Hi James,

Not an answer, but an interesting take on this issue can be found in Manuel de Landa's Philosophy and Simulation. It attempts to reconcile a certain form (admittedly esoteric) of Deleuzian social theory with cellular automata, agent-based models and other kinds of simulations. It arguably performs another kind of reductionism (of agent-based models themselves), but is a useful attempt to "square the circle" of the methodological individualism implied in agent models with broader accounts of the social.

Also a nice article by Joshua Epstein, *Why Model*, and some subsequent responses, examine some of the epistemological issues of ABM from within the field.

http://www.amazon.com/Philosophy-Simulation-Emergence-Synthetic-Reason/dp/1441170286

http://jasss.soc.surrey.ac.uk/11/4/12.html

I do not see why agent-based modelling MUST be reductionist. Any interacting distributed system may produce emergent properties - including when not designed to do so - and these can be fed back to agents giving a dynamic between the social and individual level. Thus research into emergent narrative for example. The reductionist aspect is harder to avoid at initialisation: of course real-world societies never start from a social blank-sheet as new agents are socialised as soon as they are created. This is harder to model but is also being attempted, for example in developmental robotics and  language-creation-from-interaction work. The software agents take on agent-based modelling - which does tend to methodological individualism - is therefore by no means the only one. It is lack of attention to interaction and emergence  that produces this, where work on situated agents or social agents tries to make these  fundamental drivers. Arguably purely top-down direction from social properties is also wrong: there is a dialectic between the social and the individual levels. 

I think not. It does not seem in principle that ABM could not represent entities in which things could reside that were not agents. For example, the "interest rate" posted by banks is both a determinant of individual behaviour and an outcome of borrowing and lending. No doubt the interest rate of Santander does happen to reside in the heads of some individuals but cannot be "reduced" to that.

My biggest worry is not that ABM will let us down in building non reductionist models but that social science will fight on over this even though its methodological positions are muzzy to the point of uselessness. I think we actually need to avoid the philosophy and ask what classes of properties we might want our ABM to have that they don't or risk getting hopelessly bogged down.

An attempt to do this (but for realism rather than reductionism) can be found in a paper Nigel and I wrote called:

Hunting the Unicorn: An Exploration of the Simulation of Small Group Leadership

which is on my academia.edu page.

I think Edmund is making a different distinction, that between world state and individual agent representations of world state.  In order to produce interesting interaction in an agent system, agents really do need perception and action capabilities IMO. Again, not so much thought about that in the software agents community but fundamental in embodied agent work (both graphical and robotic). Clearly this implies a world model to be perceived and acted upon: indeed, a situated agent has to be situated somewhere and interact across the boundary between itself and its environment. So the environment has to be modelled..

Yes, I wondered about this myself but arguably that illustrates my point. Can we find a workable description of reductionism (or any other reified ism) that is solid enough for us to be able to say that ABM must (or need not) follow it? Some sociologists at least would probably not think of the distinction between world state and individual agent representations as being different from reductionism (in methodological individualism for example.)

It is perhaps safe to say that ABM is reductionist -- only certain characteristics of a system are included in its specification and its dynamics.  It is also perhaps safe to say that when ABM is used in ways that aren't `merely experimental` -- that is, used to test certain effects arising from the starting characteristics and dynamics -- that this reductionism melts into the background and becomes an unstated part of the very foundation of the system.  We can get tricked by that if we're not careful. It is absolutely appropriate to suggest that society is a collection of emergent properties, but this does not necessarily mean that the emergent properties of an ABM are a model of Society in the real world.  All of that said, ABM is a tool to think with, and like any such tool, so long as we don't come to believe in it as a truth-making apparatus and instead use it as a way to experiment with bounded subsets in the same way a biologist would experiment using Petri dish populations, we have a very neat device on our hands.  This goes back to the `merely experimental` comment above. 

Hi James,

I would say it is reductionist, if we take each agent as unique unlike the usual formulation of particles in a swarm simulation. It does not contradict social theory since it can be treated as an implementation of methodological individualism-which before the dawn of computers was thought to be nothing more than idealism. If economics be a social science, then ABM provides an avenue to test the ideas of catallaxy. One of its main proponent was Friedrich Hayek whom some think was prescient enough to have thought about and set in place some of the main tenets of agent-based economics before the emergence of agent based computational economics as a field.

Hi James,

As a continuation of my previous answer. Simulation may provide a way to resolve the idiographic-nomothetic debate. By moving the debate from "either-or" to "and-as well". Agents if treated as unique (implemented through some probabilistic mechanism) may be seen as conforming to the idiographic approach in the social science. And if it is conjectured that said agents exhibit a consistent set of behaviors (e.g.utility maximization) we may want to admit a nomothetic influence. So despite my "Yes" on reductionist which entails "No" to "conflict with much of social theory", the presence of nomothetic elements may favor generalizability and therefore facilitates theory-building- another  "No" to "conflict with much of social theory."

