pith. sign in

arxiv: 2605.01794 · v1 · submitted 2026-05-03 · 📡 eess.SP · cs.AI

Discover Fast Power Allocation Solution for Multi-Target Tracking via AlphaEvolve Evolution

Zhenkang Hou , Wenqiang Pu , Junkun Yan , Rui Zhou , Hongwei Liu This is my paper

Pith reviewed 2026-05-09 16:50 UTC · model grok-4.3

classification 📡 eess.SP cs.AI
keywords power allocationmulti-target trackingradar resource managementclosed-form solutionevolutionary searchreal-time optimizationsignal processing
0
0 comments X

The pith

An evolutionary search can derive a closed-form power allocation rule for multi-target radar tracking that matches iterative solvers to within 1.51 percent accuracy loss while running over a thousand times faster.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper tries to replace slow iterative optimization with a fast, closed-form formula for dividing limited radar transmit power among several moving targets. It encodes the radar observations into a small set of physically motivated features, then evolves a compact scoring function whose outputs are mapped to valid power levels that always sum to the total available power. A sympathetic reader would care because current optimal solvers are too slow for real-time use in changing environments, so a reliable closed-form substitute would make high-performance tracking practical on ordinary hardware. If the result holds, it shows that symbolic discovery methods can produce simple, general rules for resource allocation that avoid the need for repeated numerical search.

Core claim

The paper claims that its discovered closed-form power allocation solution, obtained by encoding radar states into features and evolving a scoring function that is then transformed into feasible allocations, achieves an average relative performance loss of only 1.51 percent in tracking accuracy, generalizes reliably across diverse scenarios and different numbers of targets, and runs more than three orders of magnitude faster than conventional iterative solvers.

What carries the argument

The compact scoring function evolved from physically inspired features of radar states and converted into valid power allocations by a deterministic transformation that enforces the total-power constraint.

If this is right

  • Real-time multi-target tracking becomes feasible on platforms with modest computing resources.
  • The allocation rule continues to perform well when the number of targets changes without any retraining or retuning.
  • Tracking accuracy stays close to the theoretical optimum across a wide range of radar environments.
  • Similar closed-form solutions can be sought for other real-time optimization tasks in signal processing.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The simple form of the discovered rule may point to underlying physical relationships in radar resource allocation that iterative solvers hide.
  • Adding more radar features could further reduce the small accuracy gap in highly cluttered or rapidly changing scenes.
  • The large speedup suggests the approach could be embedded in mobile or low-power radar systems for continuous operation.

Load-bearing premise

The chosen features of the radar states must contain every essential dynamic, and the evolved scoring function must keep working for target counts, operating scenarios, and noise levels that were never seen during the search process.

What would settle it

Applying the closed-form allocation to fresh simulation runs that use target counts or noise conditions outside the range used during evolution and observing a relative performance loss well above 1.51 percent would falsify the generalization claim.

read the original abstract

Efficient radar resource allocation is a fundamental yet computationally challenging problem, as optimal solutions typically require iterative optimization with high complexity. Motivated by the need for real-time scheduling, robust generalization, and low data dependency, this paper proposes a novel paradigm that leverages large language model (LLM)-guided evolutionary search (AlphaEvolve) to autonomously discover a closed-form power allocation solution for multi-target tracking. The approach encodes high-dimensional radar states into physically inspired features, then evolves a compact and interpretable scoring function, which is transformed to feasible power allocations via a deterministic constraint-satisfying transformation. Extensive experiments demonstrate that the discovered closed-form solution achieves near-optimal tracking accuracy (average relative performance loss of only $1.51\%$), reliable generalization across diverse scenarios and target counts, and over three orders of magnitude speedup compared to conventional iterative solvers. These results highlight the potential of LLM-guided symbolic search to revolutionize not only radar resource management but also broader classes of engineering optimization problems.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The manuscript proposes AlphaEvolve, an LLM-guided evolutionary search method to autonomously discover closed-form power allocation solutions for multi-target radar tracking. Radar states are encoded into physically inspired features; a compact scoring function is evolved via symbolic search; and a deterministic constraint-satisfying transformation produces feasible power allocations. Experiments are reported to yield an average relative performance loss of 1.51%, reliable generalization across scenarios and target counts, and >1000x speedup versus conventional iterative solvers.

