@article{fahid_rowe_spain_goldberg_pokorny_lester_2022, title={Robust Adaptive Scaffolding with Inverse Reinforcement Learning-Based Reward Design}, volume={13356}, ISBN={["978-3-031-11646-9"]}, ISSN={["1611-3349"]}, DOI={10.1007/978-3-031-11647-6_35}, abstractNote={Reinforcement learning (RL) has shown significant potential for inducing data-driven scaffolding policies but designing reward functions that lead to effective policies is challenging. A promising solution is to use inverse RL to learn a reward function from effective demonstrations. This paper presents an inverse reward deep RL framework for inducing scaffolding policies in an adaptive learning environment. The framework centers on generating a data-driven model of immediate rewards by sampling high learning-gain episodes from previous student interactions and applying inverse RL. The resulting reward model is used to induce an adaptive scaffolding policy using batch constrained deep Q-learning. We evaluate this framework on data from 487 learners who completed an adaptive trianing course that provided direct instruction on principles of leading stability operations. Results show that the framework yields significantly better scaffolding policies more quickly compared to several RL baselines.}, journal={ARTIFICIAL INTELLIGENCE IN EDUCATION: POSTERS AND LATE BREAKING RESULTS, WORKSHOPS AND TUTORIALS, INDUSTRY AND INNOVATION TRACKS, PRACTITIONERS AND DOCTORAL CONSORTIUM, PT II}, author={Fahid, Fahmid Morshed and Rowe, Jonathan P. and Spain, Randall D. and Goldberg, Benjamin S. and Pokorny, Robert and Lester, James}, year={2022}, pages={204–207} }
 @article{spain_rowe_smith_goldberg_pokorny_mott_lester_2021, title={A reinforcement learning approach to adaptive remediation in online training}, volume={7}, ISSN={["1557-380X"]}, DOI={10.1177/15485129211028317}, abstractNote={ Advances in artificial intelligence (AI) and machine learning can be leveraged to tailor training based on the goals, learning needs, and preferences of learners. A key component of adaptive training systems is tutorial planning, which controls how scaffolding is structured and delivered to learners to create dynamically personalized learning experiences. The goal of this study was to induce data-driven policies for tutorial planning using reinforcement learning (RL) to provide adaptive scaffolding based on the Interactive, Constructive, Active, Passive framework for cognitive engagement. We describe a dataset that was collected to induce RL-based scaffolding policies, and we present the results of our policy analyses. Results showed that the best performing policies optimized learning gains by inducing an adaptive fading approach in which learners received less cognitively engaging forms of remediation as they advanced through the training course. This policy was consistent with preliminary analyses that showed constructive remediation became less effective as learners progressed through the training session. Results also showed that learners’ prior knowledge impacted the type of scaffold that was recommended, thus showing evidence of an aptitude–treatment interaction. We conclude with a discussion of how AI-based training can be leveraged to enhance training effectiveness as well as directions for future research. }, journal={JOURNAL OF DEFENSE MODELING AND SIMULATION-APPLICATIONS METHODOLOGY TECHNOLOGY-JDMS}, author={Spain, Randall and Rowe, Jonathan and Smith, Andy and Goldberg, Benjamin and Pokorny, Robert and Mott, Bradford and Lester, James}, year={2021}, month={Jul} }
 @article{fahid_rowe_spain_goldberg_pokorny_lester_2021, title={Adaptively Scaffolding Cognitive Engagement with Batch Constrained Deep Q-Networks}, volume={12748}, ISBN={["978-3-030-78291-7"]}, ISSN={["1611-3349"]}, DOI={10.1007/978-3-030-78292-4_10}, abstractNote={Scaffolding student engagement is a central challenge in adaptive learning environments. The ICAP framework defines levels of cognitive engagement with a learning activity in terms of four different engagement modes—Interactive, Constructive, Active, and Passive—and it predicts that increased cognitive engagement will yield improved learning. However, a key open question is how best to translate the ICAP theory into the design of adaptive scaffolding in adaptive learning environments. Specifically, should scaffolds be designed to require the highest levels of cognitive engagement (i.e., Interactive and Constructive modes) with every instance of feedback or