在現今社會中，AI扮演的角色日漸重要，它在眾多領域的應用也逐漸得到認可，從而吸引了越來越多的學者致力於AI領域的研究工作。 其中，CNN在圖像和影像處理領域上得到了顯著的成果。然而，這些CNN模型的準確性通常取決於其底層架構。為了使模型的效能上升，領域專家經常需要手動設計架構，這意味著將花費大量的時間和精力來微調模型結構。為了解決這個難題，本研究提出了一種RL-based GAN 的神經架構搜索方法，GANAS。 透過現有的 CNN 模型，GANAS 將學習如何生成更好的模型架構。並且與基於其他策略的神經架構搜索方法相比，GANAS 將能以更短的時間完成訓練。同時，GAN的多樣性和集束搜索能使其找到更多的可行解，從而解決NAS單一最優解和尋找效能不穩定的問題。;Presently, AI has witnessed widespread proliferation and prosperity across diverse domains, thereby attracting a burgeoning community of scholars engaged in AI research endeavors. Among these advancements, CNNs have attained awe-inspiring accomplishments in the domains of image and video processing. However, the accuracy of these CNN models is typically contingent upon their underlying architectures. To bolster model performance, domain experts often resort to manual architectural design, entailing significant time and effort spent on fine-tuning the model structures. To tackle this quandary, our study proposes a RL-based GAN neural architecture search methodology, GANAS. Capitalizing on pre-existing CNN models, GANAS acquires the capacity to learn how to generate enhanced model architectures within the GAN framework. In comparison to other strategy-based neural architecture search methods, GANAS demonstrates expedited training durations. Additionally, by harnessing the diversity of GAN and beam search techniques, we remedy the limitations of NAS characterized by the single optimal solution and the instability of search performance.
