lessons learned in teaching mathematics with adaptive tutoring software

63
Department of Computer Science Ivon Arroyo University of Massachusetts Amherst Lessons Learned in Teaching Mathematics with Adaptive Tutoring Software To Erica

Upload: talon

Post on 23-Feb-2016

29 views

Category:

Documents


0 download

DESCRIPTION

Lessons Learned in Teaching Mathematics with Adaptive Tutoring Software. Ivon Arroyo University of Massachusetts Amherst. To Erica. Wayang Outpost --Math Tutoring System Grades 7,8,9,10 and community colleges. http:// Wayangoutpost.com. Empirical Learning Results since 2003. - PowerPoint PPT Presentation

TRANSCRIPT

Page 1: Lessons Learned in Teaching Mathematics with Adaptive Tutoring Software

Department of Computer Science

Ivon ArroyoUniversity of Massachusetts Amherst

Lessons Learned in Teaching Mathematics with Adaptive

Tutoring Software

To Erica

Page 2: Lessons Learned in Teaching Mathematics with Adaptive Tutoring Software

2

Wayang Outpost --Math Tutoring SystemGrades 7,8,9,10 and community colleges http://Wayangoutpost.com

Page 3: Lessons Learned in Teaching Mathematics with Adaptive Tutoring Software

3

MCASpassing%WayangMCASpassing%No Wayang

77% 60% **

34% 24% *

92% 76% *

WayangPosttest ControlNo Wayang76% 67% **

d=0.25

d=0.24

d=0.52

Empirical Learning Results since 2003After short exposure (3-4 hours)

Expanding to 2000 students in 2011

Page 4: Lessons Learned in Teaching Mathematics with Adaptive Tutoring Software

4

Wayang Outpost --Math Tutoring SystemStandardized-test math problems with multimedia help

More Help

http://Wayangoutpost.com

ModalityAnimationContiguity

Page 5: Lessons Learned in Teaching Mathematics with Adaptive Tutoring Software

5

What have we learned about how to teach math with advanced technologies?

Many things….

Some are supported by experimental evidence

Some are conjectures and anecdotes…

Page 6: Lessons Learned in Teaching Mathematics with Adaptive Tutoring Software

6

What have we learned about teaching math?

Showing Progress

Adaptive Problem Selection

Affect

Training Math Fluency

Offering Help

Page 7: Lessons Learned in Teaching Mathematics with Adaptive Tutoring Software

7

Lesson learned 1

Adaptive Math Tutoring that maintains students within a “zone of proximal development”

improves learning.

Page 8: Lessons Learned in Teaching Mathematics with Adaptive Tutoring Software

88

What kind of adaptivity?Murray, T.; Arroyo, I. (2002) Toward Measuring and Maintaining the Zone of Proximal Development in Adaptive Instructional Systems, Lecture Notes in Computer Science, 2002, Volume 2363/2002, 749-758

Can we understand when we are outside of the ZPD?

Little Effort

Too much effort

Frustrated

Fatigued

Page 9: Lessons Learned in Teaching Mathematics with Adaptive Tutoring Software

99

E(Ii)

IL IH

E(Hi)

HL HH

E(Ti)

TLTH

0 1 2 3 4 0 1 2 3 4 5 6 7

Incorrect Attempts Hints Time (each bar=5seconds)

Attempts < E(Ii) — IL Hints > E(Hi) + HH Time < E(Ti) — TL

Odd behavior: too much effort, or too little effort

Few Inc. Attempts Lots of Hints Little Time< > <

Learning what is high and low effortIn any problem pi i=1, .., N N=Total problems in system

Page 10: Lessons Learned in Teaching Mathematics with Adaptive Tutoring Software

10

Scenarios outside of the ZPD

Page 11: Lessons Learned in Teaching Mathematics with Adaptive Tutoring Software

1111

Where the ZPD is

HIGH

LOW

Little Effort

Too much effort

Frustrated

Fatigued

Disengaged

Murray, T.; Arroyo, I. (2002) Toward Measuring and Maintaining the Zone of Proximal Development in Adaptive Instructional Systems, Lecture Notes in Computer Science, 2002, Volume 2363/2002, 749-758

Page 12: Lessons Learned in Teaching Mathematics with Adaptive Tutoring Software

12

How to increase/decrease problem difficulty?

Arroyo, I., Mehranian, H., Woolf, B. (2010) Effort-based Tutoring: An Empirical Approach to Intelligent Tutoring. Proceedings of the 3rd International Conference on Educational Data Mining.

Pittsburgh, PA.

