模糊邏輯在決策樹上之應用

Abstract

符號學習的策略通常可以分為兩類，第一類是批次學習策略，其中典型的 代表是ID3和PRISM；第二類是漸進學習策略，其代表是version space 。 本篇論文的焦點在於模糊邏輯和決策樹的結合，其中決策樹是一種批次學 習策略。近來模糊邏輯在人工智慧領域的應用已經愈形重要，因為它可以 處理不精確的資料而且也更接近人類的思考模式。決策樹和其所衍生的一 些演算法非常普遍地運用在從描述特徵的例子中學習推理，它之所以受歡 迎是由於決策樹是一種易於表達，易於產生知識的架構，並且很容易從決 策樹上整理出知識的規則集，並將其成功地運用在專家系統或其他知識庫 系統上；不過這種方法有些缺陷，就是對於處理不精確資料或測量錯誤上 的能力不足。這些問題是來自於資料特徵的連續性，而且對於雜訊的敏感 所致，為了解決這種問題，所以我們提出一種方法將決策樹和模糊邏輯作 結合。結合的第一部份是將輸入資料模糊化，利用模糊化後的資料來做學 習，第二部份是將原有的熵函數加以修改為模糊熵函數，用來作為選擇未 使用過的特徵的參考。另外我們將這種模糊決策樹的方法運用在一個 Iris Plant的資料庫上作一個實驗，實驗的結果顯示經過模糊化後的決策 樹系統比未經模糊前的系統稍有改進；未來的工作首先希望能在另一個適 合的領域作另一個實驗，其次希望能發展一種方法將模糊決策樹整理成一 個模糊規則集合。 Symbolic learning strategies can usually be divided into two classes. The first class is batch learning strategies, such as ID3 and PRISM. The second class is incremental learning strategies, such as version space. In recent years, the applications of fuzzy logic becomes increasing important in artificial intelligence research fields since it can deal with the problem of imprecise data and is closer to the thinking model of human than the traditional approach. The focus of this thesis is on the merger of fuzzy logic and decision trees. Decision trees and their various algorithms are popular choices in applications to learning and reasoning from feature-based examples. This popularity is due to easily understood processing mechanisms, comprehensibility of the generated knowledge structure, which might be converted to a set of rules and subsequently used in a diagnostic expert system, and wide availability of data in a form of feature descriptions. However, there are still some problems related to the traditional decision. They include the inability to cope with missing data, imprecise information, and measurement errors. The most important sources of the above problem come from continuous attributes and the sensitivity of noise (from imprecise information or measurement errors). To solve the above problem, merging traditional decision trees with the fuzzy processing is proposed. The fuzzy approach to decision trees fuzzifies the numeric data at first, and uses the fuzzy entropy function to select the best feature which is not used. Furthermore, an experiment of fuzzy and non-fuzzy decision tree on Iris Plants Database is done. The result shows that the fuzzy approach is a little better than non-fuzzy approach on accuracy and tree size since the Iris data contains only 150 cases. The future work is to do another experiment on a proper domain and develop a methodology to convert the fuzzy decision tree to a fuzzy rule set.