An Adaptive Balance Search Based Complementary Heterogeneous Particle Swarm Optimization Architecture

TL;DR

This paper tackles premature convergence in PSO when solving multimodal optimization by addressing underutilization of constructed vectors in cognitive-only variants. It introduces CHxPSO-ABS, a generic architecture combining a complementary heterogeneous PSO (CHxPSO) with an adaptive balance search (ABS) strategy, featuring two update channels that share a common constructed vector and two subswarms to separate exploration versus exploitation. ABS uses a cap limiter, an R& P box, and an adaptive selector to dynamically allocate vector usage between exploration and exploitation as the search progresses. Empirical results on CEC benchmark suites demonstrate improved generalization, convergence accuracy, and stability across varying function types and computational budgets, validating the approach and its embedding of CLPSO/cognitive-only vector constructions.

Abstract

A series of modified cognitive-only particle swarm optimization (PSO) algorithms effectively mitigate premature convergence by constructing distinct vectors for different particles. However, the underutilization of these constructed vectors hampers convergence accuracy. In this paper, an adaptive balance search based complementary heterogeneous PSO architecture is proposed, which consists of a complementary heterogeneous PSO (CHxPSO) framework and an adaptive balance search (ABS) strategy. The CHxPSO framework mainly includes two update channels and two subswarms. Two channels exhibit nearly heterogeneous properties while sharing a common constructed vector. This ensures that one constructed vector is utilized across both heterogeneous update mechanisms. The two subswarms work within their respective channels during the evolutionary process, preventing interference between the two channels. The ABS strategy precisely controls the proportion of particles involved in the evolution in the two channels, and thereby guarantees the flexible utilization of the constructed vectors, based on the evolutionary process and the interactions with the problem's fitness landscape. Together, our architecture ensures the effective utilization of the constructed vectors by emphasizing exploration in the early evolutionary process while exploitation in the later, enhancing the performance of a series of modified cognitive-only PSOs. Extensive experimental results demonstrate the generalization performance of our architecture.

Figures (7)

• Figure 1: Generic architecture CHxPSO-ABS.
• Figure 2: The correspondence between initial swarm, subswarms and slArchive.
• Figure 3: Schematic diagram of topology construction and exemplar selection in our architecture. The blue-labeled circles represent active particles. (a) Topology. (b) Exemplar selection.
• Figure 4: The number of employed particles during the evolutionary process number of iterations (up). The combination of particles employed on different iterations (down).
• Figure 5: Trend plots of swarm diversity indicator. (a) Cognitive-only PSO and CHpPSO-ABS on the Sphere function. (b) Cognitive-only PSO and CHpPSO-ABS on the Rastrigin function. (c) CLPSO and CHCLPSO-ABS on the Sphere function. (d) CLPSO and CHCLPSO-ABS on Rastrigin functions.