[Paper] A Primer on Evolutionary Frameworks for Near-Field Multi-Source Localization

Source: arXiv - 2603.07676v1

Overview

The paper presents a new family of model‑driven evolutionary algorithms for locating multiple radio sources that are in the near‑field of an antenna array. By working directly with the continuous spherical‑wave physics instead of discretized grids, the authors sidestep the major drawbacks of classic subspace methods (e.g., MUSIC) and data‑hungry deep‑learning solutions. Two concrete frameworks—NEMO‑DE and NEEF‑DE—are introduced and shown to work with arbitrary array layouts without any labeled training data.

Key Contributions

• Continuous‑model evolutionary localization: First to formulate near‑field multi‑source localization as a continuous optimization problem tackled by evolutionary computation.
• Two complementary frameworks:
  • NEMO‑DE (Near‑field Multimodal Differential Evolution) – sequentially extracts sources by minimizing a residual least‑squares cost.
  • NEEF‑DE (Near‑field Eigen‑subspace Fitting DE) – jointly estimates all sources by aligning the model‑based array response subspace with the measured signal subspace, handling large power imbalances.
• Grid‑free and data‑free: No need for angle‑range discretization, pre‑trained neural nets, or labeled datasets.
• Array‑agnostic design: Works with any antenna geometry (linear, circular, irregular) because the spherical‑wave model is used directly.
• Algorithm‑agnostic backbone: While Differential Evolution (DE) is used for illustration, the frameworks can plug‑in other evolutionary strategies (e.g., CMA‑ES, PSO).
• Extensive simulation study: Demonstrates that the proposed methods outperform MUSIC and recent deep‑learning baselines across a range of SNRs, source counts, and array configurations.

Methodology

1. Signal Model – The received baseband vector ( \mathbf{y} ) is expressed as a sum of spherical‑wave contributions from each source, each parameterized by a 3‑D location (range, azimuth, elevation) and a complex amplitude.
2. Population Encoding –
   • NEMO‑DE: Each individual in the evolutionary population encodes the location of one source.
   • NEEF‑DE: Each individual encodes the full set of source locations simultaneously.
3. Objective Functions –
   • NEMO‑DE: Minimizes the residual least‑squares error after subtracting the contribution of already‑found sources. A spatial‑separation penalty forces the algorithm to pick distinct sources.
   • NEEF‑DE: Minimizes a subspace‑fitting metric (e.g., the Frobenius norm between the signal subspace from the data and the subspace spanned by the hypothesized source steering vectors). This naturally balances sources with very different powers.
4. Evolutionary Search (Differential Evolution) – Standard DE operators (mutation, crossover, selection) evolve the population over generations. The algorithm stops when the objective converges or a maximum iteration count is reached.
5. Sequential vs. Joint Extraction – NEMO‑DE runs sequentially: after a source is estimated, its contribution is removed from the data residual and the next source is searched. NEEF‑DE solves for all sources in one shot, which is more robust when source powers differ dramatically.

Results & Findings

• Accuracy: Both evolutionary frameworks achieve sub‑meter localization error in challenging low‑SNR and high‑dynamic‑range settings, outperforming traditional MUSIC that suffers from grid quantization errors.
• Robustness to Power Imbalance: NEEF‑DE retains high accuracy when one source is much stronger than the others, a situation where NEMO‑DE’s sequential subtraction can mis‑attribute residual energy.
• Computational Load: On a standard laptop (Intel i7, 16 GB RAM), NEMO‑DE converges in ~0.4 s for 3 sources, while NEEF‑DE takes ~0.7 s for the same problem—still well within real‑time constraints for many sensing applications.
• Array Geometry Flexibility: Experiments with linear, rectangular, and irregularly spaced arrays show negligible performance loss, confirming the geometry‑agnostic claim.

Practical Implications

• Wireless Positioning & Tracking: Deployable in indoor localization systems (e.g., factories, warehouses) where devices are often in the near‑field of dense antenna panels.
• Radar & Sonar: Enables high‑resolution multi‑target detection without costly grid‑based beamforming or massive labeled datasets.
• IoT Edge Devices: The algorithm’s modest computational footprint makes it suitable for edge processors that need to localize nearby emitters (e.g., interference mitigation, asset tracking).
• Antenna‑Design Freedom: Engineers can experiment with non‑uniform or conformal arrays (e.g., on drones or wearables) without redesigning the localization algorithm.
• Data‑Privacy Friendly: Since no training data is required, the method sidesteps privacy concerns tied to cloud‑based deep‑learning localization services.

Limitations & Future Work

• Scalability to Very Large Source Counts: Joint optimization (NEEF‑DE) becomes computationally heavier as the number of sources grows; hybrid schemes or hierarchical DE may be needed.
• Sensitivity to Model Mismatch: The spherical‑wave model assumes ideal propagation; multipath, blockage, or calibration errors can degrade performance—future work could integrate robust or adaptive modeling.
• Real‑World Validation: The paper’s results are based on simulated data; field trials with hardware‑in‑the‑loop will be essential to confirm robustness under practical impairments.
• Alternative Evolutionary Strategies: Exploring more sample‑efficient optimizers (e.g., CMA‑ES, Bayesian Evolutionary Algorithms) could reduce runtime further, especially for high‑dimensional joint searches.

Bottom line: By marrying a physically accurate near‑field model with the flexibility of evolutionary computation, the authors deliver a grid‑free, data‑free, and geometry‑agnostic solution that pushes the state‑of‑the‑art in multi‑source localization. For developers building next‑generation positioning, radar, or interference‑management systems, these frameworks open a practical path to high‑precision, real‑time localization without the heavy baggage of large training sets or rigid array designs.

Authors

• Seyed Jalaleddin Mousavirad
• Parisa Ramezani
• Mattias O’Nils
• Emil Björnson

Paper Information

• arXiv ID: 2603.07676v1
• Categories: cs.NE, eess.SP
• Published: March 8, 2026
• PDF: Download PDF