Verified-reward training for syndrome-guided detector selection in one-bit QPSK MIMO under imperfect CSI

Abstract We study detector selection for a heavily loaded $$N=16$$ N = 16 , $$K=8$$ K = 8 one-bit QPSK MIMO uplink with imperfect channel state information. The physical samples are generated by the exact sign quantizer; Bussgang models are used only to design the soft-output receivers. A contextual deep Q-network (DQN) selects among linear and MMSE-PIC actions from a receiver-computable observation and is trained with a syndrome–CRC-verified event. The evaluation comprises 1392 completed runs, 6.96 million packets, three independent training seeds, and matched episode streams across Rayleigh, correlated UMi NLOS, and a common-steering UMi LOS stress family. Across the six primary channel/CSI conditions, Holm-adjusted tests detect no significant difference between DQN-V8 and Fixed MMSE in five cases. In the UMi LOS stress case with $$\sigma _e^2=0.1$$ σ e 2 = 0.1 , DQN-V8 has lower true packet success (TPS) by 0.111 percentage points (Holm-adjusted $$p=0.0096$$ p = 0.0096 ). Fixed Turbo-3 has the lowest coded BER in the Rayleigh SNR sweep; in the primary cells it also executes more detector calls and belief-propagation work. Matched covariance controls show that the diagonal residual is close to the full arcsine-colored receiver in the evaluated conditions (maximum absolute TPS contrast about 0.33 percentage points), but differs more from a scalar-white model (up to 1.41 points). A policy trained on $$K\in \{4,8,12\}$$ K ∈ { 4 , 8 , 12 } changes its held-out $$K=6$$ K = 6 TPS by only $$-0.12$$ - 0.12 to $$+0.33$$ + 0.33 points relative to the fixed-load policy over the tested grid. CRC-6, CRC-8, and CRC-16 controls quantify the trade-off among data payload, reward sparsity, and false acceptance. The selector learns from receiver-observable verification and achieves TPS estimates close to Fixed MMSE in most tested conditions, while using fewer detector calls and BP iterations than Fixed Turbo-3.