Page 13: Lessons Learned in Teaching Mathematics with Adaptive Tutoring Software

13

Does this adaptivity improve learning?Randomized Controlled Experiment (N=56) Spring 2004

ANCOVA for Posttest Score F(55,1)=8.4, p=.006

Raw percent Correct (Pre and Posttest) Accuracy over attempted problems

Arroyo, I., Mehranian, H., Woolf, B. (2010) Effort-based Tutoring: An Empirical Approach to Intelligent Tutoring. Proceedings of the 3rd International Conference on Educational Data Mining.

Pittsburgh, PA.

Page 14: Lessons Learned in Teaching Mathematics with Adaptive Tutoring Software

14

Lesson learned 1

Adaptive Math Tutoring that attempts to maintain students within a “zone of proximal development” improves learning.

Being adaptive over smaller “chunks” of similar problems (instead of the full set of problems)

yields higher learningBeing “gentle” at increasing difficulty yields

higher learning

Page 15: Lessons Learned in Teaching Mathematics with Adaptive Tutoring Software

15

Lesson learned 2

Training Basic Arithmetic, not only for accuracy but for Speed to respond, enhances mathematics learning in combination with Wayang Outpost.

Page 16: Lessons Learned in Teaching Mathematics with Adaptive Tutoring Software

16

Math Facts Retrieval (fluency) training

Royer, J. M., & Tronsky, L. N. (1998). Addition practice with math disabled students improves subtraction and multiplication performance. In T. E. Scruggs and M. A. Mastropieri (Eds.), Advances in Learning and Behavioral Disabilities (Vol 12). Greenwich, Conn.: JAI Press, Inc.

Page 17: Lessons Learned in Teaching Mathematics with Adaptive Tutoring Software

1717

True Means and SD for HARD items of the standardized pretest and posttest

ANCOVA for Hard Posttest with Hard Pretest as covariate; MFR,Wayang fixed factors:Wayang F(222,1)=6.8, p=.01; WayangxMFR F(222,1)=6.8, p=.009

Post-Hoc Contrasts: Wayang > no-Wayang? Yes. Wayang-MFR > Wayang-NoMFR? Yes.

Results on Standard Math Test (Hard Items)

Page 18: Lessons Learned in Teaching Mathematics with Adaptive Tutoring Software

18

But why? Problem solving takes place in a cognitive system constrained by

a limited capacity of working memory

StrategyBasic Math

Working Memory Capacity When Solving a Math Problem

Doing Math is like speaking a language. If you are fluent, you will concentrate better on the message.

Basic Math

Math FluencyMath Fluency helps predict performance at state-wide tests“Basic Math” could be anything… such as solving easy equations…

Royer&Tronsky (1998)Royer et al. (1999)

Page 19: Lessons Learned in Teaching Mathematics with Adaptive Tutoring Software

19

Lesson learned 2

Training Basic Arithmetic, not only for accuracy but for Speed to respond, enhances mathematics learning in combination with Wayang Outpost.

Page 20: Lessons Learned in Teaching Mathematics with Adaptive Tutoring Software

20

Lesson learned 3

Showing students their historical improvement (not just their mastery level) at math problem solving improves engagement in subsequent problems

and increases learning.

Page 21: Lessons Learned in Teaching Mathematics with Adaptive Tutoring Software

2121

5 problems earlier Last 5 problems

Similar to open learner models, but focus on progress

Progress Monitoring InterventionsSelf-monitoring feedback Self-referenced-feedback (McColskey and Leary (1985))

Page 22: Lessons Learned in Teaching Mathematics with Adaptive Tutoring Software

2222

5 problems earlier Last 5 problems

Progress Monitoring InterventionsSelf-monitoring feedback Self-referenced-feedback (McColskey and Leary (1985))

Page 23: Lessons Learned in Teaching Mathematics with Adaptive Tutoring Software

2323

Mathematics Learning Results

+0%+7%

Means and Standard Deviations Percentages

+16%

+13%

ANCOVADependent Learning (Posttest-Pretest score)GroupCovariate Pretest score

F=4.23, p=.043

Effect size: Cohen’s d=0.4

Page 24: Lessons Learned in Teaching Mathematics with Adaptive Tutoring Software

24

Changes in time per problemMedians and Quartiles for time spent in the problem (in pairs of subsequent problems)

529303 529303N =

Seconds spent in a problem

MotivationalControl

Med

ian/

quar

tiles

for t

i me

spen

t per

pro

blem

70

60

50

40

30

20

10

0Before InterventionAfter Intervention

Tutor-InterventionTutor-Control

Page 25: Lessons Learned in Teaching Mathematics with Adaptive Tutoring Software

25

Lesson learned 3

Showing students their historical improvement (not just their mastery level) at math problem solving improves engagement in subsequent problems

and increases learning.

