BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Memento EPFL//
BEGIN:VEVENT
SUMMARY:Leveraging Reinforcement Learning for Enhancing Educational Enviro
 nments
DTSTART:20240112T150000
DTEND:20240112T170000
DTSTAMP:20260407T163929Z
UID:878f6ecbcc1d8fd9a1ca816486cef65adb096247b3db1e79028d4bef
CATEGORIES:Conferences - Seminars
DESCRIPTION:Bahar Radmehr\nEDIC candidacy exam\nExam president: Prof. Anto
 ine Bosselut\nThesis advisor: Prof. Tanja Käser\nCo-examiner: Prof. Cagla
 r Gulcehre\n\nAbstract\nDigital learning environments have increasingly be
 come\na cornerstone of modern education. This shift has spurred\nextensive
  research on applied Reinforcement Learning (RL) in\neducation\, focusing 
 on enhancing various aspects such as personalized\ncurriculum design\, pro
 viding customized hints\, modeling\nstudent learning behaviors\, and auto-
 grading. Despite the growth\nin research\, RL methods are not yet widely a
 dopted in a variety\nof learning platforms\, primarily due to the challeng
 es in adapting\nagents for reasoning tasks in complex educational environm
 ents\nwith usually large\, natural language-based search spaces\, misalign
 ment\nbetween agent and human behaviors\, and limitations\non agent polici
 es by constrained instructional opportunities.\nThis doctoral candidacy pr
 oposal integrates insights from three\npapers to address these challenges 
 in RL application in education.\nThe first paper introduces a language-ena
 bled agent framework\nfor complex interactive reasoning tasks based on Beh
 avior\nCloning and Prompting Large Language Models. The second\npaper prop
 oses a behavior cloning approach to align AI behavior\nwith human strategi
 es\, using next-move and error prediction\nmodels that replicate humans’
  behavior at different skill levels.\nThe third paper explores enhancing R
 L agents in constrained\ninstructional settings by identifying critical de
 cisions for desired\noutcomes. Our proposed research agenda\, inspired by 
 these\npapers\, aims to foster RL adoption in education by developing\nlan
 guage-enabled agents with reasoning and planning capabilities\ntailored to
  complex educational environments\, aligning agent\nbehavior with human be
 havior\, and dentifying critical decisions\nneeded for constrained tutorin
 g.\n\nBackground papers\n1. Pick the Moment: Identifying Critical Pedagog
 ical Decisions Using Long-Short Term Rewards\n2. Aligning Superhuman AI 
 with Human Behavior: Chess as a Model System\n3. SwiftSage: A Generative 
 Agent with Fast and Slow Thinking for Complex Interactive Tasks\n 
LOCATION:INF 220 https://plan.epfl.ch/?room==INF%20220
STATUS:CONFIRMED
END:VEVENT
END:VCALENDAR