The development of models that replicate complex, non-linear, adaptive, constructive threats poses challenges for training simulation developers. As threats to our nation’s defensive systems mature, our ability to characterize and model them needs to improve so that we are not outmatched.

This paper reports on the progress of our Research and Development (R&D) implementing interdisciplinary methods for Big Data analytics, Artificial Intelligence (AI), and Machine Learning (ML) to capture, characterize, and replicate red player/threat behaviors. We used a progressively complex development and testing framework. First, we selected a pathfinder application that was sufficiently complex to require game play and adaptive behavior on both sides. The paper will discuss our implementation of a Mini-max algorithm, a recursive algorithm used in decision-making, and our rationale for using this algorithm to benchmark the progress of our ML algorithms, with comparisons to others’ implementations cited in relevant research. As we tested our AI against humans, we tailored several AI parameters to make the system easier or more challenging as a human opponent. To capture human behaviors, we record all moves made by both sides, per game, and per event. An event is defined as a discrete period, usually 3 days, where multiple human players’ moves are captured.

Our solution to this challenge required us to train our ML algorithms on the individual and collective behavior of human players at an event. The paper will describe how we have been able to replicate human player/threat behavior for three events, with references to others’ implementation in scientific literature. The paper will also describe logical extensions to problems of interest to the Department of Defense, Department of Homeland Security, and cyber security. Those domains include human, machine, and cyber players and opponents whose behavior can be replicated using our AI/ML methods.

Menu

Military Rank Timothy Hill