CADer: A Deep Reinforcement Learning Approach for Designing the Communication Architecture of System of Systems

Abstract

Modern war relies heavily on various combat units, such as command and control units, surveillance units, and reconnaissance units. These units are usually combined organically as a combat SoS (System of Systems) to conduct given missions. To ensure that the combat SoS executes correctly, the communication systems (e.g., communication vehicles, communication UAVs) must be planned carefully to connect all constituent systems or units. However, the <bold xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">c</b> ommunication <bold xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">a</b> rchitecture <bold xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">d</b> esign problem for combat SoS (SoS-CAD) is very challenging as the combat SoS usually resides in a dynamic and confrontational environment. In this paper, we formally formulate the SoS-CAD problem with integer programming and prove its NP-hardness. The SoS-CAD problem has high requirements for the solving speed as well as the solution quality, but traditional meta-heuristic algorithms cannot satisfy them. Thus, we propose a deep reinforcement learning-based method called CADer to solve it. Specifically, we introduce the attention mechanism and dynamic embedding mechanism into CADer with considering the characteristics of the SoS-CAD problem itself. The massive experiment results show that CADer performs well in the generalization ability and can achieve the best tradeoff between the solving speed and the solution quality against the meta-heuristic algorithm.