Download Principle of Maximum Expected Utility (von Neumann

Document related concepts

Channel coordination wikipedia , lookup

Trusted Computing wikipedia , lookup

Bayesian inference in marketing wikipedia , lookup

Transcript
Uncertainty, Utility, and
Understanding
Eric Horvitz
Microsoft Research
June 2000
A Long-Term Dream

1958: “Artificial Intelligence” (AI)



1970s: AI methods hit the real world


Focus: Thinking as symbol processing
AI as distinct from OR, Cybernetics, Management
Science
Limitations of logic; uncertainty and intelligent
behavior, diagnosis
1990-present: Beliefs and actions under
uncertainty


Fabric of probability and utility theory
Graphical representations and inference methods
Representation: Graphical Models



Bayesian networks, influence
diagrams
Encode independence, crisp
semantics
Fundamentally modular
representations
Delivering Value in the Real World
Medicine, machine diagnosis & repair
PULMONARY EMBOLUS
HYPOVOLEMIA LV FAILURE ANAPHYLAXIS
ANESTHESIA
INSUFFICIENT
SHUNT
PAP
LVED
VOLUME
CVP
HISTORY
LV FAILURE
STROKE
VOLUME
TPR
CARDIAC
OUTPUT
CATECHOLAMINE
ERROR
LOW OUTPUT
VENT MACHINE
VENT
TUBE
VENT LUNG
PCWP
BLOOD
PRESSURE
KINKED
TUBE DISCONNECTION
VENT ALV
SAO2
HEART
RATE
INTUBATION
PA SAT
ERROR
CAUTER
HR BP HR EKG HR SAT
MV SETTING
FIO2 PRESSURE
ARTERIAL
CO2
MINUTE
VENTILATION
EXPIRED
CO2
Beinlich, et al., 1988
Action under Uncertainty


Axioms of Utility (von Neumann & Morgenstern, 1947)
Principle of Maximum Expected Utility (MEU)
Take actions that maximize expected utility
Representations for Action

Influence diagrams (Howard & Matheson)
Action
Utility
World
State
E1
E2
E3
En
Delivering Value in the Real World
Beyond Medicine and Machines:
User Modeling

Some motivations:
- Adapting Bayesian diagnostic systems
to different classes of user
- Pilot’s Associate, SCI program
Bayesian models for capturing a
person’s intentions, goals, needs
 Ubiquity of uncertainty!

Vista Project
Causal Model of Propulsion Systems
He P1
He P2
Ox Temp
Ox Tank P
Ox
H
e
Fu
Fu Temp
Fu Tank P
s
Ox Inlet P
Fu Inlet P
X-over Fu P
N
N Accum P
Combust P
X-over Ox P
Bayesian Network for Shuttle Propulsion
He P1 Meas
He Tank Leak
He P1 Sens Failure
He Pressure
He P2 Meas
He P2 Sens Failure
He P1 Trend
He Ox Valve
He Fu Valve
Hx Fu Temp Prob
He P2 Trend
Hx Ox Temp Prob
Ox Tank Leak
Fu Tank Leak
Fuel Temp Prob
Ox Temp Prob
Ox Tank P
Fu Tank P
Ox Contam
Ox Inlet P
Fu Contam
Fu Inlet P
X-over Ox P
X-over Fu P
Combust P
N Tank P2
N Tank Pres
Engine Failure
N Tank P1
N Valve Fail
Comb Pres Meas
N Accum P
N Tank Leak
Comb Sens Failure
Comb P Trend
N P2 Sens Fail
N P1 Sens Fail
Challenge:
Display for Time-Critical Decisions

