Epstein and Axtell, in Growing Artificial Societies, get all hyperbolic about the impact agent-based modelling has on 'Explanation'.

For many, explanation and prediction are the same thing, and one can see why: to ask any more of explanation is impossible. Explain gravity! Well, here are some lovely predictive rules. Yeah, but what is it, really? Well... here are some lovely rules? Explanation also seems intuitively at home with reductionism: why are there so many different types of atom? Coz they're formed from smaller exchangeable parts - look, here's how it works.

But Epstein and Axtell reckon they have a new form of explanation in agent-based models: it's 'generative explanation'. Epstein has a whole book on this one idea, but it also appears in their much earlier stuff just mentioned:

What constitutes an explanation of an observed social phenomenon? Perhaps one day people will interpret the question, 'can you explain it?' as asking 'can you grow it?' (Epstein & Axtell 1996 p.20)

They reckon this is a new 'generative' approach to social science - providing initial conditions in the form of a set of rules that are 'sufficient to generate' emergent phenomena. The bumph for Epstein's book says:

After elaborating this notion of generative explanation in a pair of overarching foundational chapters, Epstein illustrates it with examples chosen from such far-flung fields as archaeology, civil conflict, the evolution of norms, epidemiology, retirement economics, spatial games, and organizational adaptation.

Coo. I often have digs at books I haven't read on this blog, so it's worth noting that maybe the foundations of the notion are stronger in the later book - one would hope they are: why write a whole book otherwise? But there's a big, fat problem. Well, two of them actually.

The first: something I was reading today noted that one reason the scientific method is so successful is that science works, bitches. It keeps planes in the air and gets men to the moon, allowing them to play golf in low gravity, etc. Reductionism is also enormously useful, since it allows us to work out how things work. What use is a generative social science? It may expand our knowledge, but can it do any more? Maybe it can - wouldn't want to rule it out. Neither would I want to start measuring everything we do by use-value.

But the second problem is a killer for me. An example: the Schelling segregation model is the daddy of all generative agent-based models. It has everything you want: it is beautifully simple, and from a very few rules can produce a profoundly interesting emergent behaviour. It shows that it's possible for people to have only a slight preference for neighbours 'like them' for areas to segregate into different racial ghettoes.

The problem comes when the modeller pats themselves on the back, powers down, goes home and puts their slippers on: 'ah, a good day's work - I've 'explained' the phenonemon of segregation, using the generative model of explanation. Go me.'

But Debbie Phillips, here at Leeds Geography School, has found evidence that Leeds estate agents have overtly steered ethnic minorities to certain areas, and may actually encourage white flight - perhaps, as the article linked suggests, because its in their economic interests, but perhaps also because they're racist fuckheads. (Not that discrimating on economic grounds stops you from being a racist fuckhead, of course... oh, but that would mean accepting only highly-skilled immigrants would make the government racist fuckheads. Surely that can't be right...?)

So has the Schelling model explained anything? No: it has suggested that there's a possibility for micro rules to lead to segregation. That's not some high-falutin' new kinda 'hexplan-hation', that's a hypothesis.

(And an odd one at that: Schelling suggests that, in a world of squares where people move one square at a time, segregation may happen. Does that look anything like the real world? No. But then that's a problem agent-based models have: their attempt to look even slightly similar to the real world damns them. Analytic models do not have that problem, though their abstraction may distance them even further from reality.)

Which is to say: the model is still useful, but it's potentially very dangerous to make over-the-top claims about it's explanatory power - because that encourages back-patting and slipper-donning, when the researcher should be thinking, 'well, this model's helped me come up with an idea. What should I do now to test it?'

Even more scarily, the Schelling model does what a lot of economics models do: they're appeals to nature. The world is mathematically / naturally prone to segregation, so - well, we don't think it's good or bad, but it's the way things are. It's a neutral truth, and there's no point trying to fight it.

Economics much more forcefully binds society in this way, but this little Schelling-Phillips example is a nice way to outline the problem. Now I should probably read Epstein's book, see if he's addressed all these problems in minute detail already...