一階邏輯知識庫的一致性修改

Abstract

最近使用一階邏輯語言的觀念學習已經受到廣泛的研究，特別是結合了歸 納式學習法和解說式學習法的系統更受到重視。然而，對修改知識而言， 觀念學習僅提供了部分的解。這主要是因為觀念學習，無論其要求知識庫 完整與否，都只做到利用知識庫而已。已經有一些真正的知識庫修改系統 出現，但大部分限制在命題式邏輯語言。Forte 是最早強調使用一階邏輯 語言做知識庫修改的系統，但是，它不能確保修改後的知識庫能完全符合 輸入的例子。本篇論文中我們提出了一種修改知識庫的方法，能夠使用一 階邏輯語言，而且確保所有的例子都和修正後的知識庫一致。我們提出二 次處理的方法。首先，對輸入的知識庫作一般化擴充，使得能涵蓋所有的 正例。理論上現有的觀念學習系統已可以達到這個目的。接下來對知識庫 作特殊化縮減以排除負例。然而，過度特殊化會造成無法涵蓋一些正例的 後遺症，必須排除這種情形。本論文主要即在設計一個特殊化的演算法處 理這個問題。實驗結果顯示此演算法可以有效的減低過度特殊化的問題。 Recently, conept learning from examples in first-order language is widely studied. Especially, systems that integrate inductive learning and explanation-based learning are flourishing. However, concept learning only provides partial solution of the knowledge revision problem. This is mainly because that concept learning methods only utilize background knowledge no matter the knowledge is perfect or imperfect. There are several true knowledge revision systems. Most of them are in the propositional language. {\bf Forte} is the first real knowledge revision system in the first-order language. However, it cannot guarantee that the revised knowledge is correct on all input examples. In this thesis, we present a real knowlege revision method in first-order language so that the revised knowledge is guaranteed to be correct on all given examples. The proposed method is a two-phase approach. In the first phase, the input theor` is generalized to cover all positive examples. Theoretically, this can be done by existing concept learning systems. The second phase then specializes the theory to exclude negative examples. The specialization must not cause excessive exclusion of covered positive examples. As a matter of fact, the thesis focuses on designing such a specialization algorithm. Experimental results showed that the proposed algorithm can reduce the over-specialization problem.