Decision Assistant for Battle Management (DABM)

The Missile Defense System (MDS) requires rapid, high-stakes decision-making to improve Probability of Engagement Success, minimize Interceptor Wastage, and extend the fight. Deep Reinforcement Learning (DRL) has demonstrated superhuman performance in real-time strategy with AlphaStar and aerial combat with AlphaDogFight. DRL's strength in fast paced and high dimensional scenarios can revolutionize Battle Management - skillfully defending against large scale raids in real-time. nou Systems (nSI) will follow the proven AlphaDogfight roadmap, first achieving superhuman performance in simulation before integrating with real systems. The DRL agent will be trained with offline-to-online methodology, first using data collected in advance (the "offline" phase) and then fine-tune it with real-time interactions (the "online" phase). This minimizes costly online training and while still utilizing online fine-tuning to mitigate the impacts of distributional shift. Our nouDefense simulation accurately captures BM/C2 tradeoffs, supports classified parameters, and enables high-run counts (>1M runs < 1 week) that support state-of-the-art RL. nouDefense provides full kill-chain execution, sensor cueing, and radar resource degradation - ensuring a robust and operationally relevant solution. With the tight feedback loop from a fast-running simulation, the Phase II will both improve the DRL neural architecture, demonstrate superhuman level performance on a high fidelity simulation, and define the probably large data requirements for the integration with the real system.