Probabilistic and decision-theoretic user modeling in the context of software customization Depth oral presentation Department of Computer Science University of Toronto Bowen Hui [email protected]
Probabilistic and decision-theoretic
user modeling in the context of
software customization
Depth oral presentationDepartment of Computer Science
University of Toronto
Bowen [email protected]
june 17th 2004 depth oral presentation 2
Need for software customization
current state of practice: one-size-fits-all
result:
bloated software, cluttered interfaces, closet-ware, user dissatisfaction
most affected users
novices
children
elderly people
people with cognitive, sensory, motor impairments
june 17th 2004 depth oral presentation 3
Software customization (SC)
customize interface or functionality
requirements of diff user groups [HLM03]
customize software based on user goals
ranked alternatives with preferences + skills
generated customization at design time
customize at run time
adaptable – user control
adaptive – system control
continuum: adaptableadaptive
automationsuggestion
june 17th 2004 depth oral presentation 4
user profile
goals
preferences
skills
user traits
uncertainty
cost modeling
expected value of information (EVOI)
sequential reasoning
unobservability
probabilistic and decision-theoretic modeling
Aspects of SC
june 17th 2004 depth oral presentation 5
General literature landscape
softwarecustomization
+user
modeling
june 17th 2004 depth oral presentation 6
General literature landscape
softwarecustomization
+user
modeling
planrecognition
utility theory
probabilitytheory
june 17th 2004 depth oral presentation 7
General literature landscape
softwarecustomization
+user
modeling
plan recognitionw/probabilities
+ utilities
planrecognition
june 17th 2004 depth oral presentation 8
General literature landscape
softwarecustomization
+user
modeling
plan recognitionw/probabilities
+ utilities
planrecognition
BNs,DBNs
june 17th 2004 depth oral presentation 9
General literature landscape
softwarecustomization
+user
modeling
plan recognitionw/probabilities
+ utilities
planrecognition
BNs,DBNs
june 17th 2004 depth oral presentation 10
General literature landscape
softwarecustomization
+user
modeling
plan recognitionw/probabilities
+ utilities
planrecognition
BNs,DBNs
POMDPs
MDPs
inverse RL
preferenceelicitation
june 17th 2004 depth oral presentation 11
General literature landscape
softwarecustomization
+user
modeling
plan recognitionw/probabilities
+ utilities
planrecognition
BNs,DBNs
POMDPs
MDPs
inverse RL
preferenceelicitation
june 17th 2004 depth oral presentation 12
goals, preferences, skills FC01,BJB98,CV01,MV00,ZC03,HBH+98,Hor99
user traits CV01,MV00,ZC03
uncertainty FC01,BJB98,GW04,LFL98,CV01,MV00,ZC03,HBH+98,Hor99
unobservability HBH+98,Hor99
cost modeling GW’04,LFL98
EVOI Hor99
sequential reasoning ?
System comparison
june 17th 2004 depth oral presentation 13
goals, preferences, skills FC01,BJB98,CV01,MV00,ZC03, Lumière: HBH+98,Hor99
user traits CV01,MV00,ZC03
uncertainty FC01,BJB98,GW04,LFL98,CV01,MV00,ZC03,HBH+98,Hor99
unobservability HBH+98,Hor99
cost modeling GW’04,LFL98
EVOI Hor99
sequential reasoning ?
System comparison
utility
automatedassistance
cost ofassistance
needs
explicitquery
user’sactions
documents,data
structures
goals
contexttask
history
skillsprofile
user’sbackground
assistancehistory
june 17th 2004 depth oral presentation 14
goals, preferences, skills FC01,BJB98,CV01,MV00,ZC03,HBH+98,Hor99
user traits CV01,MV00,ZC03
uncertainty FC01,BJB98,GW04,LFL98,CV01,MV00,ZC03,HBH+98,Hor99
unobservability HBH+98,Hor99
cost modeling GW’04,LFL98
EVOI Hor99
sequential reasoning ?
