Investigating the Prerequisites for a robust Neurotutor: The Detection of mixed User States containing Working Memory Load, Affective Valence and Affective Dominance
Year: 2017
Type: dissertation
Abstract: Intelligent tutoring systems are software environments that aim to simulate a human tutor. While current systems show effectiveness comparable to human tutors, they still suffer from the ‘assistance dilemma’. This drawback refers to the inability to infer the ongoing user state which can lead to situations where the system provides no or inadequate support. To alleviate this situation, user state detection has been implemented in some systems. However, at the current time, only behavioral indicators are used to infer the ongoing user state. Such overt behaviors are not specific enough to provide a detailed representation of the user state. This is the reason why I suggest to investigate the potential use of the electroencephalogram to infer the ongoing user state. This combination of an intelligent tutoring system and an EEG-based user state detection is called a neurotutor. EEG-based user state detection usually focusses on narrow user states which can be detected in controlled lab environments. I assume that real-life environments like a classroom evoke complex user states which consist of multiple different components. I therefore propose a three component framework that enables the tracking of different processes that are active during a complex user state. The first two studies focus on the separation of working memory load and affective valence in a highly controlled setting with the use of established measures from classical neuroscience. I found that measures used to infer working memory load can be used to track changes in working memory load under different affective valence. Furthermore, I found that said measures were also sensitive to changes in affective valence. Surprisingly, I found that measures used to infer affective valence were not sensitive to changes in affective valence under working memory load. Additional analyses revealed that working memory load and affective valence can be automatically detected with accuracies sufficient for the use in a neurotutor. The third study successfully replicated the findings from the first two studies in a more realistic, although less controlled setting. A simplified learning game was used to induce the complex user state of perceived loss of control that simultaneously evoked cognitive as well as affective processes. With the help of the framework I was able to integrate the findings from three different studies that all analyzed the same dataset. This would not have been possible without an adequate theoretical framework.
Source: Universitätsbibliothek Tübingen
Author: Grissmann, Sebastian
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W2606329649
