Areas of Expertise

Tactical AI automatically learns to enact a scheduling (resource allocation) action as a function of real-time monitoring of operational parameters such as processor utilization, available data transfer, power consumption, and computing latency by training its DRL model with feedback from other expert-based models or from scratch. Simultaneously, Tactical AI enacts its autonomous rule-based learning engine i.e. A2RL to learn the training environment as a whole so that it can autonomously detect deviations from the norm learned from training during the actual operation.

Tactical AI provides a powerful and realistic autonomous decision-making capability for tactical systems such as ANM based on a unique combination of Adaptive Autonomous Rule-based Learning (A2RL) and Deep Reinforcement Learning (DRL). Specifically, the A2RL of Tactical AI idetects “surprising, unforeseen, anomalous” trends and behaviors of situational dynamics that were not learned by DRL during its training phase, and adapt to them in a just-in-time manner with real-time local re-learning. This makes Tactical AI directly applicable to managing tactical networks which cannot afford to re-learn the policies required to handle the “surprise.”

Tactical AI agents effectively capture a wide range of spectrum holes through shifts in time and frequency via an extensive amount of data from both simulation and real-world tests. These agents will detect unforeseen trends and behaviors of the spectrum usage dynamics that were not learned during their training phase, and adapt to them on-the fly. Then Tactical AI will automatically adapt the available radio control strategies such as control channel rendezvous, frequency selection for data communication, modulation scheme, etc., to optimize the availability of ongoing, dynamic spectrum holes in the local RF landscape.

Building a practical mission-cognitive capability for future MUM-T systems is critical for its viability and effectiveness. Tactical AI naturally enables the concept of bounded autonomous agents with predefined levels of “autonomy” (we refer to this as “Bounded Autonomy” or “Bounding Box for Autonomy”) to ensure that individual Tactical AI agents can adapt to unknown/unplanned changes (“surprises”) while not stepping outside of the bounding box and become completely unpredictable for warfighters. When surprises occur, the Tactical AI agents deployed across the MUM-T assets would analyze the surprises and select alternative actions within the given bounding box and continue the planned mission. With this bounded autonomy combined with the concept of PlayBook, any deviation from the current mission will not just simply lead to a few default fallbacks such as “mission abort” under the degraded communication situations where assets (“players”) have insufficient resources to carry out the remaining missions based on the new situational awareness data (“different Play/Code in the agreed PlayBook”), if any.

Tactical AI builds and maintains comprehensive RF situational awareness knowledge in frequency and time across all wireless communication and networking protocols without computationally expensive DSP algorithms. Tactical AI learns the characteristics of important signals directly from the digitized radio signal (e.g., I/Q samples) or spectral power measurements in terms of the probability of importance with its DL component, and labeling of importance with its A2RL component. Specifically, when one or more new important signals or behaviors are added to the system, Tactical AI not only prevents itself from making a misclassification, but also “remembers” the importance features for successful detection of the same inputs.

Tactical AI is capable of exploiting not only the saliency by ignoring the signals that have been correctly identified as important/unimportant over a short period of time, but also the attention by tuning to the spectrum based on the key “features” learned during training.

Tactical AI is initially trained to observe the current situation and perform the most appropriate cyber-attack, including disengaging an ongoing attack to avoid detection. Then, Tactical AI is further trained to observe a trainer’s responses and adapt its actions to maximize training coverage. The AI Trainer equipped with the trained Tactical AI is now capable of building and storing the profile of a trainee’s response to cyber-attacks in a concise manner, so that it can determine the next course of attacks that are most likely to improve the responsiveness of the trainee.