System comparison
june 17th 2004 depth oral presentation 15
goals, preferences, skills FC01,BJB98,CV01,MV00,ZC03,HBH+98,Hor99
user traits CV01,MV00,OCC in Prime Climb: ZC03
uncertainty FC01,BJB98,GW04,LFL98,CV01,MV00,ZC03,HBH+98,Hor99
unobservability HBH+98,Hor99
cost modeling GW’04,LFL98
EVOI Hor99
sequential reasoning ?
System comparison
mathemotionoutcome
emo.self
emo.agent
interactionpatterns
studentaction
goalssatisfied
personality G
uses EO as preference
can infer M,P,G
can compare accuracy M,P to pre/post-tests
june 17th 2004 depth oral presentation 16
goals, preferences, skills FC01,BJB98,CV01,MV00,ZC03,HBH+98,Hor99
user traits CV01,MV00,ZC03
uncertainty FC01,BJB98,LFL98,CV01,MV00,ZC03,HBH+98,Hor99
unobservability HBH+98,Hor99
cost modeling GW’04,LFL98
EVOI Hor99
sequential reasoning ?
System comparison
june 17th 2004 depth oral presentation 17
goals, preferences, skills FC01,BJB98,CV01,MV00,ZC03,HBH+98,Hor99
user traits CV01,MV00,ZC03
uncertainty FC01,BJB98,LFL98,CV01,MV00,ZC03,HBH+98,Hor99
unobservability HBH+98,Hor99
cost modeling GW’04,LFL98
EVOI Hor99
sequential reasoning ?
System comparison
Bayesian: - model causal influences- belief distribution- update beliefs in principled way
june 17th 2004 depth oral presentation 18
goals, preferences, skills FC01,BJB98,CV01,MV00,ZC03,HBH+98,Hor99
user traits CV01,MV00,ZC03
uncertainty FC01,BJB98,GW04,LFL98,CV01,MV00,ZC03,HBH+98,Hor99
unobservability HBH+98,Hor99
cost modeling GW’04,LFL98
EVOI Hor99
sequential reasoning ?
System comparison
june 17th 2004 depth oral presentation 19
goals, preferences, skills FC01,BJB98,CV01,MV00,ZC03,HBH+98,Hor99
user traits CV01,MV00,ZC03
uncertainty FC01,BJB98,GW04,LFL98,CV01,MV00,ZC03,HBH+98,Hor99
unobservability Lumière: HBH+98,Hor99
cost modeling GW’04,LFL98
EVOI Hor99
sequential reasoning ?
System comparison
0
10
20
30
40
50
60
70
80
90
100
G1 G4 G7 G10 G13 G16 G19 G22
35%N:infers:
- Pr(need help now)- Pr(goal)
action: - help if Pr(N) > threshold
june 17th 2004 depth oral presentation 20
goals, preferences, skills FC01,BJB98,CV01,MV00,ZC03,HBH+98,Hor99
user traits CV01,MV00,ZC03
uncertainty FC01,BJB98,GW04,LFL98,CV01,MV00,ZC03,HBH+98,Hor99
unobservability HBH+98,Hor99
cost modeling FC01,GW04,LFL98,MV00,Hor99
sequential reasoning ?
System comparison
june 17th 2004 depth oral presentation 21
System comparison
goals, preferences, skills FC01,BJB98,CV01,MV00,ZC03,HBH+98,Hor99
user traits CV01,MV00,ZC03
uncertainty FC01,BJB98,GW04,LFL98,CV01,MV00,ZC03,HBH+98,Hor99
unobservability HBH+98,Hor99
cost modeling FC01, Supple: GW04,LFL98,MV00,Hor99
sequential reasoning ?cost( interface ) = navigation cost + manipulation cost, w.r.t. history
june 17th 2004 depth oral presentation 22
goals, preferences, skills FC01,BJB98,CV01,MV00,ZC03,HBH+98,Hor99
user traits CV01,MV00,ZC03
uncertainty FC01,BJB98,GW04,LFL98,CV01,MV00,ZC03,HBH+98,Hor99
unobservability HBH+98,Hor99
cost modeling FC01,GW04,LFL98,MV00, LookOut: Hor99
sequential reasoning ?
