GEOMETRIC DATA ANALYTICS INC. — Department of Defense SBIR Phase II: HR001119S0035-07

Description

US missions rely on robust and reliable deliveries of required supplies, making logistics planning a critical task throughout global operations. These deliveries are time-critical, as failure of on-time delivery (or detection of the supply movement by an adversary) can cause unacceptable tactical or strategic losses. These deliveries must be coordinated from many origins to many destinations via many possible routes. These deliveries may be subject to surveillance or interference, making some otherwise-efficient routes subject to adversarial threat. Today’s interconnected global economy provides an immense multitude of possible combinations of vessels, ports, and routes, all of which are subject to variation or disruption from causes that may be natural, civil, or adversarial. Moreover, these deliveries may be requested on short notice, requiring rapid solutions to this critical task. This project proposes to improve the reliability of time-critical planning systems by developing AI methods that are trustworthy in the face of disruption and designed to maintain robust performance in the presence of both natural and adversarial novelty. This will be accomplished by combining two new company innovations, OUQ robustness certification and ATD encoding, with careful engineering of existing state-of-the-art mathematical algorithms. These will be leveraged, along with professional implementation of problem-specific variants of existing adversarial Reinforcement Learning and graph-search methods to offer a compelling approach to this logistics planning problem. Robustness to novelty (natural, civil, or adversarial) is of particular concern for coordinated time-critical logistics planning. US missions are crippled when planners cannot reliably deliver resources due to unexpected natural events, sudden civil disruptions, or unplanned maintenance. If adversaries can detect and identify delivery routes, then they can introduce new disruptions that appear novel to existing planning systems. Likewise, many commercial applications require time-critical coordination of multiple resources where failure is unacceptably expensive. A study of novelty and robustness in CHAIN Phase-I established that AI systems can be designed via certification for robustness. This certification occurs through a process of verification and validation adapted from optimal uncertainty quantification. An essential step of this process is to examine sensitivity of event threshold via geometric and topological methods to determine margins of uncertainty. This allows smooth degradation of performance when novelties inevitable arise.