Applying Probabilistic Models for Knowledge Diagnosis and Educational Game Design

摘要: Computer-based learning environments offer the potential for innovative assessments of student knowledge and personalized instruction learners. However, there are a number challenges to realizing this potential. Many psychological models not specific enough directly deploy in instructional systems, computational can arise when considering implications particular theory learning. While learners' interactions with virtual encode significant information about their understanding, existing statistical tools insufficient interpreting these interactions. This research develops teaching combines machine algorithms interpret actions customize based on interpretations. approach results frameworks that be adapted variety educational domains, clearly separating components shared across tasks customized content. Using approach, dissertation addresses three major questions: (1) How one diagnose from behavior games laboratories? (2) predict whether game will diagnostic knowledge? (3) computer-based tutor model domain?The first question involves automatically assessing via observed complex interactive environments, such as laboratories games. These require students plan take multiple achieve goals. Unlike many traditional assessments, students' independent given each individual action cannot classified correct or incorrect. To address issue, I develop Bayesian inverse planning framework inferring observing actions. The is variation reinforcement uses Markov decision processes how people choose knowledge. Through behavioral experiments, show infer stated beliefs, accuracy similar human observers, feedback improves efficiency. extend applications outside laboratory, extended algebra skills worked solutions linear equations, different sources mathematical errors. tested by developing an online provides opportunity practice solving equations diagnoses understanding after they have solved sufficient equations. Preliminary experiments demonstrate good fit majority participants' behaviors, its consistent more conventional assessment.The previous studies showed all result learner used perfectly In cases, may ambiguous, resulting diagnosis places some probability possible state another. developed optimal design much gained player players' if were play game: gaining means less ambiguous. extends experiment methods statistics. It limit trial error necessary create education suggesting choices while still leveraging designer initial design. Behavioral concept predicted gain correlated actual best designs twice uninformed design.The final part considers personalize computer tutor, relying domain estimate builds idea broadly sequence assessment instruction. cost time spent assessment, could alternatively been allowing work through new material; however, also beneficial providing allow material effectively. partially observable tutoring process decide what pedagogical learner. automated policies faster numeric concepts than baseline policies.My demonstrates applying modeling diverse set problems computer-assisted responding data. provide systematic scalable way responses technologies only content learners but guidance