PSO-X: Modular Framework for PSO Analysis

• PSO-X framework is a modular, extensible system that designs, synthesizes, and benchmarks particle swarm optimization variants using discrete, swappable components.
• It employs functional ANOVA and hierarchical clustering to quantify module influence and reveal key interactions among algorithm components.
• Rigorous empirical protocols using CEC’05 benchmarks yield actionable guidelines, emphasizing adaptive inertia control and random transformation for optimal performance.

The PSO-X framework is a modular, extensible system for synthesizing and analyzing particle swarm optimization (PSO) algorithms by treating all major design components as discrete, swappable modules. Originating as an empirical and analytical platform, PSO-X supports the automatic generation, benchmarking, and interpretation of PSO variants tailored for continuous optimization tasks. Each algorithm instantiation within PSO-X is defined by a selection of module options across diverse aspects of swarm behavior, enabling systematic study of their contributions—singly and in interaction—to overall optimization performance. Rigorous quantitative approaches, including functional analysis of variance (f-ANOVA) and hierarchical clustering, are employed to extract module influence profiles and cluster problem classes according to shared algorithmic effect patterns (Camacho-Villalón et al., 7 Jan 2026).

1. Modular Design and Architecture

At the core of PSO-X, the algorithmic workflow is constructed by composing a fixed set of design modules, each parameterizing a fundamental aspect of the standard PSO update rule or its extensions. In the studied framework, each variant is characterized by a vector

$\vec{m} = (m_1, m_2, \dots, m_n)$

with $n = 11$ modules (8 varied, 3 fixed in recent empirical studies), and a total of 58 possible options (26 in the main benchmark) (Camacho-Villalón et al., 7 Jan 2026).

Principal modules include:

Module	Function	Option Examples
Velocity Update Rule	Governs $v_{i,t+1}=ω_1 v_{i,t}+ω_2 C_1+ω_3 C_2$	(ω₁, ω₂, ω₃): inertia/cognitive/social scaling
Next-Position Distribution (NPP)	Method for sampling displacements	hypercube, hypersphere, Gaussian
Random Transformation Matrix	Rotates/scales velocity vectors	none, uniform, parameterized
Inertia Control (ω₁CS)	Strategy for updating inertia term	constant, linearly decreasing, self-adaptive
Acceleration Coefficient Control (φCS)	Strategy for cognitive/social terms	fixed, linearly varying, adaptive
Model of Influence	Aggregates neighbors' influence	best-of-N, average, tournament
Neighborhood Topology (Topo)	Defines particle connectivity	global, ring, von Neumann
Perturbation modules	Adds (non-)informative stochasticity	on/off, adaptation strategies

Each configuration is defined as a categorical product of options over the target module set. This modular abstraction allows large-scale, systematic exploration of algorithmic variations.

2. Functional ANOVA for Module and Interaction Analysis

To attribute performance effects to individual modules and specific combinations, PSO-X employs functional ANOVA (f-ANOVA). Given the mapping $g(\vec{m})$ from module composition to performance (e.g., median log-error versus optimum), the f-ANOVA decomposition is:

$$g(\vec{m}) = g_0 + \sum_{i} g_i(m_i) + \sum_{i<j} g_{i,j}(m_i, m_j) + \sum_{i<j<k} g_{i,j,k}(m_i, m_j, m_k) + \dots$$

The variance is partitioned as

$$\mathrm{Var}[g(\vec{m})] = \sum_{U \subseteq \{1,\ldots,n\}} \mathrm{Var}_U,$$

where $\mathrm{Var}_U$ captures the contribution (main or interaction) of module subset $U$. The normalized importance $I_U = \mathrm{Var}_U / \mathrm{Var}[g(\vec{m})]$ quantifies the relative influence.

Practically, $g(\vec{m})$ is estimated using a random forest regressor trained on empirical results from large batches (e.g., 1,424 variants $\times$ 25 problems $\times$ 10 repeats). Marginalization over other modules yields unbiased effect estimations (Camacho-Villalón et al., 7 Jan 2026).

3. Experimental Protocol and Evaluation Benchmarks

PSO-X has been evaluated using comprehensive protocols:

• Benchmark Set: 25 CEC’05 continuous optimization functions, spanning unimodal, basic multimodal, expanded multimodal, and hybrid composition classes, each with $D=10$ and $D=30$ dimensions.
• Configuration Sampling: Cartesian product over 8 modules with 26 options (1,424 variants).
• Repetitions: 10 independent runs per configuration-function pair.
• Evaluation Metric: Median distance to known optimum, as $\log_{10}(\Delta)$; $\Delta$ capped at $10^{-9}$.

