Integrating Mixture of Experts into Transformers Architecture to Control UAV Swarms

This article explores the application of Mixture of Experts technology to improve the efficiency of controlling autonomous swarms of unmanned aerial vehicles (UAVs). The authors analyze and develop methods for integrating Mixture of Experts into neural transformers used for data processing in U unmanned aerial vehicle swarms. The authors propose the use of specialized submodels ("experts") to solve specific tasks, such as visual data processing, navigation, meteorological data analysis, and communication between UAVs. The study demonstrates that integrating Mixture of Experts with transformers significantly improves the performance of unmanned aerial vehicle swarms through more efficient task allocation and adaptation to changing environmental conditions. This is achieved by ensuring that each "expert" processes data relevant to its specialization, and the system dynamically selects "experts" depending on the situation. In addition, the authors consider the problem of limited communication channel bandwidth and propose a solution based on quadrature amplitude modulation to ensure reliable data transmission. The scientific novelty of the study is the development of an optimized Mixture of Experts architectures that consider the computational and resource limitations of unmanned aerial vehicles. The practical significance is to create more efficient autonomous systems for reconnaissance and rescue missions. The research opens up new avenues for developing algorithms to determine the relevance of "experts" to specific tasks and study the impact of different Mixture of Experts architectures on the overall system performance.