Embedded Files
This is the content of a Wireless @ VT Seminar talk I gave in April 2023. It's not meant to be comprehensive, but rather, to introduce the reader to the fundamental concepts supporting Cognitive Radar Networks and briefly discuss current research directions.
For a more in-depth analysis of these topics, please check out my publications or reach out to me.
Cognitive Radar Networks (CRNs) are a fundamentally intersectional technology, drawing from several sub-fields.
Radar is clearly foundational, with supporting techniques like tracking Kalman filters and coherent integration of radar pulses.
Cognitive radar is a layer up, borrowing optimization approaches like reinforcement learning (particularly bandit models). Reinforcement learning is not the focus of this post; feel free to check out my reinforcement learning intro post.
Cognitive Radar Briefly
Others (namely my co-author Charlie Thornton) have made great strides in cognitive radar (CR). I might refer to CR as single-node CR to emphasize the fact that CRNs could just be groups of single-node CRs. However, as will be discussed later, CRNs do require some modifications beyond just copy/pasting single-node CRs around a target. Anyways, here's a quick discussion of cognition and one way it has been implemented in single-node CR.
Cognitive technologies have biological inspirations. Drawing mostly from echolocation capabilities of bats and other mammals, researchers have suggested the following definitions of cognition:
"Possessing the ability to learn from continuing interactions with the environment" - Simon Haykin
“Displaying intelligence and modifying operating and processing parameters” - IEEE Standards
Cognitive decision making has been described a number of times over the years. These two cycles shown here are essentially the same, but were defined in difference scenarios. The perception/action cycle just shows this basic structure, where an agent observes an environment and makes decisions.
The OODA loop was first described by an air force officer to indicate how pilots make decisions – his key point was that agents which are able to execute this OODA loop faster will outperform ones who go through it slower.
Our key takeaways should include that notion of efficiency, but also the fact that decision-making in this sort of problem is iterative. Once you make a choice, you then go back and make another.
Channel selection, or more broadly waveform selection, provides a lot of utility to radar systems. In the scenario where a radar has several actions available which perform differently in an environment, an agent which makes “good” decisions has been shown in the literature to outperform less-informed techniques in a lot of cases.
We'll see this example more in this post, so here's some more details:
The action set consists of multiple channels, each providing a different "reward" in the form of signal-to-interference-plus-noise (SINR) ratio.
Optimal channel selection can greatly improve the average SINR experienced by a radar. This directly improves performance.
Machine Learning in Cognitive Radar Networks
The main avenue through which cognition is implemented in modern systems is reinforcement learning. As mentioned above, this post assumes some high-level familiarity with reinforcement learning. Here's the important part: reinforcement learning consists of an agent interacting iteratively with an environment, selecting one of several actions and observing a corresponding reward.
Physically, the purpose of a radar network is to obtain more spatial diversity to observe a target. So, in this figure, we see a target moving along with some velocity.
Each radar node can make an observation of this target, pinging a waveform off of its surface and processing the reflection. Of course this is a simplification – the environment may contain multiple targets as well as clutter (anything that isn’t a target). So, the returned signal will contain a bunch of information that the receiver must process. We’ll just image this gets reduced down to a single vector of the target’s estimated position and velocity.
Differently positioned radar nodes can be thought of as drawing these measurements from a distribution, with a variance that is impacted by such things as the true target range, angle, and apparent cross section.
These measurements can then be combined by some fusion center that has access to node position and measurements. As you might expect, it’s been shown in the literature that such a network can obtain higher quality estimates than angle single radar node could acquire.
The term centralized describes networks which fuse measurements centrally, and the central coordinator also controls all of the nodes. This is a good model – the amount of processing power is limited and concentrated to one single spot. There’s downsides of course - having a single processing center causes the network to be more vulnerable to damage.
The second major geometry discussed in the literature are distributed networks, where each node handles its own processing, and the only thing handled centrally is data fusion. This is the more typical “network of monostatic radars” that is considered more resilient. Of course, now we need to have signal processing at each radar node, but this can still be cheap considering todays electronics.
Decentralized networks are those with nothing handled centrally, using ad-hoc message-passing to share observations. These are interesting too, but are not the focus of this work.
Centralized
▪Tracks fused centrally.
▪All nodes operated by central coordinator.
Distributed
▪Tracking at each node.
▪Fusion center for global aggregation.
Decentralized
▪No central fusion.
▪Ad-hoc networking to share observations.
Network Cognition Models
Distributed Cognition
Observations from individual nodes are combined at a fusion center but no feedback is provided to nodes.
Centralized Cognition
A central coordinator is the only cognitive agent, collecting information from each node and dictating future actions.
We discussed earlier how OODA loop and PAC models help us understand what cognition could look like in a system.
Haykin further provides these definitions to illuminate the ways in which a network could be considered cognitive. These really describe where cognitive loops are occurring – where the decision making agents are.
The first is “distributed cognition”, where each node has its own perception-action cycle – there is cognition at each node.
Centralized cognition on the other hand envisions some central node as the only cognitive agent; the nodes are just sensors and make no “decisions”.
Hakyin described centralized control and cognition as “troublesome” due to delayed feedback – i.e., a network being controlled on a pulse-to-pulse level would have additional latency due to the necessity to transmit observations to the coordinator and then to return feedback to the nodes.
Page updated
Google Sites
Report abuse