What is the most important information to
display to decision makers?
Action,t
Utility
System
Faults
E1
E2
E3
En
Considering the Fundamental
Decision Problem
Action, t
Utility
System
Faults
E1
E2
E3
En
Taking System’s Perspective on
Display Actions
 Decision maker actions as uncertain variable
?
Action,t
Utility
System
Faults
E1
E2
E3
En
Taking System’s Perspective on
Display Actions
 Decision maker actions as uncertain variable
?
Action,t
Utility
System
Faults
E1
E2
E3
En
Taking System’s Perspective on
Display Actions
 Decision maker actions as uncertain variable
User’s
Action
Displayed
Information
User’s
Delay
Action,t
Utility
System
Faults
E1
E2
E3
En
Expected Value of Displayed Information
How will displaying additional information
enhance an user’s decision?
Gold standard
User model
A*
t(e)
Construct User Models via
Selective Pruning of Expert Model
H2
H1
E1
E2
E3
E4
H2
H1
E1
E2
E3
E4
Construct User Models via
Selective Pruning of Expert Model
H2
H1
E1
E2
E3
E4
H2
H1
E2
E3
E4
Power of normative representations
to capture descriptive challenges
Information Highlighting Decisions

Output: Highlighting data in situ
Oxygen
Fuel Pres
Chamb Pres
He Pres
Delta v
15.6
10.5
14.2
11.8
5.4
4.8
17.7
33.3
14.7
63.3
Oxygen
Fuel Pres
Chamb Pres
He Pres
Delta v
10.2
12.8
0.0
15.8
32.3
10.6
12.5
0.0
15.7
63.3
Information Highlighting Decisions

Output: Highlighting data in situ
Oxygen
Fuel Pres
15.6
10.5
14.2
11.8
5.4
4.8
He Pres
17.7
14.7
Delta v
33.3
63.3
Oxygen
Fuel Pres
Chamb Pres
He Pres
Delta v
10.2
12.8
0.0
15.8
32.3
10.6
12.5
0.0
15.7
63.3
Chamb Pres
Information Highlighting Decisions

Output: Highlighting data in situ
Oxygen
Fuel Pres
15.6
10.5
14.2
11.8
5.4
4.8
He Pres
17.7
14.7
Delta v
33.3
63.3
Oxygen
Fuel Pres
10.2
12.8
10.6
12.5
0.0
0.0
15.8
15.7
32.3
63.3
Chamb Pres
Chamb Pres
He Pres
Delta v
Uncertainty at Microsoft
Uncertainty at Microsoft
Intelligent
Interface
Intelligent
Systems
Diagnosis and Repair
On
the
Web…
www.microsoft.com
Goals, Understanding, and Uncertainty
Infer likelihoods of user’s goals, attention,
understanding and take ideal actions
• User query
• User activity
• User location
• User profile
• Data structures
• Vision, speech, sound
Pr(Goals, Understanding)
*
Value-Focused Action
Foci
 Sensing
events & content
 Building / learning models of goals,
knowledge, attention
 Principles of decision making
 Psychological studies
 Prototyping apps and systems
Big Picture
Learning
Models
Events
Events
Synthesis
Inference
about
User, World
Control
Computation
of Ideal UI Action
New
Perceptual
Actions?
Ideal
Actions
Lumière Project
User’s Goals
User’s Profile
User’s Needs
User Activity
Actions + Words  Goals
Studies with Human Subjects
Studies with Human Subjects

“Wizard of OZ” experiments at
MS Usability Labs
User Actions
Typed Advice
Expert Advisor
Inexperienced user
Assistance Informed by a Keyhole View
Video
Activities with Relevance to
Informational Needs
Several categories of evidence

Search: e.g., exploring of multiple menus

Introspection: e.g., sudden pause, slowing of
command stream

Focus of attention: e.g, selected objects

Undesired effects: e.g., command/undo, dialogue
opened and cancelled

Inefficient command sequences

Syntactic / semantic content of file

Goal-specific sequences of actions
Building Bayesian User Models
User Needs
Assistance
User Distracted
Pause after
Activity
Building Bayesian User Models
Recent
Menu Surfing
User Needs
Assistance
User Distracted
Pause after
Activity
Building Bayesian User Models
User Expertise
Difficulty of
Current Task
User Needs
Assistance
User Distracted
Recent
Menu Surfing
Pause after
Activity
Portion of Lumière Bayesian Net
User background
Primary goal
Chart wizard
Repeated chart
create/delete
Consolidation
Hierarchical
presentation
Pivot wizard
Group mode
3D cell reference
Database defined
Leading spaces
External reference
Repeated chart
change
Use query
Multicell selection
Adjacent conceptual
granularity
Repeated
print / hide Rows
Sensing Context and Content: Eve

