AI-assisted optimization of multi-agent system parameters: a model of iterative hyperparameter tuning in swarm intelligence algorithms

Abstract

The relevance of the study is driven by the necessity for rapid and safe analysis of territories affected by emergencies, which stimulates the search for innovative approaches in the field of autonomous systems, in particular the use of swarms of unmanned aerial vehicles for terrain scanning. The aim of the article is to present an intelligent multi-agent superstructure for the dynamic adjustment of existing swarm system parameters, based on a combination of Swarm Chemistry self-organization mechanisms and global optimization using an evolutionary algorithm. The task consists in the development of an iterative tuning technology that allows for real-time adaptation of the behaviour of individual agents and the entire formation in response to dynamic environmental changes without interfering with the base structure of the control algorithm. Among the methods employed is an architecture consisting of a set of specialized agents organized into four functional layers, performing continuous monitoring, sensitivity analysis, and mission hyperparameter adjustment, together with a combined methodology of a genetic algorithm and local optimization methods that take into account swarm heterogeneity and the individual scanning widths of each device. The scientific novelty of the work lies in the dynamic tuning system architecture that ensures the survivability and high performance of existing swarm solutions in complex conditions. The practical significance is determined by the significant potential for application in disaster zone monitoring, search for victims, and real-time damage assessment. The results of simulation demonstrated a significant reduction in mission execution time and maximization of the coverage area while simultaneously increasing the stability of swarm behaviour, even under conditions of partial loss of agents or the appearance of unpredictable obstacles. The conclusions point to further development of the model, which involves the integration of machine learning methods for an in-depth analysis of previous mission experience and the expansion of the system for operation in three-dimensional space.