Date on Master's Thesis/Doctoral Dissertation

12-2016

Document Type

Doctoral Dissertation

Degree Name

Ph. D.

Department

Electrical and Computer Engineering

Degree Program

Electrical Engineering, PhD

Committee Chair

Graham, James H.

Committee Co-Chair (if applicable)

Welch, Karla

Committee Member

Welch, Karla

Committee Member

Pennington, Robert

Committee Member

Inanc, Tamer

Committee Member

Hieb, Jeff

Author's Keywords

intelligent tutoring systems; affective computing; wearable technology; small-group instruction; machine learning; observational learning

Abstract

Autism spectrum disorder (ASD) is a neurodevelopmental disorder characterized by ongoing problems in social interaction and communication, and engagement in repetitive behaviors. According to Centers for Disease Control and Prevention, an estimated 1 in 68 children in the United States has ASD. Mounting evidence shows that many of these individuals display an interest in social interaction with computers and robots and, in general, feel comfortable spending time in such environments. It is known that the subtlety and unpredictability of people’s social behavior are intimidating and confusing for many individuals with ASD. Computerized learning environments and robots, however, prepare a predictable, dependable, and less complicated environment, where the interaction complexity can be adjusted so as to account for these individuals’ needs. The first phase of this dissertation presents an artificial-intelligence-based tutoring system which uses an interactive computer character as a pedagogical agent (PA) that simulates a human tutor teaching sight word reading to individuals with ASD. This phase examines the efficacy of an instructional package comprised of an autonomous pedagogical agent, automatic speech recognition, and an evidence-based instructional procedure referred to as constant time delay (CTD). A concurrent multiple-baseline across-participants design is used to evaluate the efficacy of intervention. Additionally, post-treatment probes are conducted to assess maintenance and generalization. The results suggest that all three participants acquired and maintained new sight words and demonstrated generalized responding. The second phase of this dissertation describes the augmentation of the tutoring system developed in the first phase with an autonomous humanoid robot which serves the instructional role of a peer for the student. In this tutoring paradigm, the robot adopts a peer metaphor, where its function is to act as a peer. With the introduction of the robotic peer (RP), the traditional dyadic interaction in tutoring systems is augmented to a novel triadic interaction in order to enhance the social richness of the tutoring system, and to facilitate learning through peer observation. This phase evaluates the feasibility and effects of using PA-delivered sight word instruction, based on a CTD procedure, within a small-group arrangement including a student with ASD and the robotic peer. A multiple-probe design across word sets, replicated across three participants, is used to evaluate the efficacy of intervention. The findings illustrate that all three participants acquired, maintained, and generalized all the words targeted for instruction. Furthermore, they learned a high percentage (94.44% on average) of the non-target words exclusively instructed to the RP. The data show that not only did the participants learn nontargeted words by observing the instruction to the RP but they also acquired their target words more efficiently and with less errors by the addition of an observational component to the direct instruction. The third and fourth phases of this dissertation focus on physiology-based modeling of the participants’ affective experiences during naturalistic interaction with the developed tutoring system. While computers and robots have begun to co-exist with humans and cooperatively share various tasks; they are still deficient in interpreting and responding to humans as emotional beings. Wearable biosensors that can be used for computerized emotion recognition offer great potential for addressing this issue. The third phase presents a Bluetooth-enabled eyewear – EmotiGO – for unobtrusive acquisition of a set of physiological signals, i.e., skin conductivity, photoplethysmography, and skin temperature, which can be used as autonomic readouts of emotions. EmotiGO is unobtrusive and sufficiently lightweight to be worn comfortably without interfering with the users’ usual activities. This phase presents the architecture of the device and results from testing that verify its effectiveness against an FDA-approved system for physiological measurement. The fourth and final phase attempts to model the students’ engagement levels using their physiological signals collected with EmotiGO during naturalistic interaction with the tutoring system developed in the second phase. Several physiological indices are extracted from each of the signals. The students’ engagement levels during the interaction with the tutoring system are rated by two trained coders using the video recordings of the instructional sessions. Supervised pattern recognition algorithms are subsequently used to map the physiological indices to the engagement scores. The results indicate that the trained models are successful at classifying participants’ engagement levels with the mean classification accuracy of 86.50%. These models are an important step toward an intelligent tutoring system that can dynamically adapt its pedagogical strategies to the affective needs of learners with ASD.