Telescopic STAR-RIS for ISAC Networks: An LSTM-Based Multi-Agent DRL Framework

Conventional simultaneously transmitting and reflecting reconfigurable intelligent surfaces (STAR-RISs) employ fixed-height elements and therefore adapt poorly to fast block ages, Doppler shifts, and mutual coupling in dense integrated sensing-and-communication (ISAC) scenarios. We propose a telescopic STAR-RIS (T-STAR-RIS) mounted on a high-rise facade whose passive elements can extend or retract vertically while adjusting phase and transmit–reflect power split. This extra degree of freedom enables fine beam shaping for full-duplex non orthogonal multiple access (NOMA) ISAC. An integrated ISAC based model that incorporates Doppler, RIS feature adjustments, path loss, and imperfect channel state information is formulated as a non-convex optimization. To solve it in a stochastic setting, we design a multi-agent deep reinforcement-learning (DRL) framework with a single global critic, actor regularization for stability, and a long short-term memory (LSTM) network that forecasts channel evolution. With 128 T-STAR-RIS elements and a 4-antenna base station, simulations show 6.4% higher downlink throughput (bps/Hz) than an equal-size fixed-height STAR-RIS. They also indicate an ≈ 9 dB sensing-SNR improvement under AR(1) fading with imperfect CSI. The LSTM keeps channel estimation error below 0.055 in more than 80% of coherence intervals, and the DRL scheme converges in under 200 episodes.