Even the term "unique" is not without hazard. Agents in an ABM can be "mundanely" unique in respect to a set of common attributes (i. e. the only 55 year old divorced male farmer) but uniqueness in terms of idiographic-nomothetic debates probably means also with a unique history/memory and perhaps a unique cognitive process pertaining to that. ABM usually do not have the environmental and social richness (or the right social processes like socialisation) to generate such uniqueness (though they probably could in principle.) But as I argue in my "Why Sociology Should Use Agent-Based Modelling" (on ResearchGate and also my academia.edu page in final pre publication form) the reason for not building such models is not technical but methodological. They would be useless "maps as big as the world". Even though social scientists fiercely disagree about how we find social generalisations and what they ought to look like, it is hard to think of people who don't believe in _some_ sort of empirically backed generalisation as social scientists at all. As such, it is not clear what the point of such ABM would be.

A couple more thoughts for this Q/A thread.

If agent-based modelling is reductionist (and acceptably so to at least a substantial minority of social scientists) then a question that must surely be asked is how far the reduction of the social needs to be taken. My impression is that agents in modelling studies typically correspond to groups of people (e.g. a household, a university department) or to actual individual people (e.g. a police office or a hospital patient). But certainly the model could be pitched below the level of the individual e.g. {Freudian: id, ego, superego}, or {Transaction analysis: parent, adult, child} or just {Neuroscience: right or left brain hemisphere, cerebellum etc}. Why not, if going further “down” delivers a more effective model as it surely might?

There is a further and maybe more radical point. If the choice of the level to which the social is to be reduced is to be determined ultimately by the effectiveness of the resulting model, then the choice of level can be handled as a primarily computational matter – the model is a computational structure/process, and the reduction level at which it is best built can be determined by trial and error against effectiveness NOT determined, perhaps not even guided, by social science theory/insight.

Best

Jim

Is this not a matter to be established empirically? To what extent does an agent model based on "self report" of decision processes fail to explain what we observe such that a model can be improved by postulating jealousy or repression for example? (This is a definite advantage of ABM because mathematical models, for example, just don't have the option to "add in" cognitive content and see what difference it makes ceteris paribus.) Most discussions of reductionism I have seen tend to be pointless (like the KISS one) because such issues cannot be settled in the abstract. Having said that, this point is only a programme not a solution. It is far from trivial actually to distinguish ABM on data to say convincingly that repression adds to a model (especially when the default approach to data in ABM seems to be to ignore it.)

Perhaps Edmund and I can agree that there is such a thing as a mathematically precise theory of computational “artificial societies” -- this phrase as used in “Journal of Artificial Societies and Social Simulation”.  This theory (alas, currently underdeveloped!) addresses, for example, the properties an Art Soc must have if it is to display a particular behavioural property, P  – i.e. what alternative sets of restrictions an Art Soc must meet on its structures and processing for P to emerge. The investigation of questions such as this within Art Soc theory can be and arguably should be ENTIRELY A MATH/COMP MATTER.

The theory of artificial societies then stands, amongst other things, as a source of suggestive theorems and structural/processual insights that can help us understand natural societies i.e. help for the social scientists.

Yes?          

Best

Jim

Simply because it has no standard theory. Everything is still under study

And, of course, it has only just occurred to me in the way the question is worded. To what extent should we worry about a contradiction to "social theory" unless that is, itself, empirically backed? Maybe it is the social theorists who, in fact, have it wrong ...

"Reductionism", as opposed to "complexity" refers to the idea that all truth derives from ONE, or at least a reduced system of undisputable truths such as Newton's mechanics or Maxwell's equations. The underlying idea of ABM is abstraction, not reductionism: abstraction refers to simplification by usin models. As a consequence of complexity paradigms, the search for an absolute truth is left aside, instead, knowledge is expressed in terms of models in order to be used i simulations which a priori produce a reduced set of possible scenarios of reality in order to help in decision meking under uncertainty.

Since system-level properties can arise from agent interaction ABM is not inherently reductionist in my view. Analytical categories are not themselves reductionist unless one adds the idea that ALL the properties of a system can be determined by inspecting them individually.

Your question goes to the core of complex systems theory and design. There are two opposing directions in the design of complex systems expressing emergent properties: reductionism, and emergence.

1) Reductionism is present in the design of agents as they encompass only a limited number of variables and reactions to any action/input. The task of the model/agent designer is to localize the most critical variables and interactions that lead to observed emergence.

2) Emergence is what is the goal of your model. Emergent properties produced by the model must explain the studied and described real system.

It sounds easy but the opposite is true. The interplay of the two above declared parts of the model is the part of the design of complex systems models.

I realized that you can study their interplay between reductionism and emergence in this book by Ilachinski, which enables readers to dive deeply into the design of ABM models:

https://www.worldscientific.com/worldscibooks/10.1142/5531

It speaks volumes as you read through it all.