⚡ electrical engineering

Behavior-Aware Multi-Objective Action Selection for Cognitive Radar against Cognitive ESM Threats

This paper presents a behavior-aware multi-objective action selection framework for cognitive radars that optimizes mission performance and target estimation quality while minimizing predictability to cognitive ESM threats by employing a "Random Near-Best" policy that balances high reward with reduced behavioral regularity.

Original authors: abbas fadavi

Published 2026-09-11

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: abbas fadavi

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the invisible world of electromagnetic waves, a silent contest plays out between radar systems and the sensors designed to find them. A radar works by sending out pulses of energy and listening for the faint echo that bounces back from a target, like a ship or an aircraft. To do this job well, the radar needs to be smart enough to change its signals on the fly, adapting to weather, interference, or the movement of the target. This is the realm of cognitive radar, a system that learns from its environment to make better decisions. However, there is a catch: every time the radar speaks, it risks being heard by an enemy listening device known as an Electronic Support Measures, or ESM, system. These devices scan the skies, looking for the unique signature of a radar's transmission to identify and locate it. Traditionally, engineers have tried to hide radars by making their signals weak or hard to detect, a strategy called low probability of intercept. But a new kind of threat has emerged: the cognitive ESM. Unlike older sensors that just look for a specific signal, these intelligent systems watch the radar over time, learning its habits and predicting its next move based on patterns in its behavior.

This is the challenge that researchers at the Islamic Azad University of Semnan set out to solve. They asked a difficult question: how can a radar stay effective at its job while also hiding not just its signal, but its very behavior? If a radar always picks the best signal for a given situation, it might become predictable, allowing a smart enemy to guess what it will do next. The researchers developed a new way for the radar to choose its actions that balances high performance with the need to remain unpredictable. They created a system where the radar evaluates its options based on two main goals: how well it can track a target and how hard it is to detect by a standard sensor. But they added a third, crucial layer: they measured how predictable the radar's sequence of choices would look to a smart, learning enemy.

To test this, the team built a detailed simulation involving a radar, a moving target, and a population of one hundred different listening devices, each with its own unique capabilities and detection methods. They defined a radar "action" by four specific settings: how wide its signal spreads in frequency, how often it sends pulses, how many pulses it sends in a row, and the specific shape of the wave it uses. In their simulation, the radar had to choose from 2,400 possible combinations of these settings. The researchers first taught a computer model to predict which of these actions would work best for the mission, using data gathered from thousands of simulated scenarios. This model learned to look at the immediate results of a radar pulse—such as the clarity of the echo and the level of background noise—and estimate how good that choice would be for tracking and stealth.

The core of their discovery lies in how the radar actually picks its next move. The most obvious strategy would be to always choose the single action that the computer model predicts will be the best. However, the researchers found that doing this repeatedly creates a rigid pattern that a cognitive enemy could easily learn and exploit. On the other hand, choosing actions completely at random makes the radar unpredictable but often leads to poor performance in tracking the target. The team proposed a middle path they call "Random Near-Best." In this approach, the radar first identifies the top six actions that are predicted to perform well. Instead of picking the single best one, it randomly selects one of those six. This simple twist keeps the radar highly effective at its mission while introducing enough variety in its behavior to confuse an intelligent observer.

The results of their simulations showed that this approach works remarkably well. When the radar always picked the single best predicted action, it achieved a high level of mission performance, but its behavior was so regular that a cognitive enemy could model it almost perfectly. When the radar chose completely at random, it was very hard to predict, but its mission performance dropped significantly. The "Random Near-Best" strategy struck a near-perfect balance. It maintained 94% of the peak mission performance while making the radar's behavior almost as hard to predict as the random strategy. In their tests, the new method reduced the behavioral regularity that a cognitive enemy could exploit by a large margin, without sacrificing the radar's ability to do its job.

The researchers also developed a way to measure this behavioral stealth. They used four different mathematical models, including advanced neural networks and statistical tools, to analyze the sequence of radar emissions. These models looked for patterns in the timing and structure of the signals, trying to see if the radar was falling into a routine. They found that the "Random Near-Best" policy successfully broke these patterns, keeping the radar's behavior fluid and unpredictable. This work suggests that for modern radars to survive against smart, learning enemies, they cannot just focus on the quality of a single signal. They must also consider the rhythm and variety of their entire sequence of actions, ensuring that even when they are performing at their best, they do not become too easy to read.

Technical Summary: Behavior-Aware Multi-Objective Action Selection for Cognitive Radar

Problem Statement
Cognitive radars adapt their operational parameters (bandwidth, pulse repetition interval, number of pulses, and waveform mode) to optimize mission performance. However, traditional action-selection strategies often focus solely on maximizing instantaneous metrics like Target State Estimation Quality (TSEQ) and Low Probability of Intercept (LPI). This paper identifies a critical vulnerability: repeatedly selecting the highest-performing actions can create temporal regularities in emission behavior. Cognitive Electronic Support Measures (ESM) systems can exploit these patterns to model and predict radar behavior, thereby compromising stealth. The core challenge is to select actions that maintain high mission performance while minimizing the behavioral predictability exploitable by a Cognitive ESM.

