| 摘要: | 學習同伴系統 (Learning Companion System) 是一種電腦輔助學習系統，其中電腦藉由虛擬一個或多個的擬人化成員來提供學生一個社會學習的環境。其中由系統所虛擬的非權威、非專家的擬人化成員稱之為「學習同伴」。這些「學習同伴」可以成為學生的競爭者、指導者、被指導者或是合作者。有兩種方法可以來製作「學習同伴」的領域以及學習能力：機器學習(machine learning)與模擬(simulation)。採用模擬方法的優點是比較容易製作多個的「學習同伴」，也比較容易控制這些「學習同伴」來符合學習理論或是個別學生的需求。再者「量化模擬」(quantitative simulation)適合於製作學習能力中的「技能熟練」(skill refinement)。然而目前尚未有相關研究是採用「量化模擬」來製作一個扮演被指導者的「學習同伴」。另一方面有一些學者採用「量化模擬」來製作一個扮演競爭者的「學習同伴」，然而這些「學習同伴」只擁有整體性的熟練度，而欠缺個別詳細的技能熟練度。 本研究的目的是要提出一個採用「量化模擬」的「學習同伴」製作方法。這個方法必須適用於解題領域(problem solving)，支援製作多個能夠扮演被指導者和競爭者、且具備個別詳細技能熟練度的「學習同伴」。解題領域是一個「學習同伴」需要運用「技能熟練」學習能力的領域。本論文所提出的方法稱之為QSLCM(Quantitative-Simulation based Learning Companion Modeling)，此論文介紹了QSLCM方法的架構、步驟、解題模型以及扮演被指導者、競爭者和指導者的模擬機制和規則。QSLCM的架構可以視為傳統「智慧型導學系統」(Intelligent Tutoring System)架構的擴充。此架構同時也提供了一個途徑來比較「學習同伴系統」與「智慧型導學系統」。QSLCM採用兩個「學生模型製作」(user modeling)的技術：「覆蓋模型」(overlay model)與「錯誤模型」(bug model)。「學生模型製作」是觀察學生的行為來建立學生模型。而QSLCM則是「學生模型製作」的相反機制，也就是建立一個代表「學習同伴」的「學習同伴模型」(learning companion pattern)用來虛擬「學習同伴」的行為。採用「覆蓋模型」，「學習同伴」的可能解法是專家的子集合。而採用「錯誤模型」可以讓「學習同伴」犯錯。QSLCM採用機率來表現「學習同伴」技能熟練度與解題的不確定性。QSLCM藉由多個「學習同伴模型」來支援製作多個「學習同伴」。藉由修改「學習同伴模型」就可以改變該「學習同伴」的解題行為。 本論文也介紹了兩個採用QSLCM來製作的「學習同伴系統」。DwestAgent系統展現QSLCM支援製作多個具備不同程度與扮演不同角色「學生同伴」的能力。DwestAgent提供三個分別具備專家、中等生與新手程度的「學生同伴」。每一個「學生同伴」都可以扮演競爭者、被指導者與指導者。RTS系統是一個支援「教學相長」的「學習同伴系統」。RTS提供一個輪流扮演指導者與被指導者的「學習同伴」。 LCS (Learning Companion System), a type of computer assisted learning systems, simulates one or more artificial human-like agents to provide users with a social learning environment. The non-authoritative or non-expert agents of LCSs, called learning companions, can act as the competitor, tutor, tutee, or collaborator of the user. Two approaches exist for implementing the knowledge and learning ability of a learning companion: machine learning and simulation. The advantages of the simulation approach are that implementing multiple learning companions, and tailoring the learning companion to the requirements of learning theory and individual users is more convenient. Furthermore, quantitative simulation is suitable for supporting refinement of learning skills. However, applying quantitative simulation to modeling a learning companion as the tutee of a user remains little researched. On the other hand, some researchers have applied quantitative simulation to modeling a learning companion as a user’s competitor in their LCSs. However, these systems only provide a learning companion with an overall proficiency. The learning companion does not have special proficiencies in different skills. This thesis aims to propose a quantitative simulation approach to modeling multiple learning companions in the domain of problem solving, particularly, modeling learning companions as tutees or competitors of users, and with specialized skill proficiencies. Problem solving is a domain, in which learning companions require implementation of skill refinement. The approach, called QSLCM (Quantitative-Simulation based Learning Companion Modeling), involves architecture, steps, problem solving modeling, and simulation of the roles of competitor, tutee, or tutor. Several heuristic rules for the approach are also proposed herein. QSLCM architecture is a generalization of the typical Intelligent Tutoring System (ITS) architecture. This architectural view provides a means of comparing the basic differences between ITSs and LCSs. QSLCM enables modeling learning companions by applying two techniques of user modeling, the overlay model and the bug model. User modeling aims to base the construction of the user’s model on observation of the user’s behavior. QSLCM is a reverse process of user modeling; namely, constructing a learning companion pattern to simulate the behavior of a learning companion. Applying the overlay model, the learning companion’s solutions to problems are a subset of the solutions of an expert. Meanwhile, applying the bug model allows the learning companion to make mistakes. In QSLCM, probabilities represent the uncertain skill proficiency behavior of the learning companion. Meanwhile, QSLCM supports modeling different kinds of learning companions by setting different learning companion patterns. Furthermore, QSLCM makes it convenient to change the behavior of learning companions simply by adjusting the learning companion pattern. Two applications are developed by applying QSLCM to simulate learning companions. The first application, called DwestAgent, presents that QSLCM can support modeling multiple learning companions with different roles and different skill levels. DwestAgent has three learning companions, with skill levels of an expert, average student, and novice, respectively. Each learning companion can act as competitor, peer tutor, or tutee. The second application, called RTS, supports the learning activity of reciprocal tutoring. RTS provides users with a learning companion, which can act as tutee and peer tutor. |
