具有處理模糊語言能力之類神經網路系統

Abstract

本篇論文提出一具有處理模糊語言之類神經網路系統。我們使用α- level sets形式的模糊集合(fuzzy number)來表示一個語言性質的資料 (linguistic data)。本論文首先提出一個五層的類神經網路架構，這個 架構能處理數值性(numerical)的輸入以及模糊語言的輸入，而且針對不 同性質的應用，產生所需要的數值性的輸出或模糊的輸出。這個架構並不 需要太多的空間及記算時間，而能夠表示任何形態的模糊集合。至於學習 方式，我們討論了有指導性 (supervised) 的學習方式及評論性質 (reinforcement learning) 的學習方式。對於有指導性的學習方式，一 個重要的應用–知識庫壓縮 (rule base concentration) 被展示。至於 評論性質的學習方式，我們強調評論訊息 (reinforcement signal) 是一 種模糊資料，而不像傳統的方式只使用數字來表示。而且針對兩種不同性 質的問題，single-step及 multi- step 的預測方法分別地被提出。常見 的倒單擺 (cart-pole balancing)問題被實際地模擬來顯示我們所提之系 統的功效。 A neural fuzzy system with linguistic teaching signals is proposed in this thesis. We use fuzzy numbers based on α-level sets to represent linguistic information. At first, we propose a five-layered neural network which can process numerical information as well as linguistic information. Moreover, the inputs and outputs of this five-layered connectionist architecture can be a hybrid of fuzzy numbers an numerical numbers. The important characteristics of the proposed model are that the network weights can be fuzzy numbers of any shaped and the performance of this model is superior to several other methods both in learning speed and memory requirement. Two kinds of learning schemes are discussed: supervised learning and reinforcement learning. With supervised learning, the proposed model can be used for rule base concentration to reduce the number of rules in knowledge base. For reinforcement learning, we consider that the reinforcement signal from environment is linguistic information such as "good" , "very good", or "bad". We discuss two kinds of reinforcement learning learning problems:single-step prediction problems and multi- step prediction problems. Simulation results of the cart- pole balancing problem are presented to illustrate the performance and applicability of the proposed reinforcement system.