Page 26: Lessons Learned in Teaching Mathematics with Adaptive Tutoring Software

26

Lesson learned 4

Emotions/Attitudes/Affect as a more important long-term outcome than learning.

In general, students are really bored about mathematics.

Also, there are important group differences in students emotions, before tutoring.

Page 27: Lessons Learned in Teaching Mathematics with Adaptive Tutoring Software

27

Who needs more affective supportLow achieving students (95% of students with IEPs)

Table 1 . Affective self-reports of high-achieving vs. low-achieving students prior to Tutoring

Af fective Criterion Means, standard deviations and between-

subjects test Lo w-achie ving: N=64; High-achie ving: N=43

Self-concept of math abili ty (in comparison to other stud ents,

other subjects, 3 items)

Lo w-achie ving: M=3.2 SD=1.1 High-achie ving: M=4.1 SD=1.0

***F(106,1)=18.2, p=.000

How confident do you feel when solvi ng math prob lems?

Lo w-achie ving: M=3.1 SD=1.3 High-achie ving: M=4.0 SD=1.3

***F(105,1)=11.5, p=.001

How frus trating is it to solve math problems?

Lo w-achie ving: M=3.6 SD=1.2 High-achie ving: M=3.0 SD=1.1

** F(106,1)=7.6, p=.007

How exciting is it to solve math prob lems?

Lo w-achie ving: N=64 , M=2.2 SD=1.2 High-achie ving: N=43, M=2.7 SD=1.4

*F(106,1)=3.64, p=0.05

Page 28: Lessons Learned in Teaching Mathematics with Adaptive Tutoring Software

28Figure 1: Results for a pre-tutor survey in two public schools: Girls

develop negative feelings for mathematics, including decreased confidence (left) and increased frustration (right), between middle and

high school.

Who needs more affective supportHigh School Girls

Page 29: Lessons Learned in Teaching Mathematics with Adaptive Tutoring Software

29

Lesson learned 4

Emotions/Attitudes/Affect as a more important long-term outcome than learning.

In general, students are really bored about mathematics.

Also, there are important group differences in students emotions, before tutoring.

Page 30: Lessons Learned in Teaching Mathematics with Adaptive Tutoring Software

30

Lessons learned 5

We can understand and TRACE students at the affective level, understanding their emotions.

How? from very recent behaviors, and with the help of physiological sensors.

Page 31: Lessons Learned in Teaching Mathematics with Adaptive Tutoring Software

3131

Students Self-Report EmotionsEvery 5 minutes, students would report emotions

Page 32: Lessons Learned in Teaching Mathematics with Adaptive Tutoring Software

32

Affective Tracing

Anxiety BoredomFrustration

Predicting Emotions in Real-Time

Linear Models to Predict Emotions from last Problem SeenModels Created using Stepwise Regression

# Hints Seen

Solved? 1st Attempt

# Incorrectattempts

CharacterPresent?

Seconds to1st Attempt

Time in Tutor

Seconds To Solve

Tutor Context Variables (for the last problem)

R2=0.3 R2=0.15 R2=0.18R2=0.19

Enjoyment

Accuracy of a YES/NO prediction of each emotion, compared to TRUE self-report

86% 88% 78% 83%

Possible that looking at longer episodes of recent history will achieve as good accuracy as sensors.

SitForwardStdev

“Concen-trating”

SitForwardMean

R2=0.38 R2=0.31R2

=0.40

R2=0.44

“Interest”Min

MaxPressure

Page 33: Lessons Learned in Teaching Mathematics with Adaptive Tutoring Software

33

Lessons learned 5

We can understand and TRACE students at the affective level, understanding their emotions.

How? from very recent behaviors, and with the help of physiological sensors.

Page 34: Lessons Learned in Teaching Mathematics with Adaptive Tutoring Software

34

Lessons learned 6

Affective Characters that talk about the importance of Effort and Perseverance improve affect

towards math for all, particularly for girls and low achieving students.

However, they don’t impact learning.

Page 35: Lessons Learned in Teaching Mathematics with Adaptive Tutoring Software

35

Animated Pedagogical Agents

“Cognitive”Pedagogical

Agents

Cognitive Outcomes(Retention, Transfer)

Affective Outcomes(Motivation, Attitudes, Emotions)

AffectivePedagogical

Agents?