knowledge component? To answer this question, in this paper we investigate a data-driven pedagogical modeling framework based on batch-constrained deep Q-networks, a type of deep reinforcement learning (RL) method, to induce policies for delivering ICAP-inspired scaffolding in adaptive learning environments. The policies are trained with log data from 487 learners as they interacted with an adaptive learning environment that provided ICAP-inspired feedback and remediation. Results suggest that adaptive scaffolding policies induced with batch-constrained deep Q-networks outperform heuristic policies that strictly follow the ICAP model without RL-based tailoring. The findings demonstrate the utility of deep RL for tailoring scaffolding for learner cognitive engagement.}, journal={ARTIFICIAL INTELLIGENCE IN EDUCATION (AIED 2021), PT I}, author={Fahid, Fahmid Morshed and Rowe, Jonathan P. and Spain, Randall D. and Goldberg, Benjamin S. and Pokorny, Robert and Lester, James}, year={2021}, pages={113–124} }
 @article{min_spain_saville_mott_brawner_johnston_lester_2021, title={Multidimensional Team Communication Modeling for Adaptive Team Training: A Hybrid Deep Learning and Graphical Modeling Framework}, volume={12748}, ISBN={["978-3-030-78291-7"]}, ISSN={["1611-3349"]}, DOI={10.1007/978-3-030-78292-4_24}, abstractNote={Team communication modeling offers great potential for adaptive learning environments for team training. However, the complex dynamics of team communication pose significant challenges for team communication modeling. To address these challenges, we present a hybrid framework integrating deep learning and probabilistic graphical models that analyzes team communication utterances with respect to the intent of the utterance and the directional flow of communication within the team. The hybrid framework utilizes conditional random fields (CRFs) that use deep learning-based contextual, distributed language representations extracted from team members' utterances. An evaluation with communication data collected from six teams during a live training exercise indicate that linear-chain CRFs utilizing ELMo utterance embeddings (1) outperform both multi-task and single-task variants of stacked bidirectional long short-term memory networks using the same distributed representations of the utterances, (2) outperform a hybrid approach that uses non-contextual utterance representations for the dialogue classification tasks, and (3) demonstrate promising domain-transfer capabilities. The findings suggest that the hybrid multidimensional team communication analysis framework can accurately recognize speaker intent and model the directional flow of team communication to guide adaptivity in team training environments.}, journal={ARTIFICIAL INTELLIGENCE IN EDUCATION (AIED 2021), PT I}, author={Min, Wookhee and Spain, Randall and Saville, Jason D. and Mott, Bradford and Brawner, Keith and Johnston, Joan and Lester, James}, year={2021}, pages={293–305} }
 @article{geden_smith_campbell_spain_amos-binks_mott_feng_lester_2019, title={Construction and Validation of an Anticipatory Thinking Assessment}, volume={10}, ISSN={["1664-1078"]}, url={http://www.scopus.com/inward/record.url?eid=2-s2.0-85077306542&partnerID=MN8TOARS}, DOI={10.3389/fpsyg.2019.02749}, abstractNote={Anticipatory thinking is a critical cognitive skill for successfully navigating complex, ambiguous systems in which individuals must analyze system states, anticipate outcomes, and forecast future events. For example, in military planning, intelligence analysis, business, medicine, and social services, individuals must use information to identify warnings, anticipate a spectrum of possible outcomes, and forecast likely futures in order to avoid tactical and strategic surprise. Existing methods for examining anticipatory thinking skill have relied upon task-specific behavioral measures or are resource-intensive, both of which are challenging to scale. Given the increasing importance of anticipatory thinking in many domains, developing a generic assessment of this skill and identifying the underlying cognitive mechanisms supporting it are paramount. The work reported here focuses on the development and validation of the anticipatory thinking assessment (ANTA) for measuring the divergent generative process of anticipatory thinking. Two-hundred and ten participants completed the ANTA, which required them to anticipate possible risks, opportunities, trends, or other uncertainties associated with a focal topic. Responses to the anticipatory thinking and divergent thinking tasks were rated by trained raters on a five-point scale according to the uniqueness, specificity, and remoteness of responses. Results supported the ANTA’s construct validity, convergent validity, and discriminant validity. We also explored the relationship between the ANTA scores and certain psychological traits and cognitive measures (need for cognition, need for closure, and mindfulness). Our findings suggest that the ANTA is a psychometrically valid instrument that may help researchers investigate anticipatory thinking in new contexts.