SBIR/STTR Award attributes
Military operations require fast, decisive, and accurate decision making to accomplish missions with optimal performance and minimization of exposure to risk. Military leaders are forced to make these decisions under constant pressure, changing circumstances, incomplete information, and very short time frames with minimal margin for error. Advancements in Artificial Intelligence (AI) and, specifically, Deep Learning (DL) are helping train computers to accomplish tasks under these same conditions by recognizing patterns across massive data streams. Research in competitive gaming, especially Real-Time Strategy (RTS) games, have shared an interesting symbiosis with the advancement of AI approaches to complex decision-making and hierarchical planning. RTS games are characterized by their imperfect information, large state and action spaces, and necessary balancing of high and low-level planning. The conceptual underpinnings of these game dynamics are highly relevant to problems that are familiar to the military intelligence community, such as knowing how to parameterize plans and when to execute them. Partial observability of environments compounds these difficulties by adding a degree of uncertainty to the mix. However, a question that remains to be answered is how well recent accomplishments in gaming agents, can improve real-world decision-making aids that require situational awareness over high-dimensional observations. DL approaches are making it possible for machines to learn from experiences, adapt to new data, and provide recommendations on optimal behavior. Insights from these models can be communicated to users and analytics providing an intelligence and command advantage. However, the lack of transparency and explainability have made it impossible for these decision makers to put lives at risk based on a black box opinion on the best path forward. In addition, current AI and DL solutions focus on individual decisions based on the current situation with minimal concern for longer term strategic impacts. To solve these two shortcomings, the Machina Cognita Technologies team proposes to develop the State-based Machine Aided Real Time Strategy (SMARTS) engine. The SMARTS engine will provide the ability to analyze an array of potential sequences of actions (or decision tracks), the risks associated with each of these actions, and the required capabilities and effectiveness for units to execute the actions. In addition, the SMARTS engine will provide clear explanations as to the reasoning behind the recommended actions, the impact on mission effectivities, and the possible outcomes for the recommended actions and alternative paths. Specifically, we seek to create a machine learning framework for distilling mission plans into interpretable strategic and tactical decision tracks using a combination of behavioral cloning, unsupervised learning, and attention networks.
