RELATE: REsearch in Learning, Assessing, and Tutoring Electronically RELATE.mit.edu Postdocs: Phil Dukes Sofia Morote Rasil Warnakulasooriya PI: Dave Pritchard $: MIT, NSF, DEP 1 Also known as by EET and CyberTutor.MIT in publications an expert system based on educational expertise, not AI The most advanced tutorial and assessment system in the world Made by Effective Educational Technologies, a Pritchard company. 2 Outline • Objective • Pedagogy that Works • Feedback – Closed Loop Education • Research from MIT • Revolution in Assessment • Closing Thoughts 3 Perspective Digital Education Future?! Broadcast Radio Passive Class Uniform Style Next Edition Teacher Author High Stakes Tests Two-way Radio Interactive Student Stylized (e.g. Audio) Next Day Coach Authors/Researchers Embedded Assessment 4 Why Homework?? Teachers’ Priorities: 1. Lectures 2. Exams 3. Notes and Demonstrations 4. Homework Students spend most time and learn most from 1. Homework 5 Books, lectures, most WWW education TWO WAY LEARNING Students, Teachers, Authors, Researchers Learn from each other DATA, EXPERIMENT, ANALYSIS, CONCLUSIONS 6 Teacher Student SOCRATIC LEARNING System Authors, Researchers 7 Outline Objective • Pedagogy that Works • Feedback – Closed Loop Education • Research from MIT • Revolution in Assessment • Closing Thoughts 8 Pedagogy Design Philosophy of myCyberTutor Emulate the interaction between a human tutor and a student. The tutor informs the teacher. The process informs the author. Results: an effective interactive learning tool you can author, deliver and improve content an expert program embodying your expertise 9 Pedagogy Student-Centered Instruction Mastery Learning The amount learned should be constant and time allowed to vary Constructivists Allow students to construct knowledge in their own way Student Others Socratic pedagogy and learning styles (to be implemented) 10 Pedagogy Pedagogical Principles • • • • • Actively engage the student Adapt problem to less skillful students with hints Prompt feedback addresses wrong answers Mastery Learning >90% get solution Declarative and procedural knowledge are both important hints and subproblems • Solidify and extend the solution followups • Free response answers reduce guessing 11 12 13 14 15 16 Results from MIT Gain on the MIT Final Exam December 2000 to May 2001 P-value 0.69 0.69 0.35 0.010 17 Results from MIT Gain on Force Concept Inventory data C. Ogilvie 2000 FCI gain=0.41 for course P-value 0.854 0.807 0.198 0.087 0.015 18 Outline Objective Pedagogy that Works • Feedback – Closed Loop Education • Research from MIT • Revolution in Assessment • Closing Thoughts 19 TWO WAY LEARNING Two Way Learning - Feedback myCyberTutor interactions: • Typical student returns to server 10 times during course of each problem (cf. Web Assign ~4 times per assignment) • Students achieve the correct answer 90% of the time (cf. ~60% first time right) • Students comment on ~3% of all problems – More if problem has flaws 20 Student Feedback – Closed Loop Education Comments 21 Feedback – Closed Loop Education Wrong Answers 22 Feedback – Closed Loop Education Feedback to Author Improves Problems • < 90% correct: Need more hints • Wrong Answers: Respond to common ones • Comments: Revise wording, remove confusion, revise program • Time: Is this problem worthwhile? 23 Feedback Enables Revisions that Improve Problems Percent Answering Correctly Percent Requesting Solutions Average Wrong Answers /part Average hints/part Median Minutes/ part 83.5 % 12.3 % 1.51 0.76 1.5 Spring 2002 91.6 % 6.0% 0.89 0.83 1.5 Fall 2003 93.4% Spring 2001 Room for even more improvement!! 