Modeling real-world social complexity in a virtual-world setting could result in a better understanding of emergent behaviors. (AFP)
Paul Cummings is a senior fellow in the ICF International Modeling and Simulation group. / Courtesy Photo
Disturbed by the devastating global conflicts of the early 20th century, Lewis Fry Richardson, a pacifist English Quaker and mathematician, decided to study the causes of war.
He began to examine complex and seemingly chaotic data related to deaths from warfare, and found these deaths had some interesting, almost bizarre statistical regularities. In particular, Richardson found that everything from large wars to much smaller battles, skirmishes and even murders followed a consistent pattern: There were many events with only a few casualties, some with larger death tolls and a very small number with an enormous number of deaths, such as World War II. In other words, doubling the severity of wars leads to a decrease in frequency by a constant factor, regardless of the size.
This rather odd finding was a seminal moment in our understanding of how social interactions that seem chaotic may actually conform to some form of regularity. Enter the field of social complexity science.
Social complexity is the study of social phenomena through the lens of complex systems. This often includes building computational models of social behavior. Social complexity problems range from understanding the spread of disease to political science, economics and game theory. A great deal of research in the field is based on self-organizing agents that interact and form new emergent behaviors.
These agent-based models represent one of the most interesting and advanced approaches for simulating a complex system — something useful for training, simulations and potentially prediction, if the fidelity is good enough. In a social context, the single parts and the whole are often difficult to describe in detail. But by using intelligent agents as basic building blocks, we can study the emergence of social behavior through the creation of artificial societies. These artificial societies are grown in the technology petri dish and model wars, genocide, economics, civil unrest and a host of other social phenomena.
Interestingly, very little social complexity agent modeling finds itself into immersive virtual world training. I’m not speaking about traditional constructive semi-automated forces and artificial intelligence algorithms, but rather cognitive agents acting both independently and collectively in a simulated world, forming new behaviors that emerge in ways that are unique and often unexpected.
These emergent behaviors can help us understand a number of hard social problems, whether we consider complex counterinsurgency operations or try to predict where the next mass atrocity is going to take place. This is all well and good, but without some sort of measurement, our results may simply be chaotic. At worst, negative learning may ensue.
Enter the Power Law
A power law refers to a specific empirical regularity of highly skewed concentration, in which the small values occur often, medium events are less frequent and extreme events are rare but occur with greater frequency than would be “normally” expected.
Although the bell-shaped “normal” distribution is much better known, this pattern of “many-some-rare” occurrences is found throughout the social universe, along with other distributions that follow certain rules. The fact that social power laws exist at all gives character — and some degree of unity — to the social universe across many levels of complexity, including language groups, organizations, cities and the world system.
One of the fascinating characteristics of the power law is scaling. It is a misconception to think that small and large things that correlate to a power law distribution share little or nothing in common. They are all, small and large, part of the same overall pattern, just different ranges governed by the same parameters, as in Richardson’s deaths data.
If we knew our models were designed to produce validation on a small scale, could the results scale upward to thousands or millions? Could we train complex multidimensional thinking on a small scale, and let new ideas emerge that can be scaled into the real world?
Another interesting aspect of power laws is the concept of metastability. This means that some systems can transition to one or more states based on a few simple parameters. We have been studying metastable states in other types of systems, such as earthquakes, where seismic upheavals transition the earth’s plates into any number of metastable states.
Power laws allow the investigation of metastability because they model social situations where a broad range of states — not just the stable state — have the potential of taking place. So we can potentially study social events, determine where the tipping points are and measure what caused an event to go from “looks OK here” to “oh my, what just happened?”
Applying the Rules
So how can we apply social complexity models to virtual world training? Back in the 1990s, a very close friend with a surly Scottish brogue and attitude to match told me in no uncertain terms I’d better read a book called “Ender’s Game” if we were going to remain friends. One of the many interesting aspects of Orson Scott Card’s science-fiction novel was the young protagonist Ender Wiggins’ ability to win increasingly difficult war games that evolved over time. Building games for training, I was always rather enthralled by the idea of creating an evolving war-game tool with complex multivariable outcomes based on any number of decisions made by the player.
More importantly, I wanted to know, can results be measured not only in how effective the individual was in performing his task, but also in what type of variable-state outcomes took place? And could we measure the validity of those outcomes?
After several years of research I came to enter a computational research program that helped me understand the nature of social science, complexity, agent-based modeling and the mathematical formulas that guide them. Although it wouldn’t make a lot of sense to get into the gory details of computational model development, there are a few simple takeaways on developing valid agent models:
■Review the literature. There are a lot of interesting small models that have been developed and tested, including segregation, social networks bargaining, game theory, social norms and epidemics. The attributes of one model might be retrofit to work in a variety of social simulation circumstances. A simple sand pile Self-Organized Criticality model, showing how individual grains of sand may eventually cause an avalanche, can be a new way to model crowd responses.
■Start small. One of the common mistakes of new social complexity modelers is realizing that models are very hard to measure. If we build an agent with a hundred parameters describing his motives, culture, behaviors, etc., the problem space will be extremely difficult to measure. It may seem realistic to include everything that describes human behavior, but in the end, get used to a feeling of confusion on an exponential scale.
■Measure. Use probability models to determine if the social complexity model conformed to something that wasn’t completely chaotic. Not every complexity model will conform to a power law distribution, but many other interesting things may arise.
■Rinse and repeat. Sometimes agent models will seem so interesting, developers may think they are witnessing something that has never emerged in computer modeling. That may be true, but running the model again and again and again can test whether the model is truly consistent over a variety of parameters and circumstances.
Innovations to Come
Although I have socialized the idea to several organizations, I have met with a healthy amount of skepticism. On the one hand, the computational science field isn’t always keen on moving out of the “small models for social research analysis,” and the training world has a bit of a hard time picturing how we can grow artificial societies within virtual platforms that behave like real people.
But this is a very exciting area of discovery. Imagine a new type of virtual world exercise where agents (avatars) are able to exist independently — and in groups, if they choose to join. New groups with new behaviors will grow as the agent collective evolves. AI agents can communicate with one another and make decisions about how to interact with each other and man-in-the-loop players.
Now, imagine a training task developed in a truly socially complex domain. Decisions by the learner can have immediate impacts but can also spread, producing several potentially unintentional multi-order effects. With power law analysis we can investigate the patterns from the disorder, and even help our learners understand how to observe and react to these patterns. These patterns can be applied to organizational modeling, socio-cultural problems, Lean Six Sigma and emergency management crises, to name a few.
And as our understanding of complexity evolves, we can generate harder problems and ask the learner to consider more complicated and contrasting information in order to properly size up a situation. From here, we can develop a library of best practices in complex environments where we can learn and grow as multidimensional thinkers.
We know these are hard problems to solve, but they are by no means impossible. Our understanding of complex problems is growing, evolving and is ripe for innovation. It is important for all of us to recognize that just because it may be hard to see the pattern, it doesn’t mean it’s not there.
Paul Cummings is a senior fellow in the ICF International Modeling and Simulation group. He is currently leading the development of simulation, gaming and social media projects for the Army Research Institute, Navy Bremerton (Post Traumatic Stress Disorder Virtual Worlds), Environmental Protection Agency, Department of Health and Human Services, Army Leadership’s Multi-Source Assessment and Feedback Program, and Air Force Air Combat Command.