ADAPTIVE MARL ALGORITHM FOR UAV SWARM CONTROL UNDER PARTIAL OBSERVABILITY AND COMMUNICATION DISRUPTIONS

Abstract

Abstract. The article addresses the problem of cooperative control of an unmanned aerial vehicle swarm operating under partial observability, unreliable communication channels, packet loss, transmission delays, and incomplete information, all of which significantly complicate coordinated decision-making in dynamic environments. The problem statement section substantiates the relevance of the study by emphasizing the growing use of UAV swarms in monitoring, search-and-rescue, infrastructure inspection, and special missions, where conventional centralized or rigidly preprogrammed control schemes fail to provide sufficient flexibility and robustness. The literature review summarizes recent advances in multi-agent reinforcement learning, including centralized training with decentralized execution, value function factorization methods, recurrent memory architectures, inter-agent communication protocols, training stabilization techniques, and graph-based coordination models, while also identifying the limitations of existing approaches in scenarios with degraded connectivity. The purpose of the paper is to develop and justify an adaptive MARL algorithm for UAV swarm control that improves training stability, resilience to external disturbances, and coherence of collective agent behavior under restricted information. In the main body of the study, the task is formalized within a partially observable multi-agent Markov decision process framework that includes the set of agents, global environment states, joint action space, local observations, transition, reward, and observation functions, as well as internal memory states. The proposed architecture follows the CTDE paradigm and integrates local actor networks, a centralized critic, a communication reliability assessment module, an adaptive policy reconfiguration block, and an LSTM-based collective memory mechanism. A dedicated section describes the communication disruption adaptation mechanisms, namely trust-based weighting of inter-agent messages, local reconstruction of the global state, adaptive switching between cooperation modes, and prediction of neighboring agents’ behavior. Another section presents the learning stabilization mechanism that combines policy regularization, target network smoothing, adaptive learning-rate control, and a shared experience replay buffer. The experimental section reports simulation results obtained in a specialized multi-UAV flight environment with sensor noise, packet loss, delays, dynamic obstacles, and variable swarm topology. Comparative evaluation against MADDPG and MAPPO demonstrates that the proposed approach achieves higher learning stability, greater average cumulative reward, fewer inter-agent conflicts, and a higher mission success rate. The conclusions confirm the effectiveness of the developed adaptive framework for UAV swarm coordination in complex and unstable conditions and outline future research directions related to self-organization, multi-agent transformer models, transfer to real-world platforms, and energy-aware optimization.

Keywords: multi-agent reinforcement learning, UAV swarm, partial observability, communication disruptions, adaptive control, cooperative learning, decentralized systems.