On the learning patterns and adaptive behavior of terrorist organizations

The threat to national security posed by terrorists makes the design of evidence-based counter-terrorism strategies paramount. As terrorist organizations are purposeful entities, it is crucial to understand their decision processes if we want to plan defenses and counter-measures. In particular, there is evidence that terrorist organizations are both adaptive in their behavior and driven by multiple objectives in their actions. In this paper, we use insights from learning theory and compare several different reinforcement learning models regarding their ability to predict terrorist organizations’ actions. Using data on target choices of terrorist attacks and two different objectives (renown and revenge), we show that a total reinforcement learning with power (Luce) choice probabilities and information discounting can be used to model the adaptive behavior of terrorist organizations. The model renders out-of-sample predictions which are comparable in their validity to those observed for learning in laboratory studies. We draw implications for counter-terrorism strategies by comparing the predictive validity of the different models and their calibrated parameters. Our results also offer a starting point for studying the convergence process in game theoretic analyses of conflicts involving terrorists.