LLM-Guided Evolution Discovers Near-Optimal Radar Power Allocation

Researchers have introduced AlphaEvolve, an innovative framework that employs large language model (LLM)-guided evolutionary search to autonomously uncover a closed-form solution for power allocation in multi-target tracking. This method encodes complex radar states into features inspired by physical principles, subsequently evolving a concise and interpretable scoring function. This function is then converted into feasible power allocations through a deterministic constraint-satisfying transformation. Comprehensive experiments reveal that the resulting closed-form solution delivers near-optimal tracking accuracy, with an average relative performance loss of just 1.51%, and demonstrates dependable generalization across various scenarios. This approach meets the demands for real-time scheduling, robust generalization, and minimal data reliance in radar resource allocation, a task that has typically involved complex iterative optimization.