Reinforcement learning for adaptive interference suppression in a phased array antenna with discrete phase control

Problem statement. Phased array antennas (PAA) with low-cost multi-bit phase shifters, providing only a few (e.g., eight) discrete phase states per element, are widely used in communication and radio surveillance systems. When a source of intentional or unintentional interference appears, a null must quickly be formed in the radiation pattern (RP) in the direction of the jammer while maintaining gain toward the desired signal. Classical analytical adaptive-nulling methods are designed for continuous phase control; after rounding the resulting solution to the nearest discrete state (quantization), the null depth often degrades sharply, while an exhaustive search over all phase combinations is computationally infeasible even for a moderate number of array elements. Goal. To develop and experimentally investigate a method for adaptive nulling of the RP of a linear PAA with discrete (8-position) phase control based on reinforcement learning, operating directly in the space of discrete phase-shifter states, and to quantitatively compare it with a classical method (analytical nulling followed by phase quantization) and with exhaustive search as a global-optimum benchmark. Results. A linear PAA of eight radiating elements with a spacing of 0.19 m at a carrier frequency of 800 MHz and eight discrete phase states per element was considered. The nulling problem was formulated as a combinatorial optimization over 88 = 16,777,216 phase-state variants and solved using the policy gradient method (REINFORCE) with an independent categorical policy per array element. For a jammer in the angular range 15°–65°, the classical method after phase quantization provides only 0.6–8.3 dB of jammer suppression relative to the main-lobe level, whereas the proposed method, using only 3000 model evaluations (0.018% of exhaustive search), provides 15.2–32.2 dB of suppression with a main-lobe gain loss of no more than 5.2 dB. Practical relevance. The proposed approach makes it possible to find practically applicable solutions to the adaptive nulling problem directly in the space of discrete states of real phase shifters 3-4 orders of magnitude faster than exhaustive search, making it suitable for embedding into adaptive PAA beam-control systems with limited computational resources.