What is cognitive electronic warfare?
Part of the Field Guide to Cognitive Electronic Warfare.
Cognitive electronic warfare combines perception, decision-making, and learning to adapt behavior from experience. Its distinguishing feature is the use of feedback to improve an internal model or decision process, rather than the speed of an automated response.
In this guide
When a known signal changes
Electronic warfare (EW) uses the electromagnetic spectrum to gather information, interfere with an opposing system, or preserve the operation of friendly equipment. A receiver supporting these functions must interpret the signals around it. Before it can report what another system is doing, it has to distinguish the transmitter's behavior from changes introduced by distance, interference, and its own receiving equipment.
A threat library helps with this task by storing information about known signals and their sources. When a new observation resembles a stored description, the receiver has a basis for identifying it. That accumulated knowledge is useful because the system can begin with an established interpretation instead of learning every signal from the beginning.
The difficulty arises when the observation has several plausible explanations. A familiar transmitter may have entered a new operating mode. A different transmitter may produce a similar signal. The propagation path may have changed an otherwise familiar waveform. In each case, matching the observation to a stored description leaves a further question: what additional evidence would make the interpretation more reliable?
Cognitive EW addresses this connection between interpreting observations, choosing a response, and learning from what follows. A model represents relationships used to explain observations or predict behavior. The system uses that model in its decisions and learns from later experience to improve the model or decision process. Signal libraries, conventional processing, and learned models can all contribute.
Following one observation through the system
Consider a receiver that is uncertain about the category of an observed emitter, the equipment producing the signal. It begins with a state estimate: its current interpretation of the source and its behavior. That estimate can include several possible explanations with different levels of confidence.
The receiver can use the estimate to decide what to observe next. It might continue observing the same source for longer to determine whether an apparent change persists. The decision affects the evidence it will receive. Later measurements then allow it to retain, revise, or reject its earlier interpretation. This is a feedback loop because the result of one decision becomes information for a subsequent decision.
The system also needs a reason to prefer one observation over another. That reason comes from its objective and constraints. In this example, improving the classification may justify spending more observation time on the source, while another task may require that time elsewhere. The decision process connects the available evidence to the purpose of the receiver's work.
Learning adds a longer-term change to this cycle. If later evidence establishes the correct classification, the system can use the example to improve the model used on future observations. Revising the current estimate changes its interpretation of this encounter. Updating the model changes how it will interpret later encounters.
How learning differs from automation and adaptation
These distinctions become clearer when applied to the same receiver. Automation allows it to carry out detection and classification with reduced human intervention. The procedure can run quickly and consistently while remaining unchanged throughout its use.
An adaptive detector changes its behavior with the conditions. For example, it may adjust a detection threshold as the estimated background noise rises or falls. The adjustment follows a governing rule that can remain fixed even though the threshold changes continuously.
Learning changes a model or decision process using experience. A classifier trained on confirmed examples may learn which signal characteristics help separate categories. Training can take place during development, between operating periods, or while the system is running. The location and timing of training determine how new knowledge reaches the operational system.
An implementation can therefore be automated, adaptive, and learning at the same time. Describing the actual change is more informative than assigning it a label: the receiver adjusts a threshold, revises a state estimate, or updates a classifier, and each operation has a different purpose.
Why the research connects recognition to assessment
The need to handle unfamiliar behavior motivated two Defense Advanced Research Projects Agency (DARPA) programs. Adaptive Radar Countermeasures examined the problem for radars, while Behavioral Learning for Adaptive Electronic Warfare examined communications signals. Both investigated responses to systems whose behavior had not been fully characterized in advance.
In the radar work, recognizing signal properties was an intermediate step. The proposed system also had to develop a response and assess its effect from later observable behavior. Recognition supplied information for the decision; assessment supplied evidence about the decision's outcome. The research therefore connected tasks that are easy to discuss separately but have to exchange information in a learning system.
The quality of the assessment determines what can be learned. If a signal disappears after an action, the receiver has observed a change, but it still has to determine why the change occurred. The next chapter follows that problem from a received measurement to an estimate of another system's behavior.