Agentic AI enabled pinching-antenna systems

A pinching-antenna system (PASS)-enabled downlink communication framework is investigated, where pinching activation (PA) and pinching sliding (PS) protocols are employed to dynamically adjust the positions of pinching-antennas against dynamic wireless channels. The optimization goal is to maximize the achievable sum data rate by jointly optimizing the pinching configuration and power allocation over time-varying fading blocks. Although there are precedents for developing AI algorithms to optimize PASS, the inference latency caused by the AI algorithm is explicitly taken into account. To reduce the inference cost of employing an AI agent for PASS optimization, an agentic deep learning (DL) approach and an agentic deep reinforcement learning (DRL) approach are proposed for the PASS employing the PA protocol and the PS protocol, respectively. For the PA protocol, the agentic DL framework employs a predictor model to select pinching positions and power decisions, together with a light-weight self-controller regulating the inference frequency to reduce inference overhead. For the PS protocol, the positions of moving pinching-antennas form a Markov Decision Process (MDP), where the agentic DRL framework leverages a dualpolicy actor–critic architecture to jointly learn long-term beam-forming control and self-control strategies. Numerical results indicate that both agentic DL and agentic DRL are able to outperform the performance of inference-per-step baselines by significantly reducing the average inference time. Additionally, in scenarios with low user mobility, the DRL-driven PS protocol is capable of achieving a sum rate advantage via more precise control. In the opposite case, the DL-driven PA protocol is superior due to high model generalization performance and prompt pinching-antenna position switching.