Machine Learning for Defense Applications

Towards the goal of building a robust (protecting from surprises and anomalies) and intuitive (enabling seamless human-AI interaction) AI solution with a broad range of military’s autonomous system command and control applications, we propose to design & develop a novel hybrid AI architecture coined Tactical AI which uniquely combines three powerful machine learning techniques: Surprise-Based Learning (SBL) [1], Generative Hierarchical Symbolic Reinforcement Learning (GH-SRL) [2], and Deep Learning (DL) [3]. The core novelty of the Tactical AI is to be able to simultaneously learn the structure of underlying data (discover generating functions, grammars, relationships, and not simply classification) in an unsupervised manner, and be taught by a human in an intuitive way when data is sparse, unavailable, or surprising. Specifically, Tactical AI will enable the AI and humans to seamlessly exchange knowledge (prior knowledge, laws of physics, common-sense rules, etc.) and intent (commander’s planned actions, etc.) while preserving the benefits of learning directly from observational/experimental data, simulated data, and other sources of prior knowledge. If successful, the resulting Tactical AI technology will drastically reduce the time and cost it requires to train future unmanned military systems, while significantly increasing their autonomous decision-making capability consistent with commander’s intent on the field.