Doppler-Aware Meta-Learning for Evolutive STAR-RIS in Cognitive Autonomous Networks
In real-time optimization of simultaneous transmission and reflection reconfigurable intelligent surfaces (STAR-RIS), limitations arise due to Doppler effects and reduced adaptation speed. Existing RIS memoryless surface models use quasi-static optimization and incur high latency, which limits their application in cognitive autonomous networks. To overcome this limitation, we propose a stateful RIS model that includes an element-wise memory, which locally stores and updates its phase history across coherence intervals. Hence, the metasurface becomes a stateful surface with element-level memory that enables the exploitation of temporal channel correlations. This element-wise memory is complemented with a meta-parameter, yielding faster convergence and giving rise to hierarchical adaptation in gradient and episodic timescales. The proposed framework supports autonomous and experience-driven learning, which enables the STAR-RIS architecture and memory to adapt according to the Doppler effect, as seen in cognitive communication systems. Results demonstrate an improvement in spectral efficiency and adaptation speed across various mobility patterns, fading models, interference levels, and array sizes. An analytical field- programmable gate array (FPGA) latency estimate indicates that the critical path fits within the coherence window.