.png){kind=icon}
.png){kind=icon}
Return
Dynamic-Weight Multi-Objective Optimization for Multi-UAV Collaborative ISCPT Networks: A Pareto-Based Learning Framework
.png){kind=icon}
PRE
AI
DOI：10.1109/JSAC.2026.3701332.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Uncrewed aerial vehicle (UAV)-assisted integrated sensing, communication, and power transfer (ISCPT) is an emerging multi-functional wireless paradigm. By leveraging UAV’s mobility and flexible deployment, together with the benefits of ISCPT, UAV-assisted ISCPT is able to simultaneously provide sensing, communication, and wireless power transfer services to ground IoT users. In order to satisfy the heterogeneous requirements of ISCPT users with limited resources, multi-objective optimization design becomes essential in UAV-assisted ISCPT. To effectively achieve the trade-offs among different objectives, a common and efficient approach is to assign task weights to the objectives for emphasizing their relative importance. In conventional fixed-weight (FW)-based deep learning (DL) methods, each weight setting requires one well-trained model. Once the weights change, the model has to be retrained. In UAV-assisted ISCPT, weights may vary over time as services progress. If traditional FW-based DL methods are applied, the system has to train and store a large number of neural network models to adapt to various weights, incurring substantial computation and storage overhead. To overcome such limitations, we propose a dynamic-weight multi-objective framework that integrates a Pareto-conditioned graph neural network (PCGNN) with a weight-space multi-partitioning (MP) training strategy (PCGNN-MP) to jointly optimize three-dimensional deployment and power control of UAVs. Particularly, by learning multiple weight-conditioned models through one-time training and constructing a compact set of non-dominated solutions via model-level Pareto filtering, PCGNN-MP is able to rapidly adapt to varying weights without retraining, and it only requires training and storing a small set of models, which is much smaller than the number of possible weight configurations. Simulation results show that PCGNN-MP achieves a Pareto-front hypervolume comparable to that of the FW-based DL methods, while notably reducing computational and storage overhead by about 96% and 90%, respectively.
Keywords:
Graph neural network (GNN)
integrated sensing communication and power transfer (ISCPT)
Pareto-based learning
uncrewed aerial vehicle (UAV)
Journal
IF:
17.2
Papers:
6.4K
Citations:
3.1W
View more organizations