provide a similar list: autonomy, heterogeneity, and activity; 87.)

O'Sullivan et al also pose an important question about what the circumstances are where the features of agents makes a difference in the social outcome:
This argument focuses attention on three model features: heterogeneity of the decision-making context of agents, the importance of interaction effects , and the overall size and organization of the system. If agents are the same throughout the system, then, other things being equal, an aggregate approach is likely to capture the same signifi cant features of the system as an agent-based approach.
Essentially the point here is a simple one: if an aggregate outcome results from homogenous individuals making a decision about something on the same basis as everyone else, then we don't need an agent-based model. ABM techniques become valuable when heterogeneous agents interact with each other to bring about novel outcomes.

There are quite a few social situations that do not fit the terms of these models well. Some social processes are not simply the aggregate outcome of choices by a set of independent autonomous agents. For example, the flow of work through an architectural design studio is determined by the rules of the firm, not the independent choices of the employees, and the behavior of an army is largely determined by its general staff and command structure. O'Sullivan et al put the point this way:
A more important question may be, “what should the agents in an ABM of this system represent?” If the interactions among individual actors in the real world are substantially channelled via institutions or other social or spatial structures, perhaps it is those institutions or social or spatial structures that should be represented as agents in an ABM rather than the individuals of which they are formed. (120)
So a general question for ABM methodology is this: where do structural social factors come into ABM models? Here I am thinking of things like a system of regulation and law; a pattern of racialized behavior; the architecture of the transport system; a tax system; .... We might treat these as parameters in the environment of choice for the agents. They are beyond the control of the agents and are regarded as constraints and opportunities. (This is one place where the framework of "strategic interactive fields" disagrees, since the SIF approach looks at institutions themselves as part of the field of strategic interaction in that individuals strive to modify the rules to their own benefit.)

It seems reasonable to judge that ABM techniques are very useful when we are concerned with phenomena that are aggregates of strategic behavior by individual actors; but they are not pertinent to many of the questions sociologists pose. In particular, they do not seem useful for sociological inquiries that are primarily concerned with the dynamics and effects of large social structures where the behavior of individuals is routine, homogeneous, or largely determined exogenously. These are the circumstances where the premises of the ABM approach -- autonomy, heterogeneity, and activity -- are not satisfied.


1 comment:

Anonymous said...

2 arguments : 1) I don't perceive heterogenity as a premise for ABM. Indeed, the ABM techniques are close from physical fluid modelling, in which case all the elements (such as molecules) are strictly identical. In the economic case, we can think of the matching models as some kind of homogenous ABMs. And results form matching models are far different from those of aggregate models...

2) In any case, you have to define more carefully what you call "homogenous" agents since it's not obvious : one often defines all the agents of her model exactly in the same way, so that, event if the agents belong to different states (and can be thought at a time as different and heterogenous), they are ontologically identical (homogenous). As an example, think of the 1994 Arthur's model of 100 people trying to avoid being more than 60 in a bar. The behavior rules are the same for everyone at the beginning (so that they are homogenous : you don't have "smarter" people capable of second-stage reasoning), but, since they randomly try new rules to forecast affluence (among a common determined set of rules), they evolve in a different way so that, a posteriori, they can be seen as heterogenous. Maybe "self-reinfocement" is a better premise thant "heterogeneity".