SOCIAL: Social network optimization algorithm via centrality and influence-aware learning

We present SOCIAL (Social Network Optimization Algorithm via Centrality and Influence-based Learning), a
structure-aware metaheuristic that reframes black-box engineering optimization as social-network analysis on a
small-world graph of candidate solutions. Each solution is a node, and edges specify local neighbor interactions;
information flow is governed by an influence-diffusion score that combines structural centrality (between
ness/bridge potential) with relative fitness, enabling agents to preferentially learn from solutions that are both
well-positioned in the network and high-quality in the search space. A time-scheduled learning policy shifts
from network-driven exploration toward elite-guided exploitation, with adaptive mutation and periodic popula
tion synchronization to prevent stagnation while preserving diversity. This networked view yields interpretable
search dynamics—identifying leaders, followers, and critical bridges—together with scalable communication over
sparse graphs. We assess SOCIAL on 23 benchmark functions and six constrained engineering design problems
(gear train, pressure vessel, welded beam, speed reducer, composite laminate, and FGM beam), demonstrating ro
... mehr
bust performance and competitive ranking-based evaluation against contemporary optimizers. SOCIAL achieves
particularly strong results on multimodal, discontinuous, and constraint-dominated problems, while maintain
ing stability and feasibility rates. The algorithm’s network-based architecture makes it particularly well-suited
for materials science and cheminformatics applications, where candidate structures can be modeled as nodes
in similarity graphs, enabling optimization of graph-model hyperparameters, sampling policies, and candidate
exploration strategies in materials design. The key novelty is using the social-network structure itself as the
learning mechanism, providing a general and explainable optimizer for engineering problems where gradients
are unavailable, objectives are nonconvex or noisy, and variables are mixed discrete–continuous.