Toward a “peripheral nervous system” for
sensing user activity


SDK with event abstraction language
Compiler for defining filters for user activity
Time
Abstraction of Events
Event
Source 1
Event
Source
2
Time
Event
Source n
Eve
Event-Specification
Language
Atomic Events
Modeled Events
Abstraction of Events
Atomic Events
Time
{Menu x visited (t)
Modeled Events
Abstraction of Events
Atomic Events
Modeled Events
Time
{Menu x visited (t), Menu y visited (t’)
Abstraction of Events
Atomic Events
Modeled Events
Time
{Menu x visited (t), Menu y visited (t’), … Menu z visited (t”) }
User menu surfing (t”)
Multiple Temporal Granularities
Atomic event
stream
Hours
Focused project
Minutes
Outlook centric
Seconds
Menu surfing
Overall Lumiere Architecture
Events
• Actions
Event Synthesis
Time
Bayesian
Inference
• Query
Control System
Inference from Actions & Words
• Sensed actions
• User’s query
Considering the Need for Assistance
• Probability user
desires assistance
Considering the Need for Assistance
• Probability user
desires assistance
Several Challenges





Building and learning user models
Sensing activity from systems and
applications
Reasoning over time
Models, persistent parameters of a
user’s competency
Decision making under uncertainty
Probabilistic Inference about a User’s
Time-Dependent Goals
Profil
e
Goalt-n
Ei,t-n
Profil
e
Profil
e
Ej,t-n
Goalto
Goalt-1
Ei,t-1
Ej,t-1
Ei,to
Time
Ej,to
Reasoning about Competency
Task
history
Context
Review
of Help Info
User
background
User's
goals
User's
actions
User’s
Competencies
Cost of
assistance
User's
acute needs
Utility
Explicit
query
Automated
Assistance
Representing and Updating a
“Competency Terrain”
User’s
Competencies
User’s Skills
Representing and Updating a
“Competency Terrain”
User’s
Competencies
“Mail Merge”
“Keyboard shortcuts”
User’s Skills
Competency, Memory, and
Volatility of Knowledge
User’s
Competencies
“Mail Merge”
“Keyboard shortcuts”
User’s Skills
Competency, Memory, and
Volatility of Knowledge
User’s
Competencies
“Mail Merge”
“Keyboard shortcuts”
User’s Skills
Several Current Directions




Mixed-initiative interaction
Principles of conversation
Attentional systems and interfaces
Learning models from data
Towards Mixed-Initiative User Interfaces





Courteous computing!
Assume user and system collaborate,
each making contributions to goal
Reason about need, timing of services
Preference for doing less correctly
Critical decision:



