Novel Population-Based Frameworks for Improving Search Diversity and Exploration in Nature-Inspired Optimisation Algorithms

Abstract

This research study proposes three novel population-management frameworks that can be implemented into almost any nature-inspired optimisation algorithm, to address common issues in metaheuristic search such as premature convergence, oscillations, and stagnation through improved diversity preservation and exploration. The work is motivated by wrapper-based feature selection in recognition systems, where high-dimensional datasets yield large combinatorial search spaces and stochastic optimisation can produce unstable or overly greedy feature subsets.
Firstly, an initial group-based augmentation is introduced through the Group-Based Firefly Algorithm (GBFA), which modifies population interaction by forming dynamic groupings at each iteration to promote continued variability in swarm movement. When evaluated on eight classical benchmark optimisation functions, the GBFA outperformed the standard Firefly Algorithm (FA) on six functions and matched performance on the remaining two. Building upon this, two algorithm agnostic frameworks are proposed: the Group-Based (GB) and Cross Group-Based (XGB) frameworks, which modify attraction and movement behaviours at the population level without relying on optimiser-specific mechanics. When implemented into the Bat Algorithm (BA), Firefly Algorithm (FA), and Particle Swarm Optimisation (PSO) for feature selection across 21 UCI datasets, the proposed variants improved mean classification accuracy from 0.79 to 0.82 when implemented into BA, from 0.78 to 0.81 when implemented into FA, and from 0.85 to 0.88 when implemented into the PSO, while maintaining comparable variability across datasets, with standard deviations remaining in the range 0.07-0.10.
Finally, the Dynamic Population Refresh (DPR) framework introduces a stagnation-control mechanism that replaces a proportion of the poorest-performing agents with newly generated candidates sampled from random probability distributions (Gaussian, Cauchy, and Lévy), and demonstrates improved optimisation and feature selection performance when implemented into the Gravitational Search Algorithm (GSA) and Harmony Search Algorithm (HSA). Across the proposed contributions, stability analysis using the Jaccard Index further supports that observed performance improvements are accompanied by more consistent feature subset selection behaviour across repeated stochastic runs.

Date of Award	19 Feb 2026
Original language	English
Awarding Institution	Northumbria University
Supervisor	Wai Lok Woo (Supervisor), Kamlesh Mistry (Supervisor) & Li Zhang (Supervisor)