This rigorous design delivers sufficient statistical power for robust effect quantification and supports subsequent meta-analyses (Camacho-Villalón et al., 7 Jan 2026).

4. Key Empirical Findings: Module Importance and Problem Clustering

f-ANOVA and hierarchical clustering have yielded several core insights:

• Variability Concentration: The first 5–10 effects (main + pairwise) explain 50–70% of performance variance; triple or higher interactions are weak (<5%).
• Top-ranked Modules:
  • Random Transformation Matrix (rotation/scaling of velocities): main effect explains ~25–35% of variance, and up to 60–70% in rotated multimodal problems.
  • Inertia Control (ω₁CS): main effect covers ~20–30% (higher in noisy/multimodal or high-dimensional cases).
  • Model of Influence: moderate role, especially on noisy or smooth unimodal problems (~10–15%).
  • Pairs: RandomMatrix+ω₁CS and RandomMatrix+ModelOfInfluence are top interactions.
• Problem Clustering: Problem classes cluster according to effect profiles, e.g.:
  • Rotated Rastrigin-like functions: RandomMatrix dominates.
  • Noisy multimodal domains: ω₁CS dominates.
  • Hybrid/nonseparable classes: Both RandomMatrix and ω₁CS are essential.
  • Unimodal/noisy unimodal: ModelOfInfluence is most sensitive.
• Dimensionality Effects: In $D=30$, importance curves flatten somewhat and interactions grow in influence; inertia control is relatively more prominent.

These findings imply that, across a diverse set of problem landscapes, PSO-X performance is largely driven by a small number of core modules and their direct interactions (Camacho-Villalón et al., 7 Jan 2026).

5. Practical Configuration Guidelines

Empirical results provide actionable recommendations for PSO-X deployment:

• Prioritize adaptive inertia control (ω₁CS)—prefer self-adaptive or linearly varying strategies for nonseparable/noisy tasks.
• Enable Random Transformation Matrix—use on rotated, nonseparable, or hybrid-structure landscapes (disable on purely separable problems for efficiency).
• Select Model of Influence per landscape: best-of-N for noisy/unimodal, average for multimodal.
• Budget allocation: Focus tuning resources on RandomMatrix and ω₁CS; other modules have incremental or cluster-specific benefit.
• Portfolio design: A set of 5–6 PSO-X variants differing primarily in rotation and inertia options will robustly cover most CEC’05 classes (Camacho-Villalón et al., 7 Jan 2026).

This suggests that for practical scenarios, a small set of well-chosen configurations provides near-optimal coverage for a wide array of problems.

6. Clustering and Meta-Algorithmic Insights

Hierarchical clustering of effect profiles (using Cosine distance and Complete linkage) enables grouping problem classes by algorithmic sensitivity:

• Number of Clusters: Optimal $k=6$ (D=10), $k=5$ (D=30), Silhouette $\approx 0.6$.
• Interpretation: Assignment to a cluster predicts which subset of modules (and options) will be most influential; these clusters correspond to landscape features such as rotation, separability, or hybrid/global structure.

A plausible implication is that automated meta-optimization systems could exploit the function-to-cluster mapping (potentially via landscape meta-features) to recommend or instantiate high-performance PSO-X variants a priori.

7. Future Directions and Open Scenarios

Several extensions are identified as promising within this modular, data-driven analysis framework:

• Expanding module sets (e.g., population size adaptation, angle-based rotations, multi-swarm architectures).
• Application to new benchmark suites (e.g., BBOB, contemporary CEC sets) to quantify the generalizability of findings.
• Incorporation of deeper landscape features for predictive function-to-configuration mapping.
• Analysis of higher-order interactions ($|U| > 3$), facilitated by sparse modeling techniques, to uncover further synergies in complex module spaces.

Overall, empirical decomposition within PSO-X has demonstrated that algorithmic complexity can be successfully managed, and performance reliably maximized, by focusing on a subset of highly influential modules while largely fixing or omitting others—a finding of both practical and scientific relevance for the development and deployment of modular metaheuristic algorithms (Camacho-Villalón et al., 7 Jan 2026).