Using Complex Adaptive Systems to Simulate Information Operations at the Department of Defense

Irregular Warfare (IW), with its emphasis on social and cognitive phenomena such as population sentiment, is a major new focus of the Department of Defense (DoD). One of the most important classes of IW action is Information Operations (IO), the use of information to influence sentiment. With the DoD’s new focus on IW comes the new need to analyze and forecast the effects of IO actions on population sentiment. Analysts at the DoD traditionally use Modeling and Simulation to analyze and forecast the effects of conventional warfare’s actions on the outcome of wars, but IW and IO in particular are far more complex than conventional physics-based simulations. DoD analysts are in the early stages of looking for scientifically rigorous methods in the Modeling and Simulation of IO’s complex effects. This paper presents the state of IO modeling and simulation in the DoD, using examples from several computer models now being used, in these early stages of IW analysis. It discusses how the ideas of Complex Adaptive Systems (CAS) and threshold events in particular may be incorporated into IO modeling in order to increase its scientific rigor, fidelity, and validity. Complex Adaptive Systems and Validity It is important to retain the principles of scientific rigor as we move from physics-based to social-based modeling. However, identifying these principles in a new application is not a straight-forward task. Many of the principles appear to conflict, because the roadmap to the epistemology of social simulation is still being negotiated. A simulation’s value is in “walking through” what is too complex to be solved analytically. In physics-based simulation, operations research (OR) analysts use conventional simulation to answer questions such as, “How many and what types of airplanes are needed to make it likely that a missing vessel is found?” Simulation makes this determination by walking through rules about the properties of simulated entities over time and space, including probabilities of detection and various other factors in the calculation. If the facts about the resources are put into the simulation correctly, many runs will give a reliable confidence interval for answers to questions like, “Given N of airplane A, will vessel B be found within M hours?” “Walking through” rules about the properties of social entities in particular complex scenarios is the value in social simulation as well. In fact, nonlinearity is an important aspect of a simulation’s epistemology: if a simulation’s output is the direct, obvious and linear result of its input, then it is not finding an answer a question, but instead, it is parroting back simple transformations of the input. However, the output of many IW and IO models currently used in the DoD seems linear and predictable in relation to the input. For example, the United Kingdom’s Peace Support Operations Model (PSOM) is an IW model that simulates a population’s consent to be governed. Analysts have shown concern about linear relations between consent and other factors, making the outcomes of actions easier to predict than they are in real life (Marling, 2009). If a model is too linear, it will not reveal unintended consequences of actions, and won’t even answer any questions that an analyst didn’t have the answers to to begin with. More importantly, it is against the principles of scientific rigor to put the answer to the question in the question itself, and so nonlinearity is actually needed in computer simulation for validity’s sake. However, there is also a tension within the DoD about nonlinearity, because it seems to violate validation principles of transparency and traceability. Transparency and traceability are important to DoD analysts because they need to know why a result occurs and explain it to their leadership. If a simulation is rigged to parrot back the input, then at least it is traceable. Exploring the natural, unrigged implications of properties in new circumstances is more difficult to trace. Historically, the DoD has taken issue with CAS and Artificial Intelligence (AI) “black box” techniques such as neural networks. There is an apocryphal story in the DoD about the use of neural nets to detect camouflaged tanks, that illustrates the reason why. A neural network was used to classify pictures of terrain based on whether they have a camouflaged tank in them or not. It was fed input of pictures of camouflaged tanks taken on a sunny day and pictures without tanks taken on a cloudy day. The neural net mistakenly classified the pictures based on light intensity rather on than on whether they were camouflaged or not (Fraser, 2003). Without knowing why, human judges cannot determine if they trust the automated reasoning behind the technique. Agent-Based Modeling (ABM) is a CAS technique that embodies this validity conundrum. An ABM models entities that have agency, the ability to perceive and react to their environment in an autonomous manner. An ABM exhibits emergent behavior: the properties of entities that are input to the simulation are supposed to recombine and result in patterns that are nonlinear, or more than the sum of their parts. The more you can explore the natural, unintended consequences of the configuration, the more valid it is because it lacks being “rigged,” but at the same time, the less traceable it is. In fact, when true emergence occurs in natural systems, one no longer needs the rules of the lower level system to describe the newly formed upper level system behaviors, because they have become new entities with new rules. For example, one need not refer to the rules of lower level quantum mechanics that apply at the molecular level, when one is describing the physics of bridges, even though bridges are ultimately composed with molecules. Instead, the properties are described with Newton’s laws, which do not reduce to quantum mechanics. Some CAS experts feel that this irreducibility implies that we cannot know the reason for true emergent behavior and that it is untraceable in principle (Bar-Yam, 2004). True, irreducible emergence, also called strong emergence, applies to social emergence as well. As in quantum mechanics vs. Newtonian mechanics in physics, the properties of the micro level in a social system, are different from, are described with a different language than, and are not referred to in the language of, the macro level. In physics and in sociology, the micro level and the macro level are different levels of description, and the macro does not reduce to the micro. For example, microsociology and macro-sociology are two independent disciplines that do not refer to each other. However, bridging the gap between the micro and macro in the social sciences is not quite as difficult as finding a unified field theory to bridge the gap in physics. In fact, in sociology, there is a field of study of how one level of description leads to the other, called “micro-macro integration.” It may be that it is difficult to imagine how the micro and macro influence each other in the social sciences, but it is not untraceable in principle. In fact, an agent-based model is a good way to bring micro-macro integration into our comprehension. If the micro and macro influence each other in ways too complicated for our minds to compute, the computer is able to do that computation for us. We could put the rules of the micro level in, see how macro patterns known to exist are derived, and then test the system using statistical and debugging techniques to tease out which particular sets of micro rules cause the macro patterns and how. Because an agent-based model walks through the implication of the properties of agents in time and space, including the agent’s reaction not only to the scenario but to each other, we are able to examine and describe the process by which the micro derives the macro and vice versa. This process is one that the simulation was not programmed to come up with in advance. If we find this process to exist in the real world, then the agent based model has helped us to discover a theory to test. It may be hard to find out exactly how the micro and macro interact, but the reason is contained “within the box,” inside of a computer program, in which all else may be held the same and cause may be traced. In fact, the cause of any emergent phenomena within an ABM should be traced in order to ensure that it is not an artifact such as occurred in the “camouflaged tanks” neural network example. Once we ensure that the input micro rules, the emergent macro patterns, and the emergent interactions between micro and macro all have fidelity with processes in the real world, then we have reason believe that the theory of Micro Macro Integration expressed by the ABM is valid. By the principle of Ockham’s razor, if a few micro phenomena known to exist derive many macro phenomena known to exist, then it is likely we have chosen the correct micro phenomena. The smallest primordial soup of micro phenomena is the most parsimonious explanation of the macro, and the most likely to be correct. Because Agent-Based Models have a good epistemology, the DoD should not over-emphasize transparency and traceability in their validations, but rather support the development of tools to trace out the cause of emergent phenomena in ABM.