SBIR-STTR Award

There is a large space of threat data, scenarios, and input variables required for understanding the behavior of Aegis Weapon System (AWS) designs like the Figure of Merit (FOM) algorithm used in the Weapon Control System (WCS) functions for Ballistic Missile Defense (BMD) intercept prediction function and pre-launch coordination. Developing an understanding of performance roll-offs and discontinuities can be exhaustive and time consuming. Machine learning, and the development of deep neural networks, has emerged as a revolutionary tool for the detection, classification, and generation of complex patterns in high-dimensional space. Due to a corresponding increase in both processing power and neural architecture efficiency, the application of machine learning to more restrictive domains have become possible. We propose an automated sensitivity analysis tool, leveraging GFI threat scenario data accessible through an integrated software environment (PULSE) to feed Machine Learning (ML) deep neural networks which generate data allowing FOM performance to be evaluated and categorized. One technique we propose is a Generative Adversarial Neural Network (GAN) that has been used to enhance simulations for subsequent training of anomaly-detection in highly sophisticated simulators for the NOAA GOES satellites. A GAN is a pair of competing neural networks that will find the slow degradation and sudden “anomalies” in performance in an intelligent manner through reinforcement learning. The integration of the ADS developed GAN framework within the SEG PULSE environment to evaluate FOM provides the opportunity to train the neural network in a robust fashion. Approved for Public Release | 20-MDA-10643 (3 Dec

This proposal describes the next phase of our technology development and maturation effort started in Phase 1 of the MDA20-001 Automated Factor-Based Sensitivity Analysis Small Business Innovation Research (SBIR) Topic. The goal of this project is to develop and demonstrate highly automated assessment of Aegis Weapons Systems (AWS) design performance over the full-design space of threats and inputs. Threat data exists in a vast multidimensional space, and understanding how these affect the behavior of AWS designs like the Figure-of-merit (FOM) algorithm used in the weapons control system (WCS) is critical for robust, reliable missile defense (MD) intercept prediction functions and pre-launch coordination. However, developing the comprehensive simulations needed to evaluate this domain with respect to such factors as performance roll-offs and discontinuities can be computationally expensive and time consuming. Simulations currently use complex calculations such as Monte Carlo (MC) methods, which are computationally expensive. Machine Learning (ML) opens up the potential to decrease the time to calculate a solution. In this approach (Developed in Phase I), a neural network model was developed and trained on an unclassified version of the Physics Unlimited scaLable Simulation Environment (PULSE) Monte-Carlo simulator. This model was able to calculate 15 multivariate simulation values for every MC calculation, a performance improvement of approximately 1,600% while maintaining over 97% accuracy. Approved for Public Release | 22-MDA-11102 (22 Mar 22) This speedup allows the generation of large numbers of scenarios to explore the threat space with respect to FOM efficacy, rather than using a few sample simulations. This speed improvement may support capabilities such as Coordination-without-Communication (CWoC). This is a concept that works under the assumption that identical Neural Network models, when presented with shared environmental information, even from different perspectives (e.g. two Aegis ships), can be relied upon to make similar decisions. This means that one ship can infer what another ship is likely to do with respect to defense against multiple targets. If the ship’s computational systems are advanced enough, they can infer what another ship is likely to do in response to a given stimulus and coordinate their actions accordingly. This effort will focus on fully developing this neural network model and producing a cost-effective implementation of it using larger sets of classified data. We intend to produce autoregressive, or predictive models as well as refining our current autoencoding, or interpolating models. This effort will also include detailed documentation of the training process, the underlying theory and justification for the approach, and the generalization capability. As we further develop and refine these models, we will begin to evaluate them with respect to the CWoC concept. Lastly, we will evaluate the Phase II results and prototypes with respect to commercialization. Approved for Public Release | 22-MDA-11102 (22 Mar 22