Do nothing.
Ask?
Just do it?
Preferences and Automated Action

Expected utility as fundamental in
decisions about services
Desired
Not desired
Act
u(A,D)
u(A,D)
User Desire
No act
D: User desires action i
D: User does not desire action i
u(A,D)
u(A,D)
Service
A: Computer takes action i
A: No action i
Preferences and Automated Action
eu(A) = Sj u(Ai,Dj) p(Dj|E)
1.0
u(A,D)
u(A,D)
u(A,D)
u(A,D)
0.0
p(D|E)
1.0
Preferences and Automated Action
eu(A) = p(D|E) u(A,D) + [1 - p(D|E)] u(A,D)
eu(A) = p(D|E) u(A,D) + [1 - p(D|E)] u(A,D)
1.0
u(A,D)
u(A,D)
u(A,D)
P*
u(A,D)
0.0
p(D|E)
1.0
Dynamic Influences on Utility
Utility of outcomes as function of context, u(A,D,k)
1.0
u(A,D)
u(A,D)
User rushed
u(A,D)
P*
u(A,D)
0.0
p(D|E)
1.0
Dynamic Influences on Utility
Utility of outcomes as function of context, u(A,D,k)
1.0
u(A,D)
u(A,D)
User rushed
u(A,D)
P*
u(A,D)
Increase in Amount
of Screen Real Estate
0.0
p(D|E)
u(A,D)
1.0
Dynamic Influences on Utility
Utility of outcomes as function of context, u(A,D,k)
1.0
u(A,D)
u(A,D)
User rushed
u(A,D)
P*
u(A,D)
Increase in Amount
of Screen Real Estate
0.0
p(D|E)
u(A,D)
1.0
Dynamic Influences on Utility
Utility of outcomes as function of context, u(A,D,k)
1.0
u(A,D)
u(A,D)
User rushed
u(A,D)
P*
u(A,D)
Increase in Amount
of Screen Real Estate
0.0
p(D|E)
u(A,D)
1.0
Dynamic Influences on Utility
Utility of outcomes as function of context, u(A,D,k)
1.0
u(A,D)
u(A,D)
User rushed
u(A,D)
u(A,D)
P*
u(A,D)
Increase in Amount
of Screen Real Estate
0.0
p(D|E)
u(A,D)
1.0
Considering Another Alternative
Expected value of engaging the user in dialogue
1.0
u(A,D)
u(A,D)
u(A,D)
P*
u(A,D)
0.0
p(D|E)
1.0
LookOut: Messaging & Scheduling
User Actions / Context
• Watch user’s behavior
• Store cases, timing info
• Learn model from data
Real-Time
Probabilistic Inference
Cost--Benefit
Analysis
UI / Service
Joint work with Andy Jacobs
LookOut
Learning Ideal Timing of Services


Timing can be critical
Record length of message and dwell time before
calendar invoked
Perform regression
Observed dwell before action
(sec)

10
8
6
4
2
0
0
500
1000
1500
2000
Length of original message (bytes)
2500
Varying Precision of Service:
Assume Mixed-Initiative
Tradeoff reduced precision for higher accuracy
Assume user will refine partial results
 Automated scope of calendar view

Specific appt.  Day  Week  Month
LookOut
LookOut
LookOut
LookOut
LookOut
Mixed-Initiative Interaction
and Tutoring Systems





What aspects of the user’s
understanding problem can be best
handled by the system?
How—and when—should the system
interject with some assistance?
What is the cost versus benefits of the
intervention? Consider expected utility
Was that intervention useful?
Do less, but with higher precision!
Toward Richer Models of Conversation
User Goals ?
!
?
State of Belief
Cost/Benefit
Analysis
Toward Richer Models of Conversation
User Goals ?
!
State of Belief
Cost/Benefit
Analysis
Conversational Architectures Project
Conversation as joint activity in pursuit of
common ground.

Consider visual and linguistic clues

Decompose task into a progression through
joint-activity sub-projects

Sub-projects completed when mutual
understanding is “good enough for current
purposes”

Expected utility to control information
gathering and progression.
Navigating a Hierarchy with Expected Utility
User’s Goal
Level 0
Goal n
Goal 1
Evidence gathering
Navigation decisions
Subgoal 11
Level 1
Subgoal 1x
Level 2
Subgoal 1x1
Subgoal 1xy
Conversational Actions:
Information Gathering and Navigation
Initial utterance, observations


Continue to
gather
information
User’s Goal
Goal n
Goal 1
Progress to next level
of precision after
confirmation
Level 0
EVI
Open request for information
Level 1

Progress to next level of
precision without
confirmation
Subgoal 11
Subgoal 1x
Subgoal 1x1


Return to previous level
of analysis
Take action in world
EVI
Level 2
Subgoal 1xy
EVI
World action
Receptionist Domain



Domain of situations, interactions,
actions handled by a building
receptionist
Receptionist at Microsoft Research
(Building 9)
Assistance with requests from visiting
researchers, internal people, telephone
calls, etc.
Receptionist Domain

