DNN-Based Phase Prediction for Distributed Multi-RIS Systems under Co-Channel Interference

Abstract

This paper investigates learning-assisted 2-bit phase control for a distributed multi-reconfigurable-intelligent-surface (multi-RIS) downlink under co-channel interference (CCI). A desired base station serves a single-antenna user while two neighboring base stations reuse the same time–frequency resource. We consider exhaustive RIS activation (ERA), where all surfaces reflect simultaneously, and opportunistic RIS activation (ORA), where the RIS providing the largest instantaneous signal-to- interference-plus-noise ratio (SINR) is selected. The proposed method uses channel-state-information (CSI)-derived desired and interfering cascaded-channel features and an offline discrete coordinate-ascent (DCA) teacher that maximizes the high-SNR signal-to-interference ratio (SIR). For ERA, the DNN learns 2 bit residual corrections relative to nearest-neighbor quantized phase alignment (QPA). Because about 90% of ERA teacher coordinates require no correction, class-balanced cross-entropy is combined with a utility-aware DCA–DNN SIR-gap loss and post-warm-up utility-based checkpoint selection. ORA retains absolute phase-state prediction. Fixed phases, QPA, DCA, and continuous-phase alignment (CPA) are used as common baselines. On an independent test set, the mean DCA–DNN SIR gaps are 0.0435 dB for ERA and 0.0843 dB for ORA. At Ps = 25 dBm, the DNN ergodic-capacity gaps to DCA are only 0.0040 b/s/Hz and 0.0063 b/s/Hz, respectively. High-SNR analysis further establishes nonzero outage floors and finite capacity ceilings when all co-channel transmit powers scale together. The results position the DNN as an iteration-free surrogate for interference-aware discrete phase control while confirming that QPA remains a strong low-complexity baseline.