Intelligent Planning of UAV Flocks via Transfer Learning and Multi-objective Optimization

Abstract Multiple UAVs have been extensively deployed recently to reduce human workload, resulting in increased automation and efficiency. Path planning of numerous UAVs is a challenging optimization problem and a key component in various applications. Traditional strategies cannot provide accurate, optimal solutions rapidly in complex mission settings. In this context, flocks of birds exhibit intricate patterns of group escape when faced with predators. Local group interactions may lead to the autonomy of these patterns. However, most nature-inspired intelligent planning techniques have slow search speeds and easily fall into local areas. An intelligent planning method emulating the behavior of pigeons to achieve intelligence, safety, and consistency in UAV flocks in a complicated environment is designed. The combinatorial approach of pigeon-inspired optimization and transfer learning (TL-PIO) is the focus of the multi-objective optimization task. On the one hand, path planning and formation control of individual clusters with a dynamic agent are dealt with combinatorial efforts of multi-agent systems (MAS) and flocking model. On the other hand, swapping and synchronization of individual clusters construct flocks in a dynamic environment. Specifically, interaction and swapping positions of the best members among all clusters are involved to plan optimized paths and configure agents in one flock. Experimental results have been validated through a detailed numerical analysis of proposed algorithm over other combinatorial approaches, namely social learning pigeon-inspired optimization (SL-PIO), social learning particle swarm optimization (SL-PSO), and social learning ant colony optimization (SL-ACO). TL-PIO achieves an improvement of 25% over SL-PIO and 18% over SL-ACO in seven test functions and 15% over SL-PSO but only in five test functions. Outcomes reveal the developed approach has the fastest convergence rate and high local optimal avoidance and exploration ability, significantly reducing costs and illustrating supremacy over other methods. The presented work practically implies researchers and practitioners adopt it for distinct benefits in real-world applications.