Observational Study





Domain model



9 hours of interactions
Audio, video analysis
Interviews
Knowledge engineering with 3 receptionists
32 mutually exclusive and exhaustive goals
Key visual cues
Linguistic cues
Typical Utterances










“I’m here to see Rick Rashid”
“Uh…Bathroom?”
“I need a shuttle...”
“27!”
“I’m late and lost..”
“Beam me to 25…!”
“Is this research..?”
“This is for Bill Dolan...”
“Is my travel stuff here...?”
“Where can I find Bill Gates?”
Visual Cues

Appearance


Behavior


appears rushed, direction of glance, etc.
Spatial configuration and trajectory


attire, class of badge if visible, etc.
mode of arrival, recent trajectory
Props

equipment, objects being carried, group
Bayesian Receptionist Components

Microsoft Bayesian Network Modeling &
Inference System


Microsoft NLPwin


Bayesian Inference and EVI
Syntactic, Logical, and Lexical Features
Microsoft Agent


Speech Recognition
Text-to-Speech Generation
Joint work with Tim Paek
Bayesian Models and Dialog
User’s Goal
Goal n
Goal 1
Level 0
VOI
Level 1
Subgoal 11
Subgoal 1x
VOI
Level 3
Subgoal 1x1
Subgoal 1xy
VOI
Bayesian Models and Dialog
User’s Goal
Goal n
Goal 1
Level 0
VOI
Level 1
Subgoal 11
Subgoal 1x
VOI
Level 3
Subgoal 1x1
Subgoal 1xy
VOI
Bayesian Models and Dialog
User’s Goal
Goal n
Goal 1
Level 0
VOI
Level 1
Subgoal 11
Subgoal 1x
VOI
Level 3
Subgoal 1x1
Subgoal 1xy
VOI
Bayesian Models and Dialog
User’s Goal
Goal n
Goal 1
Level 0
VOI
Level 1
Subgoal 11
Subgoal 1x
VOI
Level 3
Subgoal 1x1
Subgoal 1xy
VOI
Richer Decision Making about
Grounding a Conversation
Conversational Control
L is considering S’s proposal of a
S is proposing activity a to L
Intention
S is signaling that p for L
L is recognizing that p from S
Signal
S is presenting signal s to L
L is identifying signal s from S
Channel
S is executing behavior b for L
L is attending to behavior b from S
Conversation Control
Repair
Strategy
Intention
Errors
Dialog
Record
Grounding
Status (t-1)
Grounding
Status (t)
Goal (t-1)
Grounding
Status (t)
Action versus
Repair
Goal (t)
Utility
Consecutive
Lows
Consecutive
Multiples
Number of
Turns
Maintenance (t)
Consecutive
Troubleshoot
Maintenance (t)
Maintenance (t-1)
Utility
Overheard
Quit
Confidence
Explain
Confidence
Next
Confidence
Back
Confidence
Intention Module
Nothing
Heard
Listener
Attended
Computer
Activity
Signals
Per Turn
ASR
Reliability
Type of
Microphone
Maintenance Module
Challenge: Apply Similar Approach to
Identify and Address Conceptual Problems
Conceptual Problem
Problem n
Problem 1
Level 0
VOI
Level 1
Subgoal 11
Subgoal 1x
VOI
Level 3
Subgoal 1x1
Subgoal 1xy
VOI
Challenge: Decision-Theoretic Navigation to
Control Progression to Greater Sophistication
Expected-utility decision making
Level 0
Observation
VOI
Progress
Backtrack
Observation
Level 1
VOI
Progress
Level 3
Observation
VOI
Attentional User Interface (AUI)
Project
Harness inferences about a user’s
attentional focus