}, journal={FRONTIERS IN PSYCHOLOGY}, author={Geden, Michael and Smith, Andy and Campbell, James and Spain, Randall and Amos-Binks, Adam and Mott, Bradford W. and Feng, Jing and Lester, James}, year={2019}, month={Dec} }
 @article{ozer_penilla_spain_mott_woodson_lester_2019, title={HEALTH QUEST: PROMOTING ADOLESCENTS' HEALTH SCIENCE CAREER INTERESTS THROUGH TECHNOLOGY-RICH LEARNING EXPERIENCES}, volume={64}, ISSN={["1879-1972"]}, DOI={10.1016/j.jadohealth.2018.10.279}, abstractNote={The purpose of the Health Quest project is to create an intelligent game-based learning environment and associated resources to increase adolescents' knowledge of, interest in and self-efficacy to pursue health science careers. Science, technology, engineering and mathematics (STEM) fields are among the fastest growing career opportunities, yet women and ethnic minorities remain underrepresented in these fields. As adolescence offers a key window to promote interest in and increase self-efficacy to pursue health research careers and the health professions, Health Quest focuses on an adolescent student population. Health Quest will be pilot tested and used in classrooms across North Carolina and California, and will also be featured in informal learning settings including the North Carolina Museum of Natural Sciences and in after-school programs in San Francisco. To support the project's first aim of designing and developing a series of Health Quest Career Adventure Game episodes and interactive video interviews to promote students' interest in health science careers, the project team conducted an extensive review of recent NIH Science Education Partnership Award (SEPA) projects awarded over the past three years. The purpose of the review was to identify existing programs and online resources that have been developed to promote students' interest in health science careers and to identify any unique resources or game-based learning environments that aim to promote self-efficacy or mastery experiences towards building competency and interest in health science career fields. We conducted a review of existing SEPA projects awarded from 2015 to 2018 to identify online resources and game-based learning experiences that have been designed to promote students' interest in health science careers. Projects were reviewed according to content focus, instructional approach and resources provided (e.g., websites, podcasts, videos and online-games). Forty five of the 48 SEPA projects reviewed had websites. Projects focused primarily on the life sciences, such as genomics. A majority of projects involved engaging students in the classroom, including providing opportunities in the lab, and used an experiential and mentor-based approach to provide students with unique hands-on learning opportunities to promote students' knowledge and interest in STEM fields. Other projects involved teacher development. Game-based learning technologies offer significant potential for increasing students' interest in health science careers. The results of our review showed that while SEPA programs focus on promoting interests and STEM opportunities by engaging students in hands-on classroom and laboratory-based activities, there is a lack of technology-based resources that can be used to extend outreach to students who would not otherwise have access to these programs. We hypothesize that by leveraging the intrinsic motivation of game-play, participant agency, and personalized learning, we can create engaging learning experiences that enable students to explore and gain confidence in pursuing health science careers.}, number={2}, journal={JOURNAL OF ADOLESCENT HEALTH}, author={Ozer, Elizabeth M. and Penilla, Carlos and Spain, Randall D. and Mott, Bradford W. and Woodson, Donald and Lester, James C.}, year={2019}, month={Feb}, pages={S134–S134} }