Significance. If the generalization and performance claims hold with adequate verification, the work offers a promising route to real-time, interpretable closed-form solutions for radar resource management, reducing computational burden while maintaining near-optimal tracking accuracy. The approach of combining feature encoding with evolutionary symbolic discovery could extend to other constrained engineering optimization problems where iterative solvers are too slow.

major comments (2)
  1. [§4] §4 (Feature Encoding and Evolution): The claim of reliable generalization to unseen target counts and scenarios depends on the physically inspired features capturing all essential multi-target dynamics and interactions. No ablation studies, sensitivity analysis, or quantitative bounds on training versus test ranges for target number, noise levels, or scenario parameters are provided, so the 1.51% loss figure may not extend beyond the evolution distribution as required by the central claim.
  2. [§5] §5 (Experiments): The reported 1.51% average loss and 1000x speedup lack details on exact baselines, statistical significance testing, error bars, number of Monte Carlo trials, and the precise distribution of target counts/scenarios used for training versus evaluation. Without these, it is impossible to confirm that post-hoc selection or limited test coverage does not inflate the performance figures.
minor comments (2)
  1. [§3.2] The explicit form of the discovered scoring function and the precise definition of the deterministic transformation step should be stated in the main text or a dedicated table/equation for reproducibility.
  2. [Figures 3-5] Figure captions and axis labels in the experimental results should explicitly state the range of target counts and SNR values tested to support the generalization narrative.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback, which helps improve the clarity and rigor of our claims regarding generalization and experimental validation. We address each major comment point-by-point below, committing to revisions that directly strengthen the manuscript without altering its core contributions.

read point-by-point responses

  1. Referee: [§4] §4 (Feature Encoding and Evolution): The claim of reliable generalization to unseen target counts and scenarios depends on the physically inspired features capturing all essential multi-target dynamics and interactions. No ablation studies, sensitivity analysis, or quantitative bounds on training versus test ranges for target number, noise levels, or scenario parameters are provided, so the 1.51% loss figure may not extend beyond the evolution distribution as required by the central claim.

    Authors: We agree that the generalization claim would be more robust with explicit supporting analyses. In the revised version, we will add: (i) ablation studies removing or perturbing individual physically inspired features to quantify their contribution to multi-target interaction capture; (ii) sensitivity analysis varying target counts (e.g., 2–20), noise levels, and scenario parameters outside the evolution distribution; and (iii) quantitative bounds (min/max/mean) on all training versus test ranges. These additions will directly substantiate that the 1.51% average loss holds under the reported conditions and clarify the limits of extrapolation. revision: yes

  2. Referee: [§5] §5 (Experiments): The reported 1.51% average loss and 1000x speedup lack details on exact baselines, statistical significance testing, error bars, number of Monte Carlo trials, and the precise distribution of target counts/scenarios used for training versus evaluation. Without these, it is impossible to confirm that post-hoc selection or limited test coverage does not inflate the performance figures.

    Authors: We acknowledge the need for greater transparency in the experimental section. The revised manuscript will include: (i) explicit listing of all baselines with their implementations; (ii) number of Monte Carlo trials (1000 per scenario); (iii) error bars as standard deviation across trials; (iv) statistical significance testing (paired t-tests with p-values); and (v) detailed tables showing the exact distributions of target counts and scenario parameters for the evolution/training set versus the held-out evaluation set. This will eliminate any ambiguity about post-hoc selection and confirm the reported speedup and accuracy figures. revision: yes

Circularity Check

0 steps flagged

No circularity in derivation chain

full rationale

The paper's central method is an autonomous LLM-guided evolutionary search over symbolic scoring functions derived from physically inspired feature encodings of radar states, followed by a deterministic constraint-satisfying transformation to power allocations. This process generates a new closed-form expression rather than reducing any claimed result to its inputs by definition or construction. No equations or steps in the abstract or described chain exhibit self-definitional loops, fitted parameters renamed as predictions, or load-bearing self-citations. Performance and generalization claims rest on experimental validation across scenarios, not on tautological equivalence to the evolution inputs. The dependency of fitness on chosen scenarios is a standard methodological detail and does not force the discovered solution to be equivalent to those inputs.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

Based solely on the abstract, the ledger is minimal. The approach implicitly relies on standard radar tracking models but introduces no explicitly stated free parameters, axioms, or invented entities beyond the evolutionary search process itself.

axioms (1)
  • domain assumption Multi-target tracking performance can be adequately captured by standard radar equations and metrics that allow encoding into physically inspired features.
    Required for the state-to-feature encoding step described in the abstract.

pith-pipeline@v0.9.0 · 5474 in / 1419 out tokens · 42139 ms · 2026-05-09T16:50:30.578062+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

24 extracted references · 24 canonical work pages · 2 internal anchors

  1. [1]

    INTRODUCTION Radar resource management is critical for ensuring effective opera- tion of radar systems in complex and dynamic environments [1, 2]. It involves making real-time decisions on how to allocate limited resources, such as transmit power, revisit time, and beam direc- tions, among multiple targets to optimize overall sensing perfor- mance [2, 4–7...

