The latest issue of the Journal of Defense Modeling and Simulation 9, 2 (April 2012) is now out. Most of it is devoted to technical discussions of “Resuability, Interoperability and Composability in Air Warfare Simulations,” but it does also feature an interesting and well-written piece by John Sokolowski, Catherine Banks, and Brent Morrow on “Using an agent-based model to explore troop surge strategy.”
In October of 2001, the United States invaded Afghanistan and replaced the Taliban government. Since its overthrow, the Taliban has pieced together and waged an insurgency to retake Afghanistan, and that insurgency has gained momentum and grown in strength while the United States/North Atlantic Treaty Organization (NATO) effort shrank in size to about 55,000 troops in 2007. A wide range of factors contributed to the insurgency, ranging from socio-cultural to economic to political. This research applied an in-depth study of Afghanistan to an agent-based model to determine if a military troop surge emphasizing a focused security effort could be successful in battling the growing insurgency within Afghanistan. An agent-based model was created and validated against the strategy and situation on the ground in Afghanistan that existed in 2007. Three experiments were conducted representing surges of 50%, 200%, and 400%. The results indicated that a surge of 200% or greater of the existing size force would be necessary to reduce the size of the insurgency, but that a surge of only 50% (50,000 more troops) would not bring about any significant changes as compared to the existing strategy. These model results provide insight into the potential success of various sized troop surges in Afghanistan that implement a focused security effort.
The piece is, unfortunately, behind a paywall, so you’ll need a subscription to JDMS to access the while thing. The core political-military dynamics of their model, however, are captured in the diagram on the right (click to enlarge). These in turn provide the context for the pseudo-tactical model, in which the insurgent and coalition agents fight it out, with detection ranges, a version of the usual Ph and Pk (probability of hits and kills), and probabilities of collateral damage (which in turn affect local attitudes) all modelled. Unlike some of the work done in the technical M&S field, the piece is written in language that is likely to be clear and accessible to those working in very different, non-quantitative areas.
A number of questions might be raised about the model that the authors have developed. One could endlessly quibble about the key variables they have identified, and in some cases whether the relationships always have the directional values they impute to them (for example, deployment of the Afghan National Army—and even more the highly corrupt Afghan National Police, which they don’t model—can sometimes have negative effects on local attitudes, in cases where they are either seen as abusive and predatory, or because they attract Taliban attacks in areas that might otherwise be quiet). However, those criticisms hold true for any game design, and in general my own general reading of conflict dynamics in Afghanistan suggests that quite a bit of it sounds intuitively right.
The authors do validate their model, using open source reporting of changes in Taliban numbers and adjusting the model until it fits the historical record. I’m not sure that their estimates of insurgent “density” are robust enough to provide much validation, however. Moreover, to increase calibration they manipulate only a few of the variables and relationships in the model in order to provide a match against this single indicator. To my mind, that doesn’t provide very strong validation of the underlying model itself.
The simulation attempts to draw conclusions about the relationship between an increase in coalition troop strength in Afghanistan (“the surge”) and the strength of the insurgency. In this, the authors are refreshingly realistic about the limits of agent-based modelling in illuminating policy questions (emphasis added):
The purpose of this study was to provide a means of assessing if the implementation of a military troop surge designated toward a focused security effort strategy might reverse the trend of the growing insurgency in Afghanistan. The strategy using the United States/coalition/Afghan National Army troop strength of about 101,000 soldiers has failed to defeat or even stop the growth of the Neo-Taliban insurgency. This research sought to add some insight into whether or not a surge with a specific role could work within Afghanistan.
…The results of these experiments indicated that a surge of 400,000 or 200,000 troops will reduce the size and strength of the insurgency, but a surge of 150,000 troops would not. These results are not definitive or absolute, but give insight into the possible outcomes of a surge of the given size based on a model built using careful research. This research represents a tool for analysis in the decision process to determine if a surge should occur. It is not the answer to the question of whether a surge would be effective.
In my view, however, they’ve both overstated and understated the value of their analysis. Given the great many assumptions built into the model, I’m even more doubtful than they appear to be that the outcome of the experiment provides useful policy guidance. On the other hand, I think they could do far more to highlight the potential contribution of the experiment as a heuristic device—that is, as a way of helping decision-makers think about a large, complex, wicked problem. As Gary once put it, the article would be even more interesting for a broader audience interested in insurgency and counterinsurgency if there was less seer and more sage in its approach to the material. The model might offer some insight, for example, in why a limited surge might not work; what key indicators and metrics might be useful in assessing the effectiveness of increased coalition troop strength; or even what variables or nodes seems to have an especially important effect on outcomes. In other words, I think the article would be all the more interesting if rather than simply reporting experimental results, it also highlighted what the construction of the model itself may suggest about conflict dynamics (or our understanding of conflict dynamics) in Afghanistan. It would have also have been useful to report some of the more detailed simulation findings about how particular variables changed under different coalition troop strengths, or which relationships other than troop strength seemed to be most important to outcome.
Still, for the many readers of PAXsims who are interested in such issues but are rarely exposed to either agnet-based modelling or work in the M&S community on political-military issues, it is certainly worth a read.