A Dec-POMDP model for decentralized control of multi-agent groups with dynamic interaction topology and stochastic communication disturbances

DOI 10.31673/2412 9070.2026.318107

Abstract

The article is devoted to the theoretical justification and development of a mathematical model for decentralized control of dynamic groups of autonomous agents, in particular swarms of unmanned aerial vehicles, in conditions of fundamental uncertainty and a hostile operating environment. The authors propose an extension of the classical model of decentralized partially observable Markov decision process into a model that considers the dynamic topology of interaction between agents (Dynamic Topology Decentralized Partially Observable Markov Decision Process – DT-Dec-POMDP). The key difference lies in the integration of dynamic interaction topology and explicit modeling of stochastic communication channel disturbances as internal factors that determine the information environment of each agent.
It has been proven that ignoring these aspects leads to the loss of stability of theoretically optimal strategies in real-world scenarios, particularly during reconnaissance, electronic suppression, and search and rescue operations. To address information gaps caused by packet loss, we propose the Comm-Disturbance Filter architectural block, which implements a predictive compensation mechanism based on recursive extrapolation of neighboring agents' states. To process a stabilized stream of observations and context-dependent information aggregation, a Multi-agent Transformer has been introduced, capable of adaptively weighing input data in conditions of a changing network structure. The research results reveal critical thresholds of system stability depending on the intensity of communication failures and demonstrate the ability of the proposed architecture to maintain functionality in the compensation zone. The work provides a theoretical basis for the creation of a new class of stable, scalable, and adaptive autonomous systems capable of effective coordination in conditions of extreme uncertainty and counteraction.

Keywords: UAV swarm, dynamic topology, stochastic communication disturbances, transformer, multi-agent systems, cooperative autonomy.