Page 36: Lessons Learned in Teaching Mathematics with Adaptive Tutoring Software

36

Human-Like Affective Learning CompanionsAffective Experts, cognitive peers

Train the idea of “Malleability of Intelligence”Dweck, C.S., (1999) Self-Theories: Their role in motivation, personality and development

Page 37: Lessons Learned in Teaching Mathematics with Adaptive Tutoring Software

37

Human-Like Affective Learning CompanionsAffective Experts, cognitive peers

Quick-guess incorrect

Correct No EffortPraise Effort and Time on Hints

Low Effort High Effort

Incorrect “We kind of rushed to answer that one. Shall we ask the computer for help? I am sure we will get

it if we take the time to solve the problem.”

“These are the hard questions that I like. There is an opportunity to learn. Let’s click on the help button.”

Correct “That was good, however, I prefer harder questions so that we learn from the help that the

computer gives, even if we get them wrong.”

“Hey, congratulations! Your effort paid off, you got it right!”

Page 38: Lessons Learned in Teaching Mathematics with Adaptive Tutoring Software

38

Summary of ResultsAnalysis of Covariance

Affective Learning Companions are good for all

Page 39: Lessons Learned in Teaching Mathematics with Adaptive Tutoring Software

39Frustrated Pretest Frustrated Within

TutorFrustrated Posttest

1.5

2

2.5

3

3.5

4

4.5

5

No Learning CompanionLearning Companion

How

FRUS

TRAT

ED d

o yo

u fe

el w

hen

solv

ing

mat

h pr

oble

ms?

Impact of Affective LCs for allLess frustration reported within the tutor with Jane. **F(213,2)=6.1,p=.003

More Frustrated

Less frustrated

NeutralFrustration

Level

Page 40: Lessons Learned in Teaching Mathematics with Adaptive Tutoring Software

40

Impact of Affective LCs for all

Less boredom for math at posttest time in LC condition.For N~95 students, comparing LCs vs. no-LCs

Interested Pretest Interested Within Tutor Interested Posttest0.6

1.1

1.6

2.1

2.6

3.1

3.6

4.1

No Learning Companion Learning Companion

How

INTE

REST

ED a

re y

ou w

hen

solv

ing

mat

h pr

oble

ms?

+F(94,1)=3.4,p=.07

More Interested

More Bored

NeutralInterestLevel

Page 41: Lessons Learned in Teaching Mathematics with Adaptive Tutoring Software

41

Impact of Affective LCs for LOW ACHIEVING

More CONFIDENCE for math at posttest time FOR LOW ACHIEVING.For N~95 students, comparing LCs vs. no-LCs

Page 42: Lessons Learned in Teaching Mathematics with Adaptive Tutoring Software

42

Lessons learned 6

Affective Characters that talk about the importance of Effort and Perseverance improve affect towards math for all and for low achieving

students.

However, they don’t impact learning.

Page 43: Lessons Learned in Teaching Mathematics with Adaptive Tutoring Software

43

Lessons learned 7

There are important gender differences that suggest girls have more productive use of

Wayang

Keep in mind that you might be designing for a subset of the population

Page 44: Lessons Learned in Teaching Mathematics with Adaptive Tutoring Software

44Affective Learning Companions are good for allBenefit for all, but effect is stronger when considering Girls alone

Summary of Results charactersAnalysis of Covariance

Page 45: Lessons Learned in Teaching Mathematics with Adaptive Tutoring Software

45See dotted line for increased frustration without companions

Frustration Level

Impact of Affective LCs on GIRLS

More Frustrated

Less frustrated

Reduced FrustrationWith “Jane”

NeutralFrustration

Level

Page 46: Lessons Learned in Teaching Mathematics with Adaptive Tutoring Software

46

Frustrated Pre-Tutor

Frustrated Within Tutor

Frustrated Post-Tutor

1

3.5

6

Males without Learning CompanionMales with JakeMales with Jane

Mea

n Fr

ustra

tion

and

Stde

v (6

=ve

ry fr

ustra

ted)

Frustration Level

Impact of Affective LCs on BOYS

More Frustrated

Less frustrated

NeutralFrustration

Level

Page 47: Lessons Learned in Teaching Mathematics with Adaptive Tutoring Software

47Affective Learning Companions are good for all

Overall benefit, but the effect is stronger for Girls alone.Girls are the ONLY ones benefitting, and boys are clearly not.

Summary of Results about charactersAnalysis of Covariance

Page 48: Lessons Learned in Teaching Mathematics with Adaptive Tutoring Software

48

Perceptions of Wayang Outpost

48

Liked it? Did you Learn? Was Wayang concerned about your learning? Was it helpful?

Means and S.E. for overall Perception of Wayang Outpost

Positive perception

Negativeperception

Girls report a betterlearning experience

With ALCs.

