Research on Clutter Suppression Based on Complex-Valued Residual Network and Dynamic Reward Mechanism

Abstract

As deep reinforcement learning becomes increasingly applied to clutter suppression, existing methods have shown a certain level of adaptability. However, their capabilities in feature representation and generalization remain limited. To address the shortcomings associated with the static reward mechanism—namely, its limited adaptability and slow learning speed—a Complex-Valued Residual Deep Q-Network based on a Dynamic Reward Function (CV-ResDQN-DRF) is proposed in this study. In this method, complex-valued residual units are introduced into the complex-valued neural network framework. Through these units, a complex-valued residual network is constructed to enhance the representational capacity of both amplitude and phase features of signals. Simultaneously, a dynamic reward mechanism is designed, wherein the feedback is adaptively adjusted in real time according to the environmental states and the agent’s behavior, thereby accelerating the learning process. Experimental results show that the proposed CV-ResDQN-DRF model achieves an average signal-to-clutter-plus-noise ratio (SCNR) improvement of approximately 2.3 dB on simulated data and 1.8 dB on real measured data, and exhibits a significantly faster convergence speed. These results demonstrate a significant enhancement in clutter suppression performance under complex and non-stationary environments.