At the edge: What it takes to bring fast, robust decision-making to the tactical edge

In today’s battlespace, rapid and informed decision-making is more critical than ever to mission success. But at the tactical edge, where warfighters encounter isolated, unpredictable environments with limited and unreliable communication networks, traditional cloud computing simply can’t process and relay information fast enough, if at all. What warfighters in these environments need is technology that can operate independently from traditional data centers to deliver robust data quickly where and when it’s needed.

“We need edge computing solutions that are fast, efficient and capable of handling complex analytics right at the point of sensor collection,” said Katherine Martin, a chief engineer at Booz Allen who specializes in tailoring machine learning capabilities to establish decision advantage at the tactical edge, speaking at the recent Defense One Genius Machines conference. Tactical edge computing is a way to process data locally in real time. It is completely disconnected from the cloud and does not rely on centralized servers. This type of technology is critical in medical military operations, for example, where warfighters need to take rapid, decisive action. 

“In these environments, there's no time to wait for data to be sent to a data center, be processed and then get sent back. Tactical edge computing puts the decision-making power directly in the hands of the warfighter, wherever they are, even in the most isolated locations,” Martin said.

While crucial, the challenge, of course, is how to deploy this necessary technology in remote, hostile environments.

Delivering insights at the tactical edge

So, how can warfighters overcome the unique challenges presented by the tactical edge to deliver necessary insights quickly and efficiently? According to Martin, it will require the thoughtful and secure integration of multiple disciplines, including hardware, software, machine learning and network engineering.

In terms of hardware, warfighters need devices they can easily carry with them along with their plethora of essential gear. Therefore, ruggedized, lightweight and energy efficient devices are crucial. 

“Every ounce and watt of power matters in a field where portability, durability and logistic resupply are critical for continuous operation,” said Martin. “If we're asking soldiers to add an edge device for advanced machine learning capabilities, it must be lightweight, compact, optimized for power constraints, and, most importantly, effective in enhancing their mission.” For this reason, Booz Allen partnered with Carnegie Robotics to produce CardShark, an advanced compute platform for real-time processing at the edge, in a low size, weight, and power (SWaP) form factor designed to meet mission requirements. 

The software on these devices needs to allow for rapid, split-second decision-making and thus must be optimized for real-time processing. “It must be both lightweight and resilient, capable of running complex algorithms while minimizing resource and power usage,” said Martin. 

Artificial intelligence and machine learning will be necessary to enable the insights warfighters need, which will require machine learning models that can be compressed and optimized to fit with hardware constraints and deployed at the edge. Martin points to techniques like quantization that can help reduce models to the necessary size without sacrificing accuracy, ensuring that the machine learning models run effectively and efficiently at the edge.

One of the greatest challenges edge computing faces in remote environments is also the most crucial: networking.

“We must ensure reliable data exchange from warfighter to warfighter, even when bandwidth is limited or intermittent,” said Martin. “Any additional machine learning capabilities deployed on tactical networks must be carefully balanced with the fundamental need for clear and uninterrupted communication between soldiers and sensors.”

For this reason, it's essential that systems are designed to avoid congesting the network and to allow for continuous operations in integrated environments.