24 Outline Objective Pedagogy that Works Feedback – Closed Loop Education • Research from MIT • Revolution in Assessment • Closing Thoughts 25 Pedagogy RELATE: REsearch in Learning, Assessing, and Tutoring Electronically RELATE.mit.edu Postdocs: Phil Dukes Sofia Morote Rasil Warnakulasooriya 26 Objective Attractions for Researchers Data with resolution Capability for Split Class Assignments Your Data Package Log of your students’ interactions Anonymous student # vs name & ID for your class Database of your class’ assignments Plus - General Data Package Performance Data - your class vs. standard SML skeleton each problem (subparts, hints, etc.) Format key 27 Inductive vs. Deductive Instruction Inductive: Students learn by doing a problem from the hints, from the subproblems by figuring it out from feedback Transfer learning to tutorial questions?? Deductive: Students learn from a tutorial from the learning goal & text from the hints from the self-assessment questions Transfer learning to related problem?? 28 Tutorials • Tutorial problems in Mastering Physics are carefully planned and sequenced instruction with SAQ’s • They are used as instructional material to impart principles in deductive learning. Pedagogy Problems • Problems require a student to apply an already familiar concept, formula, or procedure • Socratic help is available, including explanation of the concept, a formula that is needed, etc. • Related Problems cover same topic as adjacent tutorial 30 Deductive: Related Problem Difficulty Reduced by working Tutorial First Harmonic Os. Newton 3rd Law Torque Results from MIT After working tutorial p=0.01* p=0.03* p=0.06** 31 Inductive: Tutorial Difficulty Reduced by working Related Problem First? Harmonic Os. Newton 3rd Law Torque Results from MIT 32 Results from MIT Conclusion: Deductive Works • Interactive tutorials significantly increase performance on subsequent related problems (~25% less difficult) • Students don’t learn inductively from a multi-part example • We recommend using online tutorials in the old fashioned way - preparation for subsequent deductive exercises 33 Twice as much learning per unit time spent on the tutorial compared with time spent on the preparatory problem Tutorial-first 0.020 Problem-first 0.018 Improvement per unit time 0.016 0.014 0.012 0.010 0.008 0.006 0.004 0.002 0.000 Torque Newton III SHM 34 Prior related problem reduces the hints requested on the related problem by ~12% (based on 6 problems) Unprepared group Prepared group by solving a related problem p < 0. 05 0.6 p < 0. 01 0.5 0.4 0.3 0.2 0.1 0.7 fraction of students who request hints 0.7 Prior tutorial reduces the hints requested on the related problem by ~19% (based on 5 problems) 0.6 Unprepared group Prepared group by solving a tutorial problem p < 0. 005 0.5 0.4 p < 0. 1 0.3 0.2 0.1 0.0 Finding Torque 0.0 Colliding cars Newton III Cross-section for asteroid impact 35 Time to Completion The real time environment allows us to study how long it takes students to work problems, whether good students do problems quicker or slower, etc. We have discovered that there are Three groups of students in time: 36 When Students Finish: Three distinct groups Quick solvers < 2.5 minutes Real-time solvers 2.5 min – 2.2 hours Interrupted solvers > 2.2 hours Colliding cars 0.40 0.35 Rate of completion 0.30 0.25 0.20 0.15 Quick solvers Interrupted solvers 0.10 0.05 Real-time solvers 0.00 -0.05 7s 2 20s 55s 2.5m 6.7m 18m 50m 2.2h 6h 4 6 8 Ln(t) 10 17h 2d 12 5d 14 37 0.40 Finding torque Flywheel kinematics Parallel-axis theorem 0.4 0.35 Median time ~ 11min 0.30 Rate of completion Rate of completion Median time ~ 7min 0.3 0.2 0.1 Colliding cars Shooting a block up an incline 0.25 0.20 0.15 0.10 0.05 0.00 0.0 7s 2 20s 55s 2.5m 7m 18m 50m 2.2h 6h 17h 4 6 8 2d 10 5d -0.05 12 7s 20s 55s 2.5m 7m 18m 50m 2.2h 2 14 4 6 2d 10 5d 12 14 Ln(t) Ln(t) Cross-section for asteroid impact Post-collision orbit The parallel-axis theorem A person standing on a leaning ladder Collision