Building Effective AI/ML Training Datasets for BDOC Decision Support

When a security incident occurs on base, BDOC operators have seconds to gain situational awareness of the threat, status and location of responders, data streams from sensors, and to initiate the QRC responses. In the heat of the moment, when a security incident is underway, such as a gate runner breaching the main gate or an active shooter entering the exchange, a BDOC operator has seconds to evaluate the threat posed by the adversary, determine the optimal incident response and then adjust their response dynamically based on new actions made by the adversary, responders, and security systems. Based on work done with the USAF, we know that there is a great need to use artificial intelligence/machine learning (AI/ML) to provide rapid decision support to BDOC operators who are charged to coordinate the response to rapidly evolving security incidents. Speed alone is not enough for AI/ML to improve decision support in the BDOC. The recommendations must also be able to adjust dynamically as conditions on the ground change, and most importantly incident response recommendations must be highly accurate. This proposed Phase I project is focused on the R/R&D for determining the breadth and depth of a modeling and simulation dataset required to train an AI/ML system to provide BDOC operators with accurate decision support during a security incident, and the means by which a BDOC operator can understand the level of confidence that the AI/ML system has in the recommended incident response. The R/R&D in this proposal includes: •    Using a knowledge graph of the dynamic decisions made in the BDOC in response to an incident (i.e. Gate Runner) to evaluate options for building the AI Training Data Set. This work involves analysis of the numerous “states” of an Adversary (location, vehicle, number, weapons, hostile intent, etc.) as well as numerous options available to the BDOC operator based upon ROE, responders, gates, etc. The focus of the analysis is the breadth of simulation and modeling that is required in AVERT Physical Security to analyze the different permutations in order to build a AI Training data set that would support the breadth and depth of decisions that may need to be made in the BDOC. •    R/R&D is also required to determine the confidence that a BDOC operator should have in the dynamically updated QRC. The research involves analysis of several parameters, such as the Adaptive Resonance Theory “Vigilance Parameter” and how it can be used to interpret the results provided by AVERT AI and the confidence that a BDOC operator should have in the incident response recommendation.