神經網路建構之專家建議系統–以電腦輔助鞋樣設計為例

Abstract

中文摘要在人所主導的設計過程中，許多繪畫的圖形概念潛在著多個隱含 之可能性，或是發展性。在認知上，設計師有能力識別其中的已知片斷( knowledge pieces)，並組合出浮現特性(emergent properties)。例如： 視覺上的形、需求、屬性等因素，被綜合考慮之後，便可進一步作為處理 設計過程中所面臨到形之認知與轉換和當中所蘊含的知識表達。然而，就 現行電腦輔助設計系統 (Computer-aided Design (CAD) Systems)而言， 基於(rule -based)文法規則基礎之循序式處理運算，卻全然地反映出毫 無能力處理上述所考慮之設計因素的情況。為了解決當前電腦輔助設計系 統(Computer-aided Design (CAD) Systems)無法於真實設計環境中，克 服所遭遇之設計問題和right-hand-side設計狀態的相互配合，本研究嚐 試從認知心理學與人工智慧連結作用所發展的觀點上，建立多層級連結網 路或稱為平行分散式處理模式(Parallel Distributed Processing (PDP) Models)，以解決其間設計的問題。關鍵字：知識表達、連結網路、平行 分散式處理模式、神經網路、倒傳遞迴網路、內容尋址記憶 AbstractIn human design processes, a designer is able to control different problem-solving simultaneously such as visual shapes, demands, and features, to solve the recognition and transformation of processed shapes and contented attributes. Unfortunately, current CAD (computer-aided design) systems entirely fails to deal with them. To solve the above problems, this study uses multilayer connectionist networks also called PDP (parallel distributed processing) models received outstanding attention in the field of AI, cognitive science and cognitive psychology. PDP models learn from examples: two connectionist knowledge representations such as from/demands, demands/features, will be presented as input/output pairs of neural networks. After the networks are well-trained, the back- propagation neural networks can be applied to solve the design mapping problems such as the recognition and transformation of processed shapes and contented attributes. In addition, this study is intended to explore the behaviors underlying neural networks.Keyword: knowledge representations, connectionist networks, PDP models, neural networks, back propagation, content-addressable memory.