DOL-net: a decoupled online learning network method for RIS-assisted ISAC waveform design

Abstract

The constant modulus (CM) waveform design, for Multiple-Input-Multiple-Output (MIMO) Integrated Sensing and Communication (ISAC) systems, is a key technology. The existing methods mainly include waveform design without Reconfigurable Intelligent Surface (RIS) or the sensing-based waveform design with RIS, which usually degrade the comprehensive performance. To address this issue, the comprehensive waveform design with RIS is proposed, where we jointly maximizing the Signal-to-Interference-and-Noise-Ratio (SINR) for radar and minimizing the RIS-aided Multiple User Interference (MUI). Due to the coupling effect between the phase shifts and waveform, the problem is difficult to solve. Additionally, the CM constraint on both the phase shifts and waveform along with the fractional SINR expression further aggravate the difficulty. To address these issues, a Decoupled Online Learning Network (DOL-Net) method is proposed using the characteristic of multiple input to multiple output mapping, which includes trainable network parameters module, loss calculation module and back-propagation module. Given fixed network parameters, the cost function can be calculated in the loss calculation module. Then, the network parameters are trained by the deep learning optimizer to minimize the loss function. Compared with the existing methods, the proposed method obtains 0.69 dB radar SINR enhancement and 138.8 % communication sum rate improvement.