Cognitive Anti-Jamming for UAV Communications: A Spectrum State Abstraction Empowered Online Learning Approach

Abstract

This paper studies the anti-jamming channel selection problem for multiple uncrewed aerial vehicle (UAV) communications in dynamic environments with multiple jammers. Considering the distributed and adversarial nature of the scenario, we model the multi-UAV channel selection as a distributed long-term decision-making problem, aiming for each UAV to avoid jamming signals while preventing interference with other UAVs’ communications. To enhance online learning and decision-making efficiency, this paper makes the following key contributions: 1) a time-frequency spectrum state (TFSS) abstraction algorithm that groups states with similar anti-jamming decision impacts into clusters, significantly reducing the state space and improving learning efficiency; 2) a multi-agent knowledge integration framework that aggregates distributed abstraction results into a shared TFSS classification network via supervised learning, enabling cooperative policy learning without explicit coordination; and 3) a lightweight online decision-making framework. It uses a trained value feature extractor resilient to dynamics and performs online policy learning with lightweight Q-learning, allowing quick adaptation to new jamming patterns without re-abstraction. Theoretical analysis shows the abstraction has bounded performance error. Benefiting from decoupling value feature learning from policy learning, simulation results demonstrate that the proposed approach can efficiently learn anti-jamming policies and quickly adapt to dynamic and complex jamming environments. It outperforms benchmark algorithms while maintaining lower computational complexity.