有加強式學習能力之適應性模糊語句解析

Abstract

本篇論文提出一個具有四層的「適應性模糊語句解析神經網路(Adaptive Fuzzy Language Acquisition Network)」系統，而在此系統中，「模糊神經網路(Fuzzy Neural Network)」將與「適應性語句解析網路(Adaptive Language Acquisition Network)」相互結合。我們簡稱此網路為”AFLAN”。這個新網路可以藉由與複雜環境的互動(Iteraction)來達到適應性的模糊語句解析，也就是說，它可以將一些帶的模糊字眼的自然語句訊息轉譯成回使用者預期想要的資料，而這個資料含括有「語意上的動作(semantic action)」及針對該動作所做的「言辭上的模糊修飾(Linguistic information)」（譬如說：在「以非常快的速度向前移動」的模糊語句中，『向前移動』代表著語意上的動作，而『非常快』則表示言辭上的模糊修飾）。此外，AFLAN有三個的重要的特質：第一，使用者可以任意的輸入他想表達的意思，而不受任何語句的限制；而且，這個網路並不要一些像聽覺的、韻律學的、文法上的或者是句子組織學上的結構。第二，代表言辭上的模糊修飾程度是經由適應性的學習而得來的，而且我們是以α-level sets形式的模糊集合(fuzzy number)來表示它。第三，此網路具有即時操作時線上學習的能力。 在此論文中，我們也將探討兩種學習形式，一為非線上學習(off-line learning)，另一為線上學習(on-line learning)。非線上學習主要是使用在學習用的資料是已知的環境下，而為了完成網路的非線上學習，我們利用「相互資料指導性學習(mutual information supervised learning)」法則來學習語意上的動作，此外還利用「模糊倒傳遞學習(fuzzy back-propagation learning)」法則來學習言辭上模糊修飾的程度。相對的，線學習主要是使用於在做即時操作時而所使用的資料是在不確知的環境下。在線上學習的學習法則方面我們將之歸類為兩組，除了上述在非線上學習所提到的兩種學習法則外，我們更利用「相互資料加強式學習(mutual information reinforcement learning)」法則來學習語意上的動作及「模糊加強式學習(fuzzy reinforcement learning)」法則來學習語意上的動作及「模糊加強式學習(fuzzy reinforcement learning)」法則來學習言辭上模糊修飾的程度；而這兩組學習法則的不同點主要是在評判式訊號(critical signal)的來源不同，前組的訊號是由指導者(supervisor)所提供的，所以這種學習是監督式學習(superuised learning)；而後組的加強式訊號是由使用者直接提供，因此我們稱呼這種學習方式為加強式學習，而這組學習方使得線上學習更為即時。此外，我們必須在這兩組學習法中各選出一適當的法則來做網路的線上學習。 在本篇論文的最後，我們將以所發展的「具模糊命令解析的控制系統」來實際地展現所提的系統之功效。

This thesis proposes a four-layered Adaptive Fuzzy Language Acquisition Network (AFLAN), in which a fuzzy neural network is embedded in an adaptive language acquisition network. The proposed network can adaptively acquire fuzzy language via interactions with a complex environment. More clearly, the network can translate a natural fuzzy language message into the intended information that includes meaningful semantic action and the linguistic information of that action (for example, the phrase" move forward" represents the meaningful semantic action and the phrase" very high" represents the linguistic information in the fuzzy language " Move forward in a very high speed.").The proposed AFLAN has three important characteristics. First, we can make no restrictions whatever on the fuzzy language input which is used to specify the desired information, and the network requires no acoustic, prosodic, grammar and syntactic structure. Second, the linguistic information of an action is learned adaptively and it is represented by fuzzy numbers based on α-level sets. Third, the network can learn during the course of performing the task. The AFLAN can perform off-line as well as on-line learning. For the off-line learning, mutual-information (MI) supervised learning scheme and fuzzy back-propagation (FBP) learning scheme are employed when the training data are available in advance. the former learning scheme is used to learn the meaningful semantic action, and the latter learning scheme to learn the linguistic information. The AFLAN can also perform on-line learning interactively when it is in use for fuzzy language acquisition. For the on-line learning, the MI reinforcement learning scheme and fuzzy reinforcement learning scheme are developed for on-line tuning. These two learning schemes are for the on-line learning of meaningful action and linguistic information, respectively. Finally, the application example, "fuzzy commands acquisition of a voice control system", will be conducted to illustrate the performance and applicability of the proposed AFLAN.