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  2. [2]

    Consider a single-radar system trackingNtargets in a 2D Cartesian coordinate system

    RADAR POWER ALLOCATION FORMULATION Modern radar systems often need to track multiple targets in real time, requiring efficient allocation of limited transmit power among all targets to maximize overall tracking performance. Consider a single-radar system trackingNtargets in a 2D Cartesian coordinate system. At each scheduling instant, the radar allocates ...

    work page

  3. [3]

    First, we extract physically inspired features from the high-dimensional radar state matrices, to construct a compact feature vector for each target in Section 3.1

    POWER ALLOCATION STRATEGY DISCOVERY This section presents the proposed paradigm to discover a closed- form power allocation solution. First, we extract physically inspired features from the high-dimensional radar state matrices, to construct a compact feature vector for each target in Section 3.1. Then, we leverage the AlphaEvolve framework to evolve a cl...

    work page

  4. [4]

    Evaluation is conducted from four aspects: static approxima- tion accuracy, large-scale scenario generalization, computational ef- ficiency, and dynamic closed-loop tracking

    VERIFICATION AND PERFORMANCE EV ALUATION This section constructs a multi-target tracking simulation environ- ment to verify the effectiveness of the closed-form scoring function (12). Evaluation is conducted from four aspects: static approxima- tion accuracy, large-scale scenario generalization, computational ef- ficiency, and dynamic closed-loop tracking...

    work page

  5. [5]

    CONCLUSION This work addresses the computational bottleneck of iterative opti- mization in radar power allocation by leveraging the search capa- bilities of large language models to discover a closed-form, approx- imate power allocation solution. The discovered formula achieves near-optimal performance, with an average relative performance loss of only1.5...

    work page

  6. [6]

    Mallick, B.-N

    M. Mallick, B.-N. V o, and R. Tharmarasa, Eds., Advanced Techniques in Tracking and Sensor Management: Theory and Applications. Wiley-IEEE Press, Forthcoming September 2026

    work page 2026

  7. [7]

    Sensor management: Past, present, and future,

    A. O. Hero III and D. Cochran, “Sensor management: Past, present, and future,”IEEE Sensors J., vol. 11, no. 12, pp. 3064– 3075, 2011

    work page 2011

  8. [8]

    Optimal waveform selection for tracking systems,

    D. J. Kershaw and R. J. Evans, “Optimal waveform selection for tracking systems,”IEEE Trans. Inf. Theory, vol. 40, no. 5, pp. 1536–1550, 1994

    work page 1994

  9. [9]

    Joint Beam Selection and Power Allocation for Multiple Target Tracking in Netted Colocated MIMO Radar System,

    J. Yan, H. Liu, W. Pu, S. Zhou, Z. Liu and Z. Bao, “Joint Beam Selection and Power Allocation for Multiple Target Tracking in Netted Colocated MIMO Radar System,” in IEEE Transactions on Signal Processing, vol. 64, no. 24, pp. 6417-6427, 2016

    work page 2016

  10. [10]

    Joint Transmit Resource Management and Waveform Selection Strategy for Target Tracking in Distributed Phased Array Radar Network,

    C. Shi, Y . Wang, S. Salous, J. Zhou and J. Yan, “Joint Transmit Resource Management and Waveform Selection Strategy for Target Tracking in Distributed Phased Array Radar Network,” in IEEE Transactions on Aerospace and Electronic Systems, vol. 58, no. 4, pp. 2762-2778, 2022

    work page 2022

  11. [11]

    Multidomain Resource Allocation for Asynchronous Target Tracking in Heteroge- neous Multiple Radar Networks With Nonideal Detection,

    C. Shi, Z. Tang, L. Ding and J. Yan, “Multidomain Resource Allocation for Asynchronous Target Tracking in Heteroge- neous Multiple Radar Networks With Nonideal Detection,” in IEEE Transactions on Aerospace and Electronic Systems, vol. 60, no. 2, pp. 2016-2033, 2024

    work page 2016

  12. [12]

