.png){kind=icon}
.png){kind=icon}
Return
Conditional diffusion-enhanced deep reinforcement learning for data-efficient cognitive interference
.png){kind=icon}
PRE
AI
DOI：10.1016/j.phycom.2026.103098.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Deep reinforcement learning (DRL) has shown strong potential in adaptive interference decision-making by enabling agents to dynamically adjust strategies in response to time-varying communication environments. However, DRL-based approaches typically suffer from low sample efficiency due to limited online interaction opportunities and high data acquisition costs, which significantly constrain their practical deployment.To address this challenge, this paper proposes a conditional denoising diffusion probabilistic model (CDDPM)-enhanced DRL framework that improves learning efficiency through experience data augmentation. The proposed method reformulates DRL experience tuples into diffusion-compatible training samples and leverages CDDPM to synthesize state-action-consistent experience data, thereby reducing the required number of online environment interactions. Specifically, a data transformation strategy is designed to construct both the diffusion training set and conditional inputs from existing DRL experience, enabling the generation of unseen but behavior-consistent data. To ensure effective data synthesis at the early stage of online learning, the CDDPM is pretrained using offline DRL experience data. In addition, conditional information incorporating both state and action variables is introduced to guide the diffusion process, achieving a balance between generation diversity and fidelity. Experimental results demonstrate that the proposed framework significantly accelerates DRL convergence and outperforms conventional DRL and existing generative data augmentation methods in terms of learning efficiency and interference performance.
Keywords:
Cognitive interference strategy
Deep reinforcement learning
Experience data synthesis
Conditional denoising diffusion probabilistic model
Journal
IF:
2.2
Papers:
360
Citations:
2.6K