Applications of Complex Systems to Social Sciences

I've recently returned from the GIACS summer school in Poland: Applications of Complex Systems to Social Sciences. Whilst not a social scientist, I am interested in the incorporation of aspects of human/social behaviour into models of the physical environment and its change. I thought this summer school might be an opportunity to get a glimpse at what the future of modelling these systems might be, and how others are approaching investigation of social phenomena.

The set of lecturers was composed of a Psychologist, three Physicists (P1, P2, and P3), a Geographer, and an Economist. I'm sure plenty of 'real social scientists' wouldn't be too happy with what some of these modellers are doing with their differential equations, cellular automata, agent-based models and network theory. One of the students I spoke to (a social psychologist) complained that these guys were modelling social systems but not humans; another (a computer scientist interested in robotics) suggested the models were too 'reactive' rather than 'proactive'. Pertinent comments I think, and ones that made me realise that to really understand what was going on would need me to take a step back and look at the broader modelling panorama.

Some of the toughest comments from the school attendees were levelled at the Geographer's model (or "virtual Geography") that attempts to capture the patterns of population growth observed for European cities, using a mechanistic approach based on the representation of economic processes. The main criticism was that the large parameter space of this model (i.e. a large number of interacting parameters) makes the model very difficult to analyse, interpret and understand. Such criticisms were certainly valid and have been previously observed by other modellers of geographic systems. However, the same criticisms could not be levelled at the models presented at the physicists' (and psychologist's) models, simply because their models have far fewer parameters.

And so this, I think, is the one of the problems that the social psychologist and cognitive scientist alluded to; the majority of the models arising from the techniques of physics (and mathematics) are generally interested in the system properties as whole and not individual interactions and components. One or two key state variables (a variable used to describe the state of the system) are reported and analysed. But actually, there's nothing wrong with this approach because of the nature of their models, based as they are on very simple assumptions and largely homogenous in the agents, actors and interactions they considered.

Such an approach didn't settle well with the social psychologist because the agents being modelled are supposed to be representative of humans; humans are individuals that make decisions based on their individual preferences and understandings. The computer scientists didn't want to know about broad decision-making strategies - he wants his robot to be able to make the right decision in individual, specific situations (i.e. move left and survive not right and fall off a cliff). Understanding broad system properties of homogenous agents and interactions is no good to these guys.

It's also why the Geographer's model stood out from the rest - it actually tries to recreate European urban development (or more specifically, "Simulate the emergence of a system of cities functionally differentiated from initial configurations of settlements and resources, development parameters and interaction rules"). It'a a model that attempts to understand the system within its context. [One other model presented that did model a specific system within its context was presented by the Economist's model ("virtual archaeology") of the population dynamics of the lost Kayenta Anasazi civilisation in New Mexico. This model also has a large parameter space but performed well largely (I'd suggest) because it was driven by such good data for parameterisation (though some parameter tuning was clearly needed).]

So no, there is nothing wrong with an approach that considers homogenous agents, actors and interactions with simple rules. It's just that these models are more divorced from 'reality' - they are looking at the essence of the system properties that arise from the simplest of starting conditions. What is really happening here it that the systems that have not be modelled previously because of the problems of quantitative representation of systems of 'middle numbers' (i.e. systems that have neither so many system elements and interactions that statistical mechanics is not useful, but have more elements and interactions than allows simple modelling and analysis) are now being broken down for analysis. The attitude is "we have to start somewhere, so lets start at the bottom with the simplest cases and work our way up". Such an approach has recently been suggested for the advancement of social science as a whole.

This still means our "virtual Geographies" and "virtual Landscapes" will still be hampered by huge parameter spaces for now. But what about if we try to integrate simple agent-based models of real systems into larger models of systems that we know to be more homogenous ('predictable'?) in their behaviour. This is the problem I have been wrestling with regarding my landscape model - how do I integrate a model of human decision-making with a model of vegetation dynamics and wildfire. From the brief discussion I've presented here (and some other thinking) I think the most appropriate approach is to treat the agent-based decision-making model like the physicists do - examine the system properties that emerge from the individual interactions of agents. In my case, I can run the model for characteristic parameter sets and examine the composition (i.e. "how much?") and configuration (i.e. "how spatially oriented?") of the land cover that emerges and use this to constrain the vegetation dynamics model.

So, the summer school was very interesting, I got to meet many people from very different academic backgrounds (physicists, mathematicians, computer scientists, cognitive scientists, psychologists, sociologists, economists...) and discuss how they approach their problems. I think this has given me a broader understanding of the types and uses of models available for studying complex systems. Hopefully I'll be able to use some of this understanding of different techniques in the future to good effect when studying the interaction between social and environmental systems.

The complex systems approach does offer many possibilities for the investigation of social systems. However, for the study of humans and society this sort of modelling will only go so far. We'll still need our sociologists, 'human' geographers, and the like to study the qualitative aspects of these systems, their components and interactions. After all, real people don't like being labelled or pigeon-holed.