System comparison
outcomes:u(A,G) [best], u(A,G) [worst]u(A,G), u(A,G)
EU(A) = Pr(G)u(A,G) + Pr(G)u(A,G), EU(A) similarly
june 17th 2004 depth oral presentation 23
goals, preferences, skills FC01,BJB98,CV01,MV00,ZC03,HBH+98,Hor99
user traits CV01,MV00,ZC03
uncertainty FC01,BJB98,GW04,LFL98,CV01,MV00,ZC03,HBH+98,Hor99
unobservability HBH+98,Hor99
cost modeling FC01,GW04,LFL98,MV00,Hor99
sequential reasoning ?
System comparison
june 17th 2004 depth oral presentation 24
What’s missing
modeling
estimate user traits
expected value of information (except Hor99)
user’s utility function (except Hor99)
non-myopic policies
evaluation
“convergence” of performance
online performance
usability
learning user’s utility function
june 17th 2004 depth oral presentation 25
Underlying models
goals, preferences, skills,
traits
uncertainty
cost modeling
sequential reasoning
unobservability
softwarecustomization
+user
modeling
plan recognitionw/probabilities
+ utilities
planrecognition
BNs,DBNs
POMDPs
MDPs
inverse RL
preferenceelicitation
june 17th 2004 depth oral presentation 26
Underlying models
goals, preferences, skills,
traits
uncertainty
cost modeling
sequential reasoning
unobservability
softwarecustomization
+user
modeling
plan recognitionw/probabilities
+ utilities
planrecognition
BNs,DBNs
POMDPs
MDPs
june 17th 2004 depth oral presentation 27
Underlying models
goals, preferences, skills,
traits
uncertainty
cost modeling
sequential reasoning
unobservability
softwarecustomization
+user
modeling
plan recognitionw/probabilities
+ utilities
planrecognition
BNs,DBNs
POMDPs
MDPs
june 17th 2004 depth oral presentation 28
Underlying models
goals, preferences, skills,
traits
uncertainty
cost modeling
sequential reasoning
unobservability
softwarecustomization
+user
modeling
plan recognitionw/probabilities
+ utilities
planrecognition
BNs,DBNs
POMDPs
MDPs
june 17th 2004 depth oral presentation 29
Underlying models
goals, preferences, skills,
traits
uncertainty
cost modeling
sequential reasoning
unobservability
softwarecustomization
+user
modeling
plan recognitionw/probabilities
+ utilities
planrecognition
BNs,DBNs
POMDPs
MDPs
june 17th 2004 depth oral presentation 30
Underlying models
goals, preferences, skills,
traits
uncertainty
cost modeling
sequential reasoning
unobservability
softwarecustomization
+user
modeling
plan recognitionw/probabilities
+ utilities
planrecognition
BNs,DBNs
POMDPs
MDPs
june 17th 2004 31
Sequential decision making under
uncertainty
principled way of modeling SC
user interacting with software
user has goals
user has reward/utility function
user “states” are unobservable
opportunities for agent to assist user
every action has consequences
model agent in environment who acts in expectation of the user’s utility function
partially observable Markov decision process (POMDP)
softwareenvironment
agent
interface
utilities
user
june 17th 2004 depth oral presentation 32
Research directions
typing assistant model represent POMDP model as DBN
refine model simulations
collect user data to learn parameters T,O
evaluation “convergence”
online performance
usability
learning user’s utility function formulate as POMDP
constrain belief distribution (PE)
incrementally learn utility function (IRL)
adapt POMDP approximation algorithms