Models of attention, engagement

Consider multiple sources of
information

Identify attention-sensitive costs and
benefits of services, information
Leveraging Models of a User’s Attention
 Where
is the user?
 What is the user’s attentional status?
 What is the cost of an interruption at the
current time?
 What is the value of the information, cost
of deferring assistance, notifications?
Richer Attentional Models
ONLINE
CALENDAR
USER LOCATION
DATE, TIME
AMBIENT
ACOUSTICAL SIGNAL
DEADLINE
STATUS
USER’S
ATTENTIONAL FOCUS
APPLICATION
IN FOCUS
INSPECTION INTERVAL,
AVAILABILITY
APPLICATION
USAGE PATTERN
DESKTOP
ACTIVITY
Extension to Consider Key
Temporal Dependencies
Calendar
Calendar
Location, to
Date, Time
Location, t1
Date, Time
Acoustical
Signal, to
Deadline
Status, to
Attentional
Focus, to
Acoustical
Signal, t1
Deadline
Status
Attentional
Focus, t1
App in Focus, to
Inspection
Interval, to
App Usage
Pattern, to
Desktop
Actions, to
App in Focus, t1
Inspection
Interval, t1
App Usage
Pattern, t1
Desktop
Actions, t1
Inference from Acoustical Events





Quiescence
Ambient noise
Human voice, conversation
Music
Telephone ringing
Inference from Acoustical Events





Quiescence
Ambient noise
Human voice, conversation
Music
Telephone ringing
Inference from Acoustical Events





Quiescence
Ambient noise
Human voice, conversation
Music
Telephone ringing
Visual Analysis
 Evidence
for user attending to system
Visual Analysis
 Evidence
for user attending to system
Sensing Content at
Focus of Attention

Studies of users’ gaze during
review of content.
Video
UFocus Prototype
Implicit Query
 Identify
content at user’s focus of attention
 Formulate query, provide related
information as part of normal work flow
 Background, implicit queries
Consider doc
structure,
basic scroll,
dwell patterns
General Handling of Notifications




Explosion of communications and
services
Multiple devices
Changing context, location
Varying attention, goals, needs
Scarcest resource is human attention
Notification Platform
Architecture & platform for handling
notifications from multiple services
 Deliberate
about attention-sensitive value
& costs of information, services
 Desktop,
cross device
 Leverage
rich contextual information
Notification Platform
Preferences
Sensors
Context
Analysis
Info/Service
Sources
Sensors
Devices
Notification
Decisions
Real-World Sensing
Unified Approach to Desktop and
Mobile Notification
Notification Platform hosts
multiple auxiliary services, user
assistance
 Communications
 Assistance, Tips
 New kinds of services

Life with a System-Wide
Attention Manager

Toward more courteous computing
Capture, deliberate about real-time, deferred
display, later review of journal
Fielding & Mediating Legacy
Alerts, Messages
• Defer
• Block
• Recast
Dear Mom,
Dear Mom,
Notification
Manager
!
Uninvited help
Psychological Studies of Disruption
Collaboration with Mary Czerwinski and Ed Cutrell
Video
Growing Set of Results

Measures of cost of interruption

Times to switch to alert and back to task

Total task time

User frustration
 Some
tasks more affected than others
 Costs
of interruption during different
activities, phases of work
 Relevant
vs. irrelevant alerts
Context-Sensitive Costs of Disruption

Different tasks

Finer grained actions within task

Direct interactions versus cognitive shifts
(e.g., type, select vs. file open, dialog)
4.6
4.6
4.4
4.4
4.2
4.2
4
4
3.8
3.8
3.6
3.6
3.4
3.4
3.2
3.2
3
Draw
Excel
Task Type
Word
3
File
Tool
Content
Interrupt Position Within Task
Relevant vs. Irrelevant Alerts
1
0.9
Relevant
0.8
Irrelevant
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
Total
Resume
Task Timing
Models of Attention in
Tutoring Systems






What exactly is the user’s attending to?
Where is the understanding problem?
When and how should I interject with some
assistance?
To what degree is the user engaged with the
system?
Is there a risk of the loss of a student’s
engagement?
What actions can be taken to enhance focus,
engagement?
Models of Engagement
Uncertainty, Utility, and Understanding




Rich set of problems across broad
domain of user modeling
Inference and decision making about
pedagogical actions under uncertainty
Expected utility to guide tutoring
interventions
Mixed-initiative, conversation, models of
attention: problems and opportunties