Deep Reinforcement Learning-Based Adaptive Modulation for Underwater Acoustic Communication with Outdated Channel State Information

Underwater acoustic (UWA) adaptive modulation (AM) requires feedback about channel state information (CSI) but the long propagation delays and time-varying features of UWA channels can cause the CSI feedback to be outdated. When the AM mode is selected by outdated CSI, the mismatch between the outdated CSI and the actual CSI during transmission degrades the performance and can even lead to communication failure. Reinforcement learning has the ability to learn the relationships between adaptive systems and the environment. This paper proposes a deep Q-network (DQN)-based AM method for UWA communication that uses a series of outdated CSI as the system input. Our study showed that it could extract channel information and select appropriate modulation modes in the expected channels more effectively than single Q-learning (QL) without needing a deep neural network structure. Furthermore, to mitigate any decision bias that was caused by partial observations of UWA channels, we improved the DQN-based AM by integrating a long short-term memory (LSTM) neural network, named LSTM-DQN-AM. The proposed scheme could enhance the DQN's ability to remember and process historical input channel information, thus strengthening its relationship mapping ability for state-action pairs and rewards. The pool and sea experimental results demonstrated that the proposed LSTM-DQN-AM outperformed DQN-, QL- and threshold-based AM methods.