Decision Transformers for RIS-Assisted Systems With Diffusion Model-Based Channel Acquisition

Jie Zhang, Yiyang Ni, Jun Li, Guangji Chen, Zhe Wang, Long Shi, Shi Jin, Wen Chen, H. Vincent Poor

Abstract

Reconfigurable intelligent surfaces (RISs) have been recognized as a revolutionary technology for future wireless networks. However, RIS-assisted communications must continuously tune phase-shifts relying on accurate channel state information (CSI), which is generally difficult to obtain due to the large number of RIS channels. The joint design of CSI acquisition and RIS phase-shift optimization remains a significant challenge in dynamic environments. In this paper, we propose a diffusion-enhanced decision Transformer (DEDT) framework consisting of a diffusion model (DM) designed for efficient CSI acquisition and a decision Transformer (DT) utilized for phase-shift optimizations. Specifically, we first propose a novel DM mechanism—conditional imputation based on a denoising diffusion probabilistic model—for rapidly acquiring real-time full CSI by exploiting the spatial correlations inherent in wireless channels. Then, we optimize beamforming schemes based on the DT architecture, which pre-trains on historical environments to establish a robust policy model. Next, we incorporate a fine-tuning mechanism to ensure rapid beamforming adaptation to new environments, eliminating the retraining process imperative in conventional reinforcement learning (RL) methods. Simulation results demonstrate that DEDT can enhance the efficiency and adaptability of RIS-aided communications with fluctuating channel conditions compared to state-of-the-art RL methods.

Introduction

Reconfigurable Intelligent Surfaces (RISs) are a cornerstone of 6G technology, acting as "smart mirrors" that steer wireless signals to bypass obstacles and enhance coverage. Despite their potential, two primary bottlenecks hinder their practical deployment: the high overhead of Channel State Information (CSI) acquisition and the poor adaptability of traditional Reinforcement Learning (RL) for beamforming optimization.

RIS elements are passive, meaning they cannot actively process signals. Estimating the channel for hundreds of elements requires massive pilot overhead, which often negates the bandwidth gains provided by the RIS. Furthermore, traditional RL methods used for phase-shift optimization are brittle; they often require complete retraining when environmental conditions change, such as when a user moves or a new obstacle appears.

The DEDT Framework

To address these challenges, we introduce the Diffusion-Enhanced Decision Transformer (DEDT). This framework decouples the problem into two stages: generative channel imputation and sequence-based policy optimization.

Conditional Diffusion for CSI Imputation

One of the core innovations of DEDT is treating wireless channel estimation as a generative image-imputation task. Instead of measuring every RIS element, the system estimates only a small subset, referred to as Partial CSI (PCSI). The Diffusion Model (DM) then treats the missing data as "noise" and utilizes a denoising process to recover the Full CSI (FCSI). This approach exploits the spatial correlations inherent in wireless channels—the principle that adjacent RIS elements likely experience similar channel conditions.

Architecture of the Diffusion Model (DM) designed for channel state information (CSI) imputation from masked partial data.

Sequence Modeling with Decision Transformers

Rather than relying on the Markovian trial-and-error approach of standard RL, DEDT frames beamforming as a sequence modeling problem. The Decision Transformer (DT) is pre-trained on an offline dataset of historical trajectories (states, actions, and rewards). By looking at the entire history of a connection, the DT predicts the optimal next phase-shift, making the system significantly more robust to temporal fluctuations.

System Workflow and Adaptation

The DEDT workflow spans from initial data collection to online inference and policy generalization. A key feature is the few-shot fine-tuning mechanism. When the system encounters a new environment, it freezes the feature-extracting layers of the DT and fine-tunes only the final output layer. This allows for near-instant adaptation without the need for exhaustive retraining.

Overview of the Diffusion-Enhanced Decision Transformer (DEDT) framework, illustrating the offline training and online inference phases.

Experimental Results

The DEDT framework was evaluated using a 3D geometry-based stochastic channel model across multiple simulated environments. Key findings include:

  • Convergence Speed: DEDT reached convergence in approximately 300 epochs, representing a 3x speedup over traditional RL methods like Proximal Policy Optimization (PPO), which required over 900 epochs.
  • Throughput Gains: DEDT outperformed state-of-the-art benchmarks (DM-PPO and DM-DDPG) by approximately 7.5% in effective data rate.
  • Estimation Accuracy: The Diffusion Model significantly reduced the Normalized Mean-Square Error (NMSE) compared to baseline random channel generation, particularly in high Signal-to-Noise Ratio (SNR) regimes.
  • Generalization: In unseen environments, the few-shot fine-tuning allowed the model to achieve near-optimal performance almost immediately, whereas standard RL agents struggled to adapt.

Conclusion

DEDT effectively bridges the gap between theoretical RIS potential and practical deployment. By combining the generative power of Diffusion Models for CSI acquisition with the robust sequence modeling of Decision Transformers, the framework achieves superior efficiency and adaptability. Future work will focus on validating these results on physical RIS hardware and addressing the computational intensity of training these large-scale models for edge-side network controllers.

Summary updated

Abstract

Reconfigurable intelligent surfaces (RISs) have been recognized as a revolutionary technology for future wireless networks. However, RIS-assisted communications have to continuously tune phase-shifts relying on accurate channel state information (CSI) that is generally difficult to obtain due to the large number of RIS channels. The joint design of CSI acquisition and subsection RIS phase-shifts remains a significant challenge in dynamic environments. In this paper, we propose a diffusion-enhanced decision Transformer (DEDT) framework consisting of a diffusion model (DM) designed for efficient CSI acquisition and a decision Transformer (DT) utilized for phase-shift optimizations. Specifically, we first propose a novel DM mechanism, i.e., conditional imputation based on denoising diffusion probabilistic model, for rapidly acquiring real-time full CSI by exploiting the spatial correlations inherent in wireless channels. Then, we optimize beamforming schemes based on the DT architecture, which pre-trains on historical environments to establish a robust policy model. Next, we incorporate a fine-tuning mechanism to ensure rapid beamforming adaptation to new environments, eliminating the retraining process that is imperative in conventional reinforcement learning (RL) methods. Simulation results demonstrate that DEDT can enhance efficiency and adaptability of RIS-aided communications with fluctuating channel conditions compared to state-of-the-art RL methods.
BibTeX
@article{zhang2026decision,
  author = {Jie Zhang and Yiyang Ni and Jun Li and Guangji Chen and Zhe Wang and Long Shi and Shi Jin and Wen Chen and H. Vincent Poor},
  title = {{Decision Transformers for RIS-Assisted Systems With Diffusion Model-Based Channel Acquisition}},
  journal = {IEEE Trans. Wireless Communications},
  year = {2026},
  doi = {10.1109/twc.2026.3692586},
}
Figure