Methodology
The proposed framework operates on two temporal scales: the Action Level (individual decision steps) and the Sequence/Policy Level (temporal behavior over time).

Action Space and Simulation:
- The action space consists of 2,400 combinations of four parameters: bandwidth (20–50 MHz), PRI (10–100 µs), pulse count (128–2048), and 12 waveform modes.
- A simulation environment generates radar responses incorporating target dynamics, multipath, clutter, jamming, and noise.
- Offline Training: The system generates a dataset where the "true" mission reward is calculated using ground-truth target states and a population of 100 heterogeneous ESM receivers.
- Online Operation: The radar cannot access ground truth or external receiver data. Instead, it extracts 11 measurable features (e.g., SINR, Doppler spread, spectral entropy) from the received echo to estimate performance.
Metrics and Reward Construction:
- Mission Reward ( $R_{Mission}$ ): Defined at the action level using a Maximin criterion combining two external metrics:
  - $M_{TSEQ}$ : Target State Estimation Quality (based on range/velocity estimation error).
  - $M_{LPI}$ : Conventional LPI performance (based on the proportion of a 100-receiver population failing to detect the emission).
- Feature-to-Reward Mapping: An affine least-squares mapping is trained offline to predict $R_{Mission}$ from the 11 observable features, enabling online decision-making without ground truth.
- Behavioral Stealth Metric ( $M_{CESM}$ ): Defined at the sequence level (over 10 consecutive observations). It evaluates the temporal regularity of the emission sequence using four models: Inverse-Std Baseline, Markov Model, LSTM, and GRU. These models analyze features like Kurtosis, Spectral Entropy, and Crest Factor. The outputs are normalized against two references (Random Action and Best-Predicted Action) to form a composite score where higher values indicate lower predictability.
Action Selection Strategies:
- Action Selector: Several machine learning methods (kNN, Random Forest, Direct Regression) were compared. Boosted Decision Trees (BDT) were selected as the primary selector due to their high mean performance ratio (0.96) and robust minimum performance.
- Policies Evaluated:
  - Random Action: Selects any action uniformly (high diversity, low performance).
  - Best-Predicted Action: Selects the single highest-ranked action from BDT (high performance, high regularity).
  - Random Near-Best: The proposed strategy. It ranks actions via BDT and randomly selects one from the top $K$ candidates.

Key Contributions

Three-Objective Framework: Formulation of a cognitive radar framework that simultaneously optimizes TSEQ, conventional LPI, and behavioral stealth, distinguishing between action-level and sequence-level objectives.
Heterogeneous LPI Evaluation: Development of an LPI model using a population of 100 heterogeneous ESM receivers with varying detection mechanisms (energy, sensitivity, spectral density, etc.), rather than a single receiver model.
Behavioral Stealth Metric ( $M_{CESM}$ ): Definition of a composite metric that aggregates outputs from multiple behavioral models (Markov, LSTM, GRU, Inverse-Std) normalized against reference policies to quantify predictability.
Random Near-Best Strategy: Introduction of a policy that balances mission performance and behavioral diversity by restricting random selection to a set of high-performing candidates ( $K=6$ ).
Offline-to-Online Mapping: Demonstration of a method to estimate mission rewards online using only observable features, bypassing the need for ground truth or external receiver data during operation.

Results

Feature-to-Reward Mapping: The linear mapping achieved a mean absolute relative error of approximately 8% in estimating the mission reward from observable features.
Action Selector Performance: Boosted Decision Trees achieved a mean performance Ratio of 0.96 (relative to an oracle best action) with a minimum Ratio of 0.85, outperforming other standard methods in stability.
Policy Comparison (K=6):
- Best-Predicted Action: Achieved a Ratio of 0.96 but an $M_{CESM}$ of ~0.00 (highly regular/predictable).
- Random Action: Achieved a Ratio of 0.87 and an $M_{CESM}$ of 1.00 (highly diverse/unpredictable).
- Random Near-Best: Achieved a Ratio of 0.94 (a minimal 0.02 reduction from Best-Predicted) and an $M_{CESM}$ of 0.98 (nearly matching the diversity of Random Action).
Parameter Sensitivity: Increasing the set size $K$ from 1 to 6 improved $M_{CESM}$ significantly (0.00 to 0.98) with only a marginal decrease in mission performance (0.96 to 0.94). Further increases to $K=7$ yielded no significant gains.

Significance and Claims
The paper claims that the proposed framework successfully demonstrates that cognitive radar action selection should not focus exclusively on maximizing the reward of individual actions. By adopting the Random Near-Best strategy, the radar can maintain near-optimal mission performance (Target Estimation and LPI) while drastically reducing the temporal regularity that Cognitive ESM systems rely on for behavioral modeling. The study emphasizes that the trade-off between mission performance and behavioral stealth is manageable; a slight reduction in the selection of the absolute "best" action yields a substantial gain in behavioral unpredictability without compromising the radar's primary mission objectives. The work highlights the necessity of evaluating action-selection policies at the sequence level, not just the action level, to counter advanced cognitive threats.

Technical Summary: Behavior-Aware Multi-Objective Action Selection for Cognitive Radar

More like this