以貝氏網路為基礎之線上適性測驗系統演算法

A Bayesian Networks Based Computerized Adaptive Testing Algorithm

張瑞欣(Jui-Hsing Chang)

指導教授 ： 劉湘川

亞洲大學/資訊學院/資訊工程學系碩士班/碩士(2006年)

摘要

心理方面的特性均非具體實物，變動而抽象，目前的技術是不能夠直接觀察和測量，但是這些特性可從實際的行為中表現出來，於是我們根據受試者在某一特定情境中的反應情形，以間接測量的方法，來估計或推斷各種潛在的心理特質之差異程度，如何精確與迅速的推斷出受試者的心理特質，正是我們迫切需要的。
現在的教學評量趨向於電腦化測驗，應用貝氏網路因果的理論，針對學生個人、班級、班群或學校等不同的群體，建置一個可以分析診斷問題，預測學生學習能力的系統，來幫助學生找出自己的那些基本概念不清楚，方便教師進行課後補救教學的參考。適性測驗用來診斷個人的技能，適性測驗使用貝氏網路對一個受測的學生和給予的問題模式化。
傳統用在教育心理測驗的方法是試題反應理論。貝氏網路目前的應用被認為是多維IRT的一般化。它有兩個基本優點：
1. 推論的過程更能反映出學生並且提供更佳的透視已模型化的問題。
2. 學生模式是能將技能之間的相依性模型化。因此，在保有原來精確度之下，適性測驗本質上能被縮短。
本研究以國小四年級數學科「面積」單元為例，利用試題證據訓練貝氏網路，選用ＡＯ*演算法來作為選題策略，建構試題結構，與歐氏距離法做比較，來分析了解 一.實體的線上學習診斷系統是否可行？二.適性選題的診斷在實體線上學習診斷系統上，是否具有功效？所以我們收集學生實際在線上作答的情況，以歐氏距離法與貝式網路作為分析的工具，討論其正確率，作為實體線上診斷是否可行的參考。

關鍵字

貝氏網路 ； 歐氏距離 ； 電腦化適性測驗 ； ＡＯ*演算法 ； 熵

並列摘要

Non- concrete material object of characteristic of the psychology, but the change is abstract, can not be observed and measured directly in present technology, but these characteristics can show from the real behavior , then we, according to experimenter's response situation in a certain specific situation, in order to the method that is measured indirectly , to estimate or infer the difference degree of various kinds of potential psychological specialities, how experimenters' psychological speciality has appeared in accurate and fast inference, it is exactly what we need urgently.
The present in implementing trends towards computerized adaptive testing, use shellfish's theory of the cause and effect of the network, direct different colonies , such as individual , class , group or the school of class ,etc. against students, one of construction can be analysed question of diagnosing, predict the system of student's learning ability , help students to find out those one's own basic conception without clear , help the teacher remedy the reference of teaching after class ing. In this section we present an example of an adaptive test that was used to diagnose person s skills The adaptive test uses Bayesian networks to model a tested student and the given questions。
Classical approach used in educational and psychological testing is item response theory IRT 。The presented application of Bayesian network can be regarded as a generalization of the multidimensional IRT It brings two basic advantages 。
1. It can better reflect the student reasoning process and provides better insight into the modelled problem 。
2. The student model encode dependence between skills 。Therefore adaptive tests can be substantially shortened while the test precision is kept
This paper probes into the mathematic unit of square measure of grade 4， using the evidence of test questions to train Bayesian Networks ， and uses AO* Algorithms as strategy of selecting items to construct question structures， compare with Euclidean distance that provides to implement the real learning diagnose on-line system。 By using system to analyze the following questions；First , is the real learning diagnose on-line system feasible？Second，does adaptive selecting item on real learning diagnose on-line system have any efficiency？
So we collected the situation of students answered items，using Bayesian Networks and Euclidean distance as tools of analyzing，discussing its correct rate，and doing the entity whether real learning diagnose on-line system is a feasible reference or not。

並列關鍵字

ＡＯ* Algorithms ； Entropy ； Euclidean distance

參考文獻

何明鏡，劉湘川，郭伯臣(民93)。以知識結構為主的適性診斷測驗系統之研究-以高階語言C++之算術邏輯運算及輸入輸出函數為例。網路學習評量系統。網路學習理論與應用學術研討會。新竹，交通大學。

Almond,R.G.,Mislevy, Robert J. (1999) Graphical models and computerized adaptive testing. Applied Psychological Measurement,23(3):223-237.

Johnsonbaugh,R., Schaefer,M. (2004).Dynamic programming ,algorithms,P.323.

Mislevy, Robert J., Almond, Russell G., Lukas, Janice F.(2003).A Brief Introduction to Evidence-Centered Design.

Vomlel,J.(2002)Evidence Propagation in Bayesian Networks for Computerized Adaptive Testing .Aalborg University .http://www.cs.auc.dk/~jirka