CHAracterization of Intrinsic Novelty (CHAIN)

This proposal outlines a proposed mathematical framework for the detection of and adaptation to novelty in online AI systems. This framework is founded on two existing theories in which GDA personnel have particular talents: shape analytics and reinforcement learning. Phase I is dedicated to performing the research necessary to develop this framework. Phase II would involve developing the algorithmic design into a computational tool for specific applications that offer promising commercial development and/or future government use. The entire framework is sensible in any context that satisfies the following general assumptions. (1) There is an AI system acting as one or more agents, taking actions in response to observations. (2) Each observation can be evaluated to a “reward” value, providing the agent with positive or negative feedback from its actions. (3) The agent can store and access records of previous observations and actions and resulting post-hoc observations, to develop predictive models. (4) Each observation can be placed in a metric space, so that there is a notion of the distance between any two observations. These assumptions are met by many practical examples, such as autonomous driving, tracking, DCI (detection, classification, and identification), and automated cyberdefense.