    Optimal Re- source Allocation for Asynchronous Multiple Targets Tracking in Heterogeneous Radar Networks,

    J. Yan, W. Pu, S. Zhou, H. Liu and M. S. Greco, “Optimal Re- source Allocation for Asynchronous Multiple Targets Tracking in Heterogeneous Radar Networks,” in IEEE Transactions on Signal Processing, vol. 68, pp. 4055-4068, 2020

    work page 2020

  13. [13]

    Poste- rior Cram´er-Rao bounds for discrete-time nonlinear filtering,

    P. Tichavsk ´y, C. H. Muravchik, and A. Nehorai, “Poste- rior Cram´er-Rao bounds for discrete-time nonlinear filtering,” IEEE Trans. Signal Process., vol. 46, no. 5, pp. 1386–1396, 1998

    work page 1998

  14. [14]

    Resource-efficient range-doppler map generation using deep learning network for automotive radar systems,

    T. Jeong and S. Lee, “Resource-efficient range-doppler map generation using deep learning network for automotive radar systems,” IEEE Access, vol. 11, pp. 55965-55977, 2023

    work page 2023

  15. [15]

    Deep game of escorting suppressive jamming and networked radar power allocation,

    Y . Wang, Y . Gu, Y . Liang et al., “ Deep game of escorting suppressive jamming and networked radar power allocation,” Journal of Radars, vol. 12, no. 3, pp. 642-656,2023

    work page 2023

  16. [16]

    Learning to Optimize: Training Deep Neural Networks for Interference Management,

    H. Sun, X. Chen, Q. Shi, M. Hong, X. Fu and N. D. Sidiropou- los, “ Learning to Optimize: Training Deep Neural Networks for Interference Management,” in IEEE Transactions on Signal Processing, vol. 66, no. 20, pp. 5438-5453, 2018

    work page 2018

  17. [17]

    Progress and perspective on physically explainable deep learning for synthetic aperture radar image interpretation,

    Z. Huang, X. Yao, and J. Han, “ Progress and perspective on physically explainable deep learning for synthetic aperture radar image interpretation,” Journal of Radars, vol. 11, no. 1, pp. 107-125, 2022

    work page 2022

  18. [18]

    Perspective on explainable SAR target recognition,

    W. Guo, Z. Zhang, W. Yu et al., “ Perspective on explainable SAR target recognition,” Journal of Radars, vol. 9, no. 3, pp. 462-476, 2020

    work page 2020

  19. [19]

    Discovering network dynamics with neural symbolic regression,

    Z. Yu, J. Ding, and Y . Li, “ Discovering network dynamics with neural symbolic regression,” Nature Computational Sci- ence, vol. 6, no. 2, pp. 156-168, 2026

    work page 2026

  20. [20]

    SR-LLM: An incre- mental symbolic regression framework driven by LLM-based retrieval-augmented generation,

    Z. Guo, S. Wang, Y . Tian et al., “ SR-LLM: An incre- mental symbolic regression framework driven by LLM-based retrieval-augmented generation,” Proc Natl Acad Sci U S A, vol. 122, no. 52, 2025

    work page 2025

  21. [21]

    A neural symbolic model for space physics,

    J. Ying, H. Lin, C. Yue et al., “ A neural symbolic model for space physics,” Nat. Mach. Intell., vol. 7, pp. 1726-1741, 2025

    work page 2025

  22. [22]

    AlphaEvolve: A coding agent for scientific and algorithmic discovery

    A. Novikov, N. V ˜u, M. Eisenberger et al., “AlphaEvolve: A coding agent for scientific and algorithmic discovery,” arXiv preprint arXiv:2506.13131v1 [cs.AI], 2025

    work page internal anchor Pith review arXiv 2025

  23. [23]

    M. I. Skolnik, Introduction to Radar Systems, 3rd ed. New York, NY , USA: McGraw-Hill, 2001

    work page 2001

  24. [24]

    Bar-Shalom, X

    Y . Bar-Shalom, X. R. Li, and T. Kirubarajan,Estimation with Applications to Tracking and Navigation. New York, NY , USA: Wiley, 2001. A. DERIV ATIONS OF THE BAYESIAN INFORMATION MATRIX COMPONENTS In this appendix, we detail the explicit formulations of the normal- ized measurement information matrixeJd,i (i.e., the information ob- tained per unit transmi...

    work page 2001