ILMDA: An Intelligent Learning Materials Delivery Agent and Simulation Leen-Kiat Soh, Todd Blank, L. D. Miller, Suzette Person Department of Computer Science and Engineering University of Nebraska, Lincoln, NE {lksoh, tblank, lmille, sperson} @cse.unl.edu Introduction Traditional Instruction http://battellemedia.com/archives/old%20book%206.gif ourworld.compuserve.com/homepages/g_knott/lecturer.gif Introduction Intelligent Tutoring Systems – – – – Interact with students Model student behavior Decided which materials to deliver All ITS are adaptive, only some learn Related Work Intelligent Tutoring Systems – PACT, ANDES, AutoTutor, SAM These lack machines learning capabilities – – – They generally do not adapt to new circumstances Do not self-evaluate and self-configure their own strategies Do not monitor usage history of content presented to students Project Framework Learning material components – – – A tutorial A set of related examples A set of exercise problems Project Framework Underlying agent assumptions – – A student’s behavior is a good indicator how well the student is understanding the topic in question It is possible to determine the extent to which a student understands the topic by presenting different examples Methodology ILMDA System – – – Graphical user interface front-end MySQL database backend ILMDA reasoning in-between Methodology Overall methodology Methodology Flow of operations Under the hood – – – – Case-based reasoning Machine Learning Fuzzy Logic Retrieval Outcome Function Learner Model Student Profiling – Student background – Relatively static First and last name, major, GPA, interests, etc. Student activity Real-time behavior and patterns Average number of mouse clicks, time spent in tutorial, number of quits after tutorial, number of successes, etc. Case-based reasoning Each case contains problem description and solution parameters The casebase is maintained separately from the examples and problems Chooses example or problem for students with most similar solution parameters Solution Parameters Solution Parameters Description TimesViewed The number of times the case has been viewed DiffLevel The difficulty level of the case between 0 and 10 MinUseTime The shortest time, in milliseconds, a single student has viewed the case MaxUseTime The longest time, in milliseconds, a single student has viewed the case AveUseTime The average time, in milliseconds, a single student has viewed the case Bloom Bloom’s Taxonomy Number AveClick The average number of clicks the interface has recorded for this case Length The number of characters in the course content for this case Content The stored list of interests for this case Adaptation Heuristics Adapt the solution parameters for the old case – – – Based on difference between problem description of old and new cases Each heuristic is weighted and responsible for one solution parameter Heuristics are implemented in a rulebase that adds flexibility to our design Simulated Annealing Used when adaptation process selects an old case that has repeatedly led to unsuccessful outcome Rather than remove old case SA is used to refresh its solution parameters Implementation End-to-end ILMDA – – – Applet-based GUI front-end CBR-powered agent Backend database system ILMDA simulator Simulator Consists of two distinct modules – Student Generator – Creates virtual students Nine different types student types based on aptitude and speed Outcome Generator Simulates student interactions and outcomes Student generator Creates virtual students – – Generates all student background values such as names, GPAs, interests, etc Generates the activity profile such as average time spent on session and average number of mouse clicks using Gaussian distribution Outcome Generator Simulates student interaction and outcomes – – Determines the time spent and the number of clicks for one learning material Also determines whether a virtual student quits the learning material and answers it successfully Simulation 900 students, 100 from each type – – – Step 1: 1000 iterations with no learning Step 2: 100 iterations with learning Step 3: 1000 iterations again with no learning Results – – Between Steps 1 and 3, average problem scores increased from 0.407 to 0.568 Between Steps 1 and 3, the number of examples given increased twofold Future Work Deploy the ILMDA system to the introductory CS core course – – – Fall 2004 Spring 2005 Fall 2005 (done) (done) Add fault determination capability – Students || Agent Reasoning || Content at fault Questions Responses I Blooms Taxonomy (Cognitive) – – – – – – Knowledge: Recall of data. Comprehension: Understand the meaning, translation, interpolation, and interpretation of instructions and problems. State a problem in one's own words. Application: Use a concept in a new situation or unprompted use of an abstraction. Applies what was learned in the classroom into novel situations in the workplace. Analysis: Separates material or concepts into component parts so that its organizational structure may be understood. Distinguishes between facts and inferences. Synthesis: Builds a structure or pattern from diverse elements. Put parts together to form a whole, with emphasis on creating a new meaning or structure. Evaluation: Make judgments about the value of ideas or materials. http://www.nwlink.com/~donclark/hrd/bloom.html Responses II Outcome function (example or problem) – – – Ranges from 0..1 Quitting at tutorial or example results in 0 for outcome Otherwise, compare average clicks and times for student with those for example or problem