A control theory based hybrid architecture to anticipate and shape adversarial behavior
Military, business, and political decision makers are faced with intelligent competitors or adversaries who have the power to influence the state of the environment. These decision makers seldom have any direct control or influence over their competitors or adversaries and must account for how these adversaries will react to their decisions. Also, competitors and adversaries in the military, business, and political domains will continuously change strategies to meet their objectives. Often, data in these domains are noisy, incomplete, and contain many structural breaks or break points due to an adversarial strategy shift. Methods are needed to anticipate adversarial actions and reactions to a decision maker’s actions. Once a decision maker is able to anticipate its adversaries’ actions, the decision maker needs tools to determine how to shape the reaction of its adversary and achieve a desirable state in the environment.^ A process to anticipate and shape adversarial behavior is developed in this work that allows a decision maker to meet his/her objectives while accounting for adversarial reaction. This process does not attempt to predict adversarial behavior, but identifies strategies for a decision maker that are as robust as possible to disruptive adversaries. The process for anticipating and shaping adversarial behavior is built upon a novel control framework: the multiple feedback loop controller with internal reference (MFLCIR). By analyzing the transfer function of the MFLCIR, five unique shaping (i.e. control) strategies are identified, limitations in the anticipating and shaping process are discovered, and strategies to overcome these limitations are developed. The limitations are mitigated by controller specifications that result in noise suppression, disturbance rejection, and reduced effects of modeling error.^ The MFLCIR is partitioned into four subtasks that are functional with authentic and available data in the irregular warfare and counterinsurgency domains. The combination of these subtasks and heterogeneous data sources constitutes the hybrid architecture for anticipating and shaping adversarial behavior. The subtasks within this hybrid architecture are: intent, indicators, anticipation model, and shaping. The hybrid architecture’s objectives, algorithms, and subtask linkages are identified from the control-theoretic foundation of the MFLCIR; whereas, other information fusion and command and control (C2) models are based on intuition. Current forecasting efforts of terrorist behavior and international conflict have error rates of approximately 50%. It is shown by simulation, and proven mathematically from the MFLCIR, that active shaping using the hybrid architecture can improve the forecast error by 70%. Intelligent and proactive adversaries have the potential to be very disruptive to the shaping process by causing large deviations in the behaviors or environmental states being shaped. The hybrid architecture can be tuned to mitigate these consequences without reliance on actionable intelligence or knowing beforehand the adversary’s intent as required in cogitative modeling. It is found that by maintaining steady shaping actions these effects caused by disruptive adversaries can be mitigated by more than 30%.^ The hybrid architecture is demonstrated using authentic diplomatic, information, military, and economic (DIME) data from the south-east region of Afghanistan from September 2007 to February 2008. Also, political, military, economic, social, information, and infrastructure (PMESII) data developed by SEAS-VIS (Synthetic Environment for Analysis and Simulation – Virtual International System) is used. From the Afghanistan demonstration, it is found that concurrent and alternating combat actions in different provinces have the result of reducing certain adversarial activity in the Kandahar province while improving social stability.^
Douglas E. Adams, Purdue University.
Military Studies|Artificial Intelligence
Off-Campus Purdue Users:
To access this dissertation, please log in to our