中图分类号：TP391.9文献标识码：ADOI:10.19339/j.issn.1674-2583.2026.04.011
中文引用格式：唐坤，王东升，韩兴豪，等. VG-ADP:基于价值引导与扩散对抗的模仿学习方法［J］.集成电路应用，2026，43（4）：60-64.
英文引用格式：Tang Kun, Wang Dongsheng, Han Xinghao,et al. VG-ADP: Value-guided adversarial diffusion imitation learning［J］.Application of IC，2026，43（4）：60-64.

VG-ADP: Value-guided adversarial diffusion imitation learning

Abstract： To address suboptimal expert data and complex multimodal action distributions in multi-agent imitation learning, this research proposes a value-guided adversarial diffusion multi-agent imitation learning framework (VG-ADP). The framework employs a Diffusion Probabilistic Model as the core generator, integrated with a U-Net structure to capture spatio-temporal environmental features, accurately modeling the multimodal nature of expert strategies while enhancing action sequence smoothness. To mitigate data noise, an adversarial enhancement module is introduced to identify and filter suboptimal trajectories. Furthermore, a value gradient operator is injected into the reverse denoising phase, utilizing collaborative value priors to dynamically guide agents toward high-reward regions for trajectory correction.

Key words : diffusion model；adversarial imitation learning；value guidance