Researchers have developed a novel verified-reward training framework designed to optimize detector selection in heavily loaded one-bit quadrature phase-shift keying (QPSK) multiple-input multiple-output (MIMO) uplink systems. The study explores how a contextual deep Q-network (DQN) can be trained to dynamically select between linear and minimum mean square error parallel interference cancellation (MMSE-PIC) actions, utilizing observations that are readily computable by the receiver. The DQN-based selector uses fewer detector calls and belief-propagation iterations than Fixed Turbo-3.
The evaluation of this system was extensive, comprising 1,392 completed runs and the processing of 6.96 million data packets. To ensure the robustness of the findings, the researchers employed three independent training seeds and utilized matched episode streams across various channel conditions, including Rayleigh fading, correlated urban microcell (UMi) non-line-of-sight (NLOS) environments, and a common-steering UMi line-of-sight (LOS) stress family. This diverse set of conditions allowed for a thorough examination of how the DQN-based selector performs compared to established, fixed-detector methodologies in the presence of imperfect channel state information.
Analysis of the primary channel and channel state information conditions revealed that, in five out of six cases, Holm-adjusted statistical tests showed no significant performance difference between the DQN-V8 model and a fixed MMSE detector. In the UMi LOS stress scenario with channel estimation error variance of 0.1, the DQN-V8 model exhibited lower true packet success by 0.111 percentage points (Holm-adjusted p=0.0096).
When comparing the proposed system to fixed Turbo-3 detectors, the findings indicated that the fixed Turbo-3 approach often achieved the lowest coded bit error rate during Rayleigh signal-to-noise ratio sweeps. Furthermore, in primary cell evaluations, the fixed Turbo-3 detector was observed to execute a higher number of detector calls and require more belief-propagation work compared to the adaptive DQN selector. This suggests that while fixed high-complexity detectors may offer performance advantages in specific metrics, the verified-reward training approach achieves competitive TPS estimates while successfully reducing the total computational burden.
The study also examined the impact of varying the load factor on the performance of the trained policy. A policy trained on a load factor of K in the set {4, 8, 12} showed only minor variations in held-out TPS for K=6, ranging from -0.12 to +0.33 percentage points relative to a policy trained on a fixed-load setup over the tested grid.
Additionally, the researchers utilized cyclic redundancy check (CRC) controls, specifically CRC-6, CRC-8, and CRC-16, to further quantify the intricate trade-offs between data payload, reward sparsity, and the rate of false acceptance within the system. The evidence demonstrates that the selector is capable of learning effectively from receiver-observable verification. The framework achieves true packet success estimates close to Fixed MMSE while using fewer detector calls and belief-propagation iterations than Fixed Turbo-3.
Subject of Research: Wireless communications and MIMO detector optimization
Article Title: Verified-reward training for syndrome-guided detector selection in one-bit QPSK MIMO under imperfect CSI
Article References: Dang, H. N., Do, Q. D., & Phuong, T. D. (2026). Verified-reward training for syndrome-guided detector selection in one-bit QPSK MIMO under imperfect CSI. Journal on Wireless Communications and Networking. https://doi.org/10.1186/s13638-026-02705-7
Image Credits: AI Generated
DOI: 10.1186/s13638-026-02705-7
Keywords: MIMO, One-bit QPSK, Deep Q-network, Detector selection, Verified-reward training, Imperfect CSI, Massive MIMO, Uplink systems, MMSE-PIC, Channel state information, Wireless networking, Signal processing, Adaptive algorithms, Packet success rate, Computational efficiency, Rayleigh fading, Urban microcell, Belief propagation
Cite Scienmag News
Scienmag. (November 17, 2026). Verified-Reward Training for Detector Selection in One-Bit MIMO Systems. https://scienmag.com/verified-reward-training-for-detector-selection-in-one-bit-mimo-systems/
Scienmag. "Verified-Reward Training for Detector Selection in One-Bit MIMO Systems." Scienmag, 17 November 2026, https://scienmag.com/verified-reward-training-for-detector-selection-in-one-bit-mimo-systems/. Accessed 17 November 2026.
Scienmag. "Verified-Reward Training for Detector Selection in One-Bit MIMO Systems." Scienmag. November 17, 2026. https://scienmag.com/verified-reward-training-for-detector-selection-in-one-bit-mimo-systems/








