Title: Knowledge Tracing and Prediction of Future Trainee Performance Tiffany S. Jastrzembski, Kevin A. Glu
1Knowledge Tracing and Prediction of Future
Trainee Performance Tiffany S. Jastrzembski,
Kevin A. Gluck, Glenn GunzelmannAir Force
Research Laboratory
2Research Goals
- Basic Science Aims
- Implement mathematical models based on cognitive
science empirical research and theory - Integrate mechanisms to handle spaced versus
massed learning - Applied Science Aims
- Predict warfighter performance when regimen of
training is known - When will the warfighter achieve
mission-readiness? - Prescribe warfighter training schedule to
optimize performance - How much training is required when deployment
date is known? - Is training timetable reasonable?
3Approach
- Construct a trainee model based on an existing
knowledge tracing equation - Capitalize on existing strengths of model
- Account for dynamics of spaced versus massed
practice - Extend model capabilities to have predictive
power - Calibrate trainee model parameters from
performance history - Extrapolate knowledge state transformation for
predictive and prescriptive purposes
4Mathematical Basis of Current Research
- General Performance Equation (Anderson Schunn,
2000) - Simple knowledge units are acquired according to
simple principles - Performance success relies on frequency, recency,
and amount of practice - Performance
- S scaling parameter
- N amount of practice
- c learning rate
- T time since learning
- d decay parameter
5General Performance Equation Model Fits
Skill Retention for Typewriting Aptitude (Bean,
1912)
Knowledge Acquisition for Logic-Based Facts
(Anderson Fincham, 1994)
6The Spacing Effect
- Empirical Findings
- Memory benefits accrue with increased durations
in practice - Rate of forgetting decreases as time passes
- Distributed practice consistently superior to
massed practice - Ex Learning is more durable when practices are
spaced over a month compared to practice crammed
into a week - General Performance Equation Results
- Discrete increments in learning are added at each
training point - Unable to account for differences in massed
versus spaced practice - Results in better performance for massed practice
7General Performance Equation Model Fit to
Distributed Data- Glenberg, 1976
- Data spaced at practice intervals of every 2
and 8 trials - Model fits all post-hoc
- Correlations quite poor
8Predictive Performance Equation
- Performance
- S original scaling parameter x improvement in
training history - N amount of practice
- c learning rate
- T time since learning
- a activation-based decay parameter
- Calculated from training history
- Based on original decay rate (d) and activation
level at latest data point -
-
- where m latest activation
9Predictive Performance Equation Model Fit to
Distributed Data- Glenberg, 1976
R .96 RMSD 1.47
- Captures recency, frequency, and spacing effect
- Correlation very good
10Challenges to Predictive Validity of Mathematical
Models (Part 1)
- Data Resolution
- Mathematical regularities differ at individual
operator, team, and aggregate levels of analysis - Aggregate level data eliminates noise and smooths
performance curves - True learning trends may become distorted or
averaged away - Learning trajectories of individuals may differ
vastly from each other and the average
11Unmanned Air Vehicle Synthetic Task Cognitive
Engineering Research Institute
- Three operators must coordinate over headsets to
complete missions and take reconnaissance photos
of ground targets - Performance is based on the weighted sum of
penalty scores across team members
12Predictive Performance Equation Model Fits to
Aggregate Data
- Training Scenario
- Session 1 Baseline
- Five, 40-minute missions
- Optimize model parameters for learning and decay
- Session 2 Predictive test
- Return 10-14 weeks later
- Three, 40-minute missions
- Extrapolate mathematical regularities for
prediction
R .95 RMSD 10.3
13Predictive Performance Equation Model Fits to
Individual Team Data
R .91 RMSD 15.6
14Predictive Performance Equation Model Fits to
Individual Operator Data
R .68 RMSD 62.6
15Comparison of Human Performance Curves Across
Levels of Data Resolution
16Challenges to Predictive Validity of Mathematical
Models (Part 2)
- Insufficient Training History
- Model will function at a disadvantage if the
baseline is inadequate - Model will extrapolate mathematical regularities
for future prediction that may be less certain or
valid - Model analyses will help reveal minimal training
history required for effective use
17Predictive Performance Equation Model Fits as a
Function of Training History
- Detailed training history is necessary to
baseline model parameters - Instructors need to collect sufficient training
logs - Greater amounts of training history may enhance
prediction at finer grains of resolution (e.g.
individual level)
18Notional Prediction
- Scenario How long will it take to achieve 95
proficiency under the current regimen of practice?
Desired Proficiency
- Baseline Training Days 1 and 84
- Parameterize equation from baseline performance
- At 3 missions/day, 5 days/week, an additional
1,120 40-minute missions are required to achieve
proficiency
19Potential Predictive Utility in the Military
Domain
- Mission-Readiness
- Reveals warfighter mission-readiness under
specific distributions of practice - Accounts for practice spaced apart at any length
of time - Equipped with ability to account for extended
breaks in training - How much additional training will be needed to
reach proficiency after taking a 2-month break? - Probabilistically determines if a maneuver or
mission will succeed at some future point in time
20Notional Prescription
- Deployment Scenario How much training must a
warfighter receive to be mission-ready (95
proficient) by the end of four weeks?
Desired Proficiency
- Baseline Training Days 1 and 84
- Parameterize equation from baseline performance
- Prediction dictates 24 practice missions/day, 5
days/week
Deployment
21Notional Prescription
- Deployment Scenario How much training must a
warfighter receive to be mission-ready (95
proficient) by the end of four months?
- Baseline Training Days 1 and 84
- Parameterize equation from baseline performance
- Prediction dictates 5 missions/day, 5 days/week
- 40 fewer practice sessions are needed in this
spaced scenario
Desired Proficiency
Deployment
22Potential Prescriptive Utility in the Military
Domain
- Optimal Training Regimen
- Compares training schedules to maximize
performance - Assesses how effective each training repetition
will be - Optimizes spacing of practice to result in larger
learning gains - Logistic Utilities
- Determines if training expectations are realistic
- Predicts whether mission-readiness can be
attained within a specified period of time - Provides trainers with reasonable
mission-readiness timetables
23Conclusions
- Model may
- Serve as a useful tool for warfighters and
instructors - Help determine when a warfighter has achieved
proficiency - Help streamline practice schedules to optimize
learning - Help determine realistic training timetables for
warfighters to become mission-ready
24Thank You!Questions?