at an angle 0.40 6h 17h 8 0.30 Median time ~ 30min Rate of completion 0.30 0.25 0.20 0.15 Rate of completion 0.25 Median time ~ 18-30min 0.35 0.20 0.15 0.10 0.10 0.05 0.05 0.00 -0.05 0.00 7s 2 20s 55s 2.5m 7m 18m 50m 2.2h 6h 17h 4 6 8 Ln(t) 10 2d 12 7s 5d 14 2 20s 55s 2.5m 7m 18m 50m 2.2h 4 6 8 Ln(t) 6h 17h 10 2d 12 5d 38 14 Real-time solvers make errors and ask for hints total no hints, no wrong ans. no hints, at least one wrong at least one hint, one wrong 0.25 Rate of completion 0.20 0.15 0.10 0.05 0.00 7s 2 20s 55s 2.5m 6.7m 18m 50m 2.2h 6h 17h 4 6 8 Ln(t) 10 2d 12 5d 14 39 0.25 total no hints, no wrong ans. no hints, at least one wrong at least one hint, one wrong Collision at an angle: Unprepared group total no hints, no wrong answers no hints, at least one wrong at least one hint, one wrong Collision at an angle: Prepared group 0.25 0.20 0.15 Rate of completion Rate of completion 0.20 0.10 0.05 0.15 0.10 0.05 0.00 0.00 7s 2 20s 55s 2.5m 6.7m 18m 50m 2.2h 6h 17h 4 6 8 Ln(t) 10 2d 12 5d 7s 14 2 20s 55s 2.5m 6.7m 18m 50m 2.2h 6h 17h 4 6 8 10 2d 12 5d 14 Ln(t) Note that: 1. The quick solvers do not make mistakes or ask for hints 2. The real-time solvers make mistakes and ask for hints 3. The interrupted solvers make mistakes and ask for hints 4. Fewer real-time solvers in the prepared group ask for hints 40 Fraction Finished curves with hints & feedback Collision at an angle Colliding cars 1.0 0.8 0.8 0.6 0.4 fraction fraction 0.6 64% 0.4 44% 0.2 0.2 0.0 0.0 7s 2 20s 55s 2.5m 6.7m 18m 50m 2.2h 6h 4 6 8 Ln(t) 10 17h 2d 12 5d 7s 14 2 20s 55s 2.5m 6.7m 18m 50m 2.2h 6h 4 6 8 10 17h 2d 12 5d 14 Ln(t) For 14 problems: fraction of real-time solvers = 65 4% 41 Time to completion curves without hints & feedback End-of-chapter 10.38 0.8 0.8 0.6 0.6 35% 0.4 End-of-chapter 10.40 1.0 fraction fraction 1.0 28% 0.4 0.2 0.2 0.0 0.0 7s 2 20s 55s 2.5m 6.7m 18m 50m 2.2h 6h 4 6 8 10 17h 2d 7s 5d 12 14 4 6 8 10 17h 2d 12 5d 14 Ln(t) Ln(t) End-of-chapter 10.46 1.0 2 20s 55s 2.5m 6.7m 18m 50m 2.2h 6h 0.8 For 3 typical homework problems: fraction 0.6 fraction of real-time solvers = 29 3% 0.4 0.2 0.0 7s 2 20s 55s 2.5m 6.7m 18m 50m 2.2h 6h 4 6 8 Ln(t) 10 17h 2d 12 5d 14 42 Outline Objective Pedagogy that Works Feedback – Closed Loop Education Research from MIT • Revolution in Assessment • Closing Thoughts 43 Low-Error Embedded Future - Assessment Assessment Imagine that a rich ship-owner has hired Socrates to tutor his children. At the end of the month he desires to assess the amount they have learned. Would you advise him to: a) Administer a standardized hour-long test to the children? b) Ask Socrates how much they have learned? MyCyberTutor Assessment has ~100 times less variance due to error than a good final exam! This Assessment gives ~6 times as reliable an assessment per unit of student time as a good final exam! 44 Future - Assessment Embedded Assessment Final Exam myCyberTutor vs. Final Exam SocraticTutor 0.5 Score (Odd Problems) Score (Odd Problems) 0.5 R2 = 0.986 0.4 0.3 0.2 0.1 R2 = 0.4053 0.4 0.3 0.2 0.1 0 0 0 0.1 0.2 0.3 0.4 Score (Even Problems) 0.5 0 0.1 0.2 0.3 0.4 0.5 Score (Even Problems) myCyberTutor •…is 6 times more reliable per unit time •…has ~100 times less error variance 45 Future - Assessment Assessment: Detailed Skill Profile 46 Predicting Final Exam Score Future - Assessment 1 0.9 R2 = 0.51 MIT Final Score 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 SocraticTutor Prediction Implication: MyCyberTutor can replace tests 47 Assessment myCyberTutor Assessment Implies: 1) More Accurate 2) Fine Grained Assessment on Subtopics 3) Immediate Remediation -Select Next Problem 4) JITT Guide for Teacher 5) Learning vs. Avoiding Lost Points 6) Predict Test Scores -Eliminate Tests 7) Incredible Tool for Education Research 8) Replace High Stakes Tests 48 Outline