I’m drilling down into theory about the computational modeling of social systems. In just a short amount of time trying to take this task seriously, I’ve already run into some interesting twists.
A word about my trajectory so far: my background, such as it is, has been in cognitive science and artificial intelligence, and then software engineering. For the past several years I have been training to be a ‘data scientist’, and have been successful at that. This means getting a familiarity with machine learning techniques (a subset of AI), the underlying mathematical theory, software tooling, and research methodology to get valuable insights out of unstructured or complex observational data. The data sets I’m interested are as a rule generated by some sort of sociotechnical process.
As much as the techniques of data science lead to rigorous understanding of data at hand, there’s been something missing from my toolbox, which is the appropriate modeling language for social processes that can encode the kinds of implicit theories that my analysis surfaces. Hence the transition I am attempting to go from being a data scientist, a diluted term, to a computational social scientist.
The difficulty, navigating as I am out of a very odd intellectual niche, is acquiring the theoretical vocabulary that bridges the gap between social theory and computational theory. In my training at Berkeley’s School of Information, frequently computational theory and social theory have been assumed to be at odds with each other, applying to distinct domains of inquiry. I gather that this is true elsewhere as well. I have found this division intellectually impossible to swallow myself. So now I am embarking on an independent expedition into the world of computational social theory.
One of pieces that’s grounding my study, as I’ve mentioned, is Cederman’s work outline the relationship between generative process theory, multi-agent simulations (MAS), and computational sociology. It is great work for connecting more recent developments in computational sociology with earlier forms of sociology proper. Cederman cites interesting works by R. Keith Sawyer, who goes into depth about how MAS can shed light on some of the key challenges of social theory: how does social order happen? The tricky part here is the relationship between the ‘macro’ level ‘social forms’ and the ‘micro’ level individual actions. I disagree with some of Sawyer’s analysis, but I think he does a great of setting up the problem and its relationship to other sociological work, such as Giddens’s work on structuration.
This is, so far, all theory. As a concrete example of this method, I’ve been reading Epstein and Axtell’s Growing Artificial Societies (1996), which I gather is something of a classic in the field. Their Sugarscape model is very flexible and their simulations shed light on timeless questions of the relationship between economic activity and inequality. Their presentation is also inspiring.
As a rule I’m finding the literature in this space far more accessible than I would have expected. It’s often written in very plain language and depends more on the power of illustration than scientific terminology laden with intellectual authority. What I have encountered so far is, perhaps as a consequence, a little unsatisfying intellectually. But it’s all quite promising.
Based on these leads, I was recommended David Little’s recent blog post about complexity in social science. He’s quite critical of the bolder claims of these scientists; I’d like to revisit these arguments later. But what was most valuable for me were his references. One was a book by Epstein, who I gather has gone on to do a lot more work since co-authoring Growing Artificial Societies. This seems to continue in the vein of ‘generative’ modeling shared by Cederman.
But Little references two other sources: John Holland’s Complexity: A Very Short Introduction and Miller and Page’s Complex Adaptive Systems: An Introduction to Computational Models of Social Life.
This is actually a twist. Holland as well as Miller and Page appear to be concerned mainly with complex adaptive systems (CAS), which appear to be more general than MAS. At least, in Holland’s rendition, which I’m now reading. MAS, Cederman and Sawyer both argue, is inspired in part by Object Oriented Programming (OOP), a programming paradigm that truly does lend itself to certain kinds of simulations. But Holland’s work seems more ambitious, tying CAS back to contributions made by von Neumman and Noam Chomsky. Holland is after a general scientific theory of complexity, not a specific science of modeling social phenomena. Perhaps for this reason his work echoes some work I’ve seen in systems ecology on autocatalysis and Varela’s work on autopoiesis.
Indeed the thread of Varela may well lead to where I’m going. One paper I’ve seen ties computational sociology to Luhmann’s theory of communication; Luhmann drew on Varela’s ideas of autopoeisis explicitly. So there is likely a firm foundation for social theory somewhere in here.
These are fruitful investigations. What I’m wondering now is to what extent the literatures on MAS and CAS are divergent.