在本研究中，我們考慮一個由流體天線系統（FAS）與雙面主動式同時透 射與反射重構智慧表面（DS-STAR-RIS）輔助的語意整合感知與通訊（semantic ISAC）系統，其中FAS 天線與DS-STAR-RIS 元件皆可於指定區域內調整位 置。系統同時考慮通訊、感知與語意傳輸需求，並以語意能源效率（SEE）最 大化為主要目標。為此，我們聯合最佳化通訊與感知波束成形、DS-STAR-RIS 相位位移與放大係數、語意壓縮比、FAS 與使用者關聯、FAS 天線位置，以 及DS-STAR-RIS 元件位置，同時滿足傳輸功率、DS-STAR-RIS 功率、最低資料 率、感知波束圖增益、相位耦合與最小元件間距等限制。為了解決上述高維度 且同時包含連續與離散變數的最佳化問題，本研究提出雙模型多智能體混合近 端策略最佳化（TM-MAHPPO）架構。RIS 智能體與FAS 智能體分別配置連續 型與離散型近端策略最佳化，以處理連續與離散動作，並透過雙模型結構進一 步產生與修正決策。此外，本研究在各智能體內部以及智能體之間導入資訊共 享，使不同策略能利用其他決策分支所產生的動作資訊，以提升不同決策之間 的協調能力。此外，本研究採用變分自編碼器（VAE）壓縮高維度的狀態與動 作資訊，以降低學習與資訊交換的負擔，並利用元近端策略最佳化執行超參數 最佳化（HO），使PPO 超參數能隨訓練過程動態調整。模擬結果顯示，所提 出的方法在SEE 與收斂表現上優於所考慮的基準深度強化學習方法。結果亦 顯示，智能體內資訊共享與智能體間資訊共享均能改善不同決策模組之間的協 調，而VAE 與HO 則分別在資訊維度與訓練適應性方面提供效益。此外，提高 DS-STAR-RIS 元件的移動能力能增加空間自由度並提升SEE，但隨著元件數量 持續增加，其效能增益會受到額外功率消耗與最佳化複雜度影響而逐漸趨緩。;In this thesis， we investigate a semantic ISAC system assisted by FASs and a dual-sided active STAR-RIS， where both the FAS antennas and DS-STAR elements can adjust their positions within predefined regions. The system jointly considers communication， sensing， and semantic transmission requirements， with SEE maximization as the main objective. Accordingly， communication and sensing beamforming， DS-STAR phase shifts and amplification coefficients， semantic compression ratios， FAS-user association， FAS antenna positions， and DS-STAR element positions are jointly optimized under transmit-power， active DS-STAR power， minimum-rate， sensing beampatterngain， phase-coupling， and minimum-spacing constraints. To handle the resulting highdimensional mixed continuous-discrete optimization problem， we propose a TM-MAHPPO framework. The RIS and FAS agents employ ConPPO and DisPPO to optimize continuous and discrete actions， respectively， while the twin-model structure is used to generate and refine the decisions. Intra-agent and inter-agent information sharing are further introduced so that different policies can exploit action information generated by other decision branches. In addition， VAE-based compression is employed to reduce the dimensions of the state and exchanged action information， while Meta-PPO-based HO adaptively adjusts the PPO hyperparameters during training. Simulation results show that the proposed method achieves higher SEE and favorable convergence performance compared with the considered benchmark DRL methods. The results also demonstrate that both intra-agent and inter-agent information sharing improve coordination among different decision modules. VAE compression reduces the information dimension while retaining most of the useful features， and HO improves the adaptability and stability of the learning process. Moreover， increasing the mobility of DS-STAR elements provides additional spatial DoF and improves SEE， although the performance gain gradually saturates as the number of elements increases because of additional power consumption and optimization complexity.