Objective Pedagogy that Works Feedback – Closed Loop Education Research from MIT Revolution in Assessment • Closing Thoughts 49 What you can gain »Write interactive problems »Educational Research on them »Educational Research in general What we can accomplish together »Partnerships in each - ideally all!! 50 Attractions for Problem Writers You can write truly interactive material Improve your problems from student data Author in just your area of research expertise Author tutorials, example problems, and problems involving applets New education research tools improve your problems and exercises 51 Attractions for Curriculum Developers New research tool with resolution of student’s capabilities and difficulties Improve instructional material via Feedback Experimentally Compare material and pedagogy Develop texts, tutorials, traditional exercises, and more sophisticated problem sequences 52 Credits and Thanks Effective Educational Technologies, Inc.* Alex Pritchard - sole programmer for 4 years Adam Morton - chief programmer David Kokorowski - content development Andrea Pritchard - president & treasurer & HR Postdocs and Undergrads Gabe Rockefeller - now at U. Arizona Phil Dukes - now at UTBrownsville Sofia Morote - now at Dennison College David Kokorowski - now at Effective Education Tech Rasil Warnakulasooriya - present Support: MIT - esp. Physics Dept. - support with TA’s NSF *DEP’s family has controlling financial interest 53 END 54 Results from MIT Student Opinion “How does the amount you learn per unit time with CyberTutor compare with time (including checking solutions) spent on written homework?” Much More Same Much Less MIT Semesters 55 Results from MIT Student Opinion “Would you recommend myCyberTutor for 8.01 next year?” Ratio of Yes to No MIT Semesters 56 Feedback – Closed Loop Education 57 Help for Authors Author training and manual Pedagogy instructional material Review first problems Instruction on using feedback data to improve problems, guidelines for wrong answers 58 Future - Assessment 85% Relilability is Unfair True Skill Observed Score 1/4 of those students who failed had a passing true score Another 1/4 should have failed, but passed 59 Objective Attractions for Teachers Tutoring available 24/7 Better than written Homework - not lower quality substitute Shift grading effort into instruction Teachers have instant access to detailed student performance data, facilitating Just-In-Time Teaching (JiTT) Problem Library with Metadata on difficulty, time, student rating, topics involved 60 Pedagogy Written Homework Assign Do homework Hand in Grade Learn from hw result? • No help for student when stuck • Feedback to student takes 1 week • Labor to grade • No feedback to teacher • Copying is easy 61 Pedagogy myCyberTutor is a Web-Based Homework Tutorial System Assign Learn while doing CyberTutor homework Detailed knowledge base • Provides immediate feedback and help to the student • Does the grading • Offers immediate feedback to the teacher • Supplies powerful insight into the student’s thinking 62 Pedagogy Tutorials • Tutorial problems in myCybertutor are carefully planned and sequenced instruction with SAQ’s • They are used as instructional material to impart principles in deductive learning. Outline Objective Pedagogy • Demo • Results from MIT • Feedback – Closed Loop Education • Future - Assessment • Closing Thoughts 64 Future - Assessment Assessment/Testing Error Any assessment has testing error - did you study problem that was on test? - careless mistakes? - lucky guess? How to determine reliability (reproducibility)? - compare two equivalent tests - split single test into two equivalent tests If test is reliable (error-free), split grades will correlate 65 Future - Assessment Final Exam 2001-2002 Reliability 0.85 s error =0.41 s observed 66