[Paper] Quadratic Surrogate Attractor for Particle Swarm Optimization

Source: arXiv - 2603.17163v1

Overview

The paper introduces a fresh twist on Particle Swarm Optimization (PSO) by swapping the classic “global‑best” attractor for a quadratic surrogate attractor. By fitting a simple n‑dimensional quadratic surface to the swarm’s recent positions, the algorithm obtains a better‑conditioned convergence target that reflects the local shape of the objective landscape. The authors show that this modest change yields consistently higher success rates across a wide suite of benchmark functions, especially on quasi‑convex problems, while adding only a tiny computational cost.

Key Contributions

• Quadratic surrogate attractor: Replaces the traditional global‑best (gBest) with the minimum of a locally fitted quadratic model, providing a smoother, more informative pull for particles.
• Minimal overhead: The surrogate is built from the current swarm positions, requiring only a small linear‑algebra solve per iteration (O(n³) with n = dimensionality, negligible for typical PSO dimensions).
• Extensive empirical validation: 400 independent runs on each of several benchmark functions, with statistical analysis of performance distributions.
• Robustness to premature convergence & noise: The surrogate’s curvature information helps the swarm escape shallow local minima and tolerate stochastic perturbations.
• Special advantage on quasi‑convex landscapes: Demonstrated superior convergence speed and accuracy where the underlying function behaves like a convex bowl.

Methodology

1. Standard PSO recap – Each particle updates its velocity based on its personal best (pBest) and the swarm’s global best (gBest).

2. Surrogate construction – At every iteration, the algorithm gathers the current positions of all particles and fits an n‑dimensional quadratic form

   [ f_{\text{sur}}(\mathbf{x}) = \mathbf{x}^\top \mathbf{A}\mathbf{x} + \mathbf{b}^\top \mathbf{x} + c ]

   using least‑squares regression.

3. Attractor extraction – The minimum of this quadratic (found analytically by solving (\nabla f_{\text{sur}} = 0)) becomes the surrogate attractor (\mathbf{x}^*).

4. Velocity update – The gBest term in the PSO velocity equation is replaced with (\mathbf{x}^*). The rest of the PSO dynamics (inertia, cognitive component) stay unchanged.

5. Evaluation protocol – The authors run both the classic PSO and the surrogate‑augmented version on a diverse set of benchmark functions (unimodal, multimodal, noisy, quasi‑convex). For each configuration, 400 independent runs are performed, and performance metrics (best‑found value, convergence speed, success rate) are statistically compared.

Results & Findings

Across all tested functions, the surrogate‑augmented PSO achieved statistically superior results (p‑value < 0.01). The most dramatic gains appeared on quasi‑convex problems, where the quadratic model could capture the global curvature almost perfectly.

Practical Implications

• Plug‑and‑play upgrade: Existing PSO codebases can adopt the surrogate attractor by adding a lightweight quadratic fitting step—no redesign of the core swarm dynamics is needed.
• Better out‑of‑the‑box performance: For engineers tuning hyper‑parameters on new optimization problems, the surrogate version reduces the need for delicate inertia or learning‑rate schedules, as the attractor already guides particles toward promising regions.
• Noise‑tolerant optimization: In real‑world scenarios (e.g., hyper‑parameter search on stochastic training pipelines, control parameter tuning on noisy hardware), the surrogate’s smoothing effect can yield more reliable convergence.
• Scalable to moderate dimensions: Since the quadratic fit scales with the number of dimensions, the method remains practical for typical PSO use‑cases (≤ 30–50 dimensions). For very high‑dimensional problems, a sparse or low‑rank surrogate could be explored.
• Potential for hybrid meta‑heuristics: The surrogate attractor concept can be combined with other swarm variants (e.g., constriction factor PSO, multi‑swarm systems) or even with differential evolution, offering a new avenue for robust hybrid solvers.

Limitations & Future Work

• Dimensionality ceiling: The quadratic fit becomes expensive and potentially ill‑conditioned in very high dimensions (> 100), limiting direct applicability.
• Assumption of locally quadratic behavior: For highly irregular landscapes where a quadratic model poorly approximates the true surface, the surrogate may mislead the swarm.
• Static surrogate update frequency: The paper updates the surrogate every iteration; adaptive schedules (e.g., update only when swarm diversity drops) could reduce overhead further.
• Extension to constrained problems: The current formulation handles unconstrained optimization; integrating constraint handling (penalties, repair operators) with the surrogate attractor remains open.

Future research directions suggested by the authors include exploring low‑rank or kernel‑based surrogates for high‑dimensional spaces, dynamic update strategies based on swarm diversity metrics, and cross‑pollination with other population‑based algorithms to assess broader applicability.

Authors

• Maurizio Clemente
• Marcello Canova

Paper Information

• arXiv ID: 2603.17163v1
• Categories: cs.NE, eess.SY, math.OC
• Published: March 17, 2026
• PDF: Download PDF