Abstract
This paper introduces a novel framework for selecting optimal transmitter-receiver pairs in distributed MIMO radar systems for target tracking. By modeling the selection as a Multi-Group Combinatorial Restless Multi-Armed Bandit (MG-CRB) problem, the approach accounts for non-stationary channel conditions where signal quality evolves independently of selection. Integrating an Interacting Multiple Model-Unscented Kalman Filter (IMM-UKF) allows for closed-loop feedback, enabling the system to predict future channel states based on target kinematics. Simulation results demonstrate significant improvements in tracking accuracy and Signal-to-Interference-plus-Noise Ratio (SINR) compared to traditional bandit and fixed selection methods.
Introduction
In distributed Multiple-Input Multiple-Output (MIMO) radar systems, transmitters and receivers are geographically dispersed to observe targets from diverse spatial angles. While utilizing all available transmitter-receiver (TX-RX) pairs maximizes data acquisition, it is often computationally prohibitive and resource-inefficient. Consequently, the system must intelligently select a subset of pairs to maximize the Signal-to-Interference-plus-Noise Ratio (SINR) for effective tracking.
The primary challenge lies in the non-stationary nature of the environment. As targets move, the geometric relationship between sensors and targets changes, causing the SINR of specific TX-RX pairs to fluctuate over time—a phenomenon known as "restless" channels. Traditional selection methods often assume static or perfectly known channel conditions, failing to address the exploration-exploitation dilemma in dynamic scenarios.
Methodology: The MG-CRB-CL Framework
The authors propose the MG-CRB-CL (Multi-Group Combinatorial Restless Bandit Closed-Loop) algorithm to address these challenges. The framework treats each TX-RX pair as an "arm" in a restless multi-armed bandit (RMAB) setting, where the state of an arm evolves regardless of whether it is selected.
Key Components
- Restless Bandit Modeling: Unlike standard bandits, the RMAB model allows the SINR to change constantly, reflecting the physical movement of targets.
- Combinatorial Optimization via BPSO: Selecting the optimal combination of M transmitters and N receivers is NP-hard. The system employs Binary Particle Swarm Optimization (BPSO) to efficiently identify the best "super arm" (subset of pairs).
- UCB1 Indexing: To balance exploration and exploitation, the Upper Confidence Bound (UCB1) algorithm assigns a "bonus" to under-sampled channels, ensuring the system periodically checks potentially improved paths.
- Closed-Loop Integration (IMM-UKF): The "secret sauce" of this approach is the integration of an Interacting Multiple Model-Unscented Kalman Filter (IMM-UKF). The tracker provides real-time estimates of target position and velocity, which are fed back into the bandit model to predict future SINR. This kinematics-aware prediction makes the selection process proactive rather than purely reactive.
Experimental Evaluation
The algorithm was evaluated through simulations of three distinct scenarios:
- Scenario 1: Single target with nearly constant velocity.
- Scenario 2: Single maneuvering target with changing speed and direction.
- Scenario 3: Multiple targets with varying motion models.
The simulations utilized a 100 km x 100 km surveillance area equipped with 4 transmitters and 6 receivers.
Results and Key Insights
The MG-CRB-CL algorithm significantly outperformed benchmarks such as epsilon-greedy, standard UCB1, and fixed selection strategies.
- Enhanced Tracking Accuracy: In maneuvering target scenarios, MG-CRB-CL achieved a Root Mean Square Error (RMSE) of 1.17m, compared to 1.31m for standard UCB1 and 1.70m for fixed strategies.
- Superior SINR Maintenance: The proposed method maintained an average SINR much closer to the theoretical "Best Policy" (optimal selection with perfect knowledge) than any other tested algorithm.
- Selection Efficiency: The Average Selection Rate (ASR)—the frequency with which the algorithm identifies the true optimal sensors—was 3 to 6 times higher than benchmark methods.
The core insight is the value of kinematics-aware bandits. By asking "Where is the target going?" instead of just "How good was this channel?", the system effectively manages the "restless" nature of radar channels.
Conclusion
The MG-CRB-CL framework provides a robust solution for sensor selection in distributed MIMO radars. By combining combinatorial bandit theory with closed-loop target tracking, it achieves near-optimal performance in non-stationary environments. Future research will focus on integrating power allocation and waveform design, as well as testing robustness against environmental clutter and electronic jamming.
Summary updated