Quantized IRS Phase Control under Limited Observation with Short-History State Stacking

Finite-resolution intelligent reflecting surfaces (IRSs) require element-wise quantized phase control over a high-dimensional multi-discrete action space. In mobile scenarios, the controller operates with a compact observation comprising current geometry, user activity, the applied IRS configuration, and aggregate link feedback. This paper investigates short-history state stacking for DDQN-family quantized IRS phase control under this limited-observation interface. The stacked-input variants replace a single observation with a four-step history while retaining the corresponding standard or dueling value-network family, quantized action interface, replay mechanism, and training protocol. A six-method comparison includes Fixed Random, a position-aware Geometric heuristic, DDQN, D3QN, SEQ-DDQN, and SEQ-D3QN. In the reference configuration, short-history stacking increases the trailing-window spectral efficiency by approximately 6.9% for the DDQN pair and 1.7% for the D3QN pair. The Geometric heuristic provides a complementary geometry-informed benchmark for evaluating feedback-driven value control. Sequence-length analysis further characterizes the balance among observation span, spectral efficiency, reward behavior, input dimension, and replay-storage cost. Under the fixed-factor Rician extension, both stacked-input variants yield positive spectral-efficiency differences in four of five fading streams. These results demonstrate that short-history state stacking can strengthen temporal input representation for feed-forward DDQN-family controllers and provide a compact reference design for quantized IRS phase control under limited observation.