Boys report a better

experienceWithout ALCs.

Males Females Males Females

No LearningCompanion

Learning Companion

NeutralPerception

(neither positive nor negative)

Page 49: Lessons Learned in Teaching Mathematics with Adaptive Tutoring Software

49

Gender differences in Accepting/Rejecting help

Help Offered

Help AcceptedHelp Rejected

Page 50: Lessons Learned in Teaching Mathematics with Adaptive Tutoring Software

50

Gender differences in attitudes and behaviors in Wayang

Page 51: Lessons Learned in Teaching Mathematics with Adaptive Tutoring Software

51

Lessons learned 7

There are important gender differences that suggest girls have more productive use of

Wayang

Keep in mind that you might be designing for a subset of the population

Page 52: Lessons Learned in Teaching Mathematics with Adaptive Tutoring Software

52

Lessons learned1) Adaptively Maintaining students within a “zone of proximal

development” improves learning.2) Training Math Fluency enhances mathematics learning in

combination with adaptive tutoring.3) Showing students their historical improvement at math

problem solving improves engagement in subsequent problems and increases learning.

4) Students are really bored about mathematics. Also, there are important group differences in emotions, before tutoring

5) We can understand and TRACE students at the affective level, understanding their emotions.

6) Affective Characters improve affect. However, they don’t impact learning.

7) Keep in mind that you might be designing for a subset of the population, such as girls and low achieving students.

Page 53: Lessons Learned in Teaching Mathematics with Adaptive Tutoring Software

53

Improving learning by looking at cognition alone: • Reduce working memory load• Facilitate transfer by keeping similar problems together

Keep concepts in working memory

Improving engagement (and affect?) is related to:Pacing and fatigueSupporting meta-cognition, goal setting, reflection.Looking at individual groups of students who need it most

Improving affect, emotions and long term attitudes:Being Positive, encouraging feedbackReflecting about myths and training attributions for failure

Concluding… My conjecturesHow to improve learning, affect, engagement

Page 54: Lessons Learned in Teaching Mathematics with Adaptive Tutoring Software

Department of Computer Science

Ivon ArroyoUniversity of Massachusetts Amherst

Lessons Learned in Teaching Mathematics with Adaptive

Tutoring Software

To Erica

Page 55: Lessons Learned in Teaching Mathematics with Adaptive Tutoring Software

55

Page 56: Lessons Learned in Teaching Mathematics with Adaptive Tutoring Software

56

Lessons learned

If being “gently adaptive” is better…

Shall we be adaptive over the whole pool of problems, or over smaller chunks (topics,skills)?

Page 57: Lessons Learned in Teaching Mathematics with Adaptive Tutoring Software

57

Lesson learned 3

Chunking problems (grouping problems of similar skills together) improves learning.

Page 58: Lessons Learned in Teaching Mathematics with Adaptive Tutoring Software

5858

Results on Standard Math Test (Easy Items)

ANCOVA for Easy Posttest with Easy Pretest as covariate; MFR,Wayang fixed factors:Wayang F(222,1)=10.6, p=.001; WayangxMFR F(222,1)=5.1, p=.025

Post-hoc Contrasts: Wayang > no-Wayang? Yes. Wayang-MFR > Wayang-NoMFR? No.

True Means and SD for EASY items of the standardized pretest and posttest

Page 59: Lessons Learned in Teaching Mathematics with Adaptive Tutoring Software

59

Lesson learned 1.a

How fast should we increase/decrease problem difficulty?

Page 60: Lessons Learned in Teaching Mathematics with Adaptive Tutoring Software

60

Varying the challenge in adaptive problem selection

XX

Challenge WayangGentle

Wayang

Page 61: Lessons Learned in Teaching Mathematics with Adaptive Tutoring Software

61

“Gentle Adaptive” Wayang also offers Help

Unpublished

Page 62: Lessons Learned in Teaching Mathematics with Adaptive Tutoring Software

62

Scaffolding and Help Offering

Pretest % correct (11 questions) Posttest %correct (11 questions)15%

25%

35%

45%

55%

65%

75%

Challenge Problem SelectorGentle Problem Selector

Pret

est

and

Post

test

Sco

res

Unpublished

Page 63: Lessons Learned in Teaching Mathematics with Adaptive Tutoring Software

63

Chunking similar problems together or Not

Pretest % correct (11 questions) Posttest %correct (11 questions)0%

10%

20%

30%

40%

50%

60%

70%

No chunking ChunkingPr

etes

t an

d Po

stte

st S

core

s

Chunking facilitates transfer?