Active Sensing for Communications by Learning

This paper proposes a deep learning approach to class of active sensing problems in wireless communications which an agent sequentially interacts with environment over predetermined number time frames gather information order perform or actuation task for maximizing some utility function. In such setting, the needs design adaptive strategy based on observations made so far. To tackle challenging problem dimension historical increases time, we propose use long short-term memory (LSTM) network exploit temporal correlations sequence and map each observation fixed-size state vector. We then neural (DNN) LSTM at frame next measurement step. Finally, employ another DNN final desired solution. investigate performance proposed framework channel communications. particular, consider beamforming mmWave beam alignment reconfigurable intelligent surface reflection alignment. Numerical results demonstrate that outperforms existing nonadaptive schemes.