Multiobjective multi-UAV path planning via evolutionary multitasking optimization with adaptive operator selection and knowledge fusion

Path planning is crucial for UAV task execution, underpinning effective aerial reconnaissance and precision strikes. An ideal flight path must both minimize travel distance and reduce the risk of enemy detection or destruction. Due to the inherent trade-off between these objectives, multi-UAV path planning is conventionally formulated as a multiobjective optimization problem. However, as the number of obstacles, threats, and UAVs increases, the computational complexity escalates, hindering the generation of optimal path planning solutions via conventional multiobjective optimization approaches. To address this challenge, we model a multiobjective multi-UAV path planning (MOMUPP) problem that simultaneously optimizes flight distance and threat cost, with the latter quantified using line-of-sight theory and terrain occlusion effects. We further construct an auxiliary task that approximates the MOMUPP problem and develop an evolutionary multitasking framework to facilitate effective knowledge transfer between tasks. Building on this framework, we propose the evolutionary multitasking multiobjective path planning (EMMOP) algorithm. EMMOP incorporates a double deep Q-networks-based adaptive operator selection (DAOS) mechanism that dynamically selects the optimal search operators for each task based on the current evolutionary state, thereby generating high-quality offspring. Additionally, a knowledge transfer strategy based on directional information extraction and knowledge fusion (KTDF) enables efficient exchange of critical information between the main and auxiliary tasks. Experiments on 15 benchmark instances across five map scenarios indicate that EMMOP outperforms five state-of-the-art methods, enhancing hypervolume by 2.46% and pure diversity by 28.27%, while generating shorter, safer, and collision-free UAV paths with diverse trade-off solutions for decision-makers. The source code is available at https://github.com/Leopard125/EMMOP.