Transcript Proposal
Preference Elicitation
in Scheduling Problems
Ulaş Bardak
Ph.D. Thesis Proposal
Committee
Jaime Carbonell, Eugene Fink, Stephen Smith,
Sven Koenig (University of Southern California)
Outline
Introduction
Example
Preliminary results
Plan of work
Outline – Introduction – Example – Preliminary Results – Plan of Work – Questions
Motivation
Improve resource planning by
reducing uncertainty of the
available knowledge.
Outline – Introduction – Example – Preliminary Results – Plan of Work – Questions
Hypothesis
By asking the questions with the highest
potential to reduce uncertainty, we can
improve the quality of the resource plan
while minimizing the cost of elicitation.
Outline – Introduction – Example – Preliminary Results – Plan of Work – Questions
Initial schedule
Available rooms:
Room
num.
1
2
3
Initial schedule:
Capacity Projector
500
100
80
Yes
No
Yes
Requests:
• Invited talk, 9–10am:
Needs big room
• Poster session, 9–11am:
Needs a room
1
Talk
2
Posters
3
Missing
info:
Assumptions:
•
• Invited
Invited talk:
talk:
–– Projector
need
Needs a projector
•
session:
• Poster
Poster session:
–– Room
size is OK
Small room
–– Projector
need
Needs no projector
Outline – Introduction – Example – Preliminary Results – Plan of Work – Questions
Choice of questions
Initial schedule:
2
1
Talk
Posters
3
Candidate questions:
Requests:
• Invited talk,
talk: 9–10am: Useless info: There are no
large rooms w/o a projector
× Needs a large
projector?
room
• Poster session,
session: 9–11am: Useless info: There are no
larger room? unoccupied larger rooms
× Needs a room
Potentially useful info
√ Needs a projector?
Outline – Introduction – Example – Preliminary Results – Plan of Work – Questions
Improved schedule
Requests:
• Invited talk, 9–10am:
Needs a large room
• Poster session, 9–11am:
Needs a room
Info elicitation:
System:
Does the poster session
need a projector?
User:
A projector may be useful,
but not really necessary.
Initial schedule:
2
1
Talk
Posters
3
New schedule:
2
1
Talk
Posters
3
Outline – Introduction – Example – Preliminary Results – Plan of Work – Questions
Architecture
Natural Lang.
Optimizer
Elicitor
Ask user and
get answers
Update
resource
allocation
Choose
and send
questions
Outline – Introduction – Example – Preliminary Results – Plan of Work – Questions
Inside the Elicitor
Get list of questions
Each uncertain variable is a potential question
For each question i get
utilities for possible
answers
Plug in possible answers to the utility function
to get change in utility.
Get question score
Score i utility , i cost i
Return top N questions
Outline – Introduction – Example – Preliminary Results – Plan of Work – Questions
Optimizer
Uses hill climbing to allocate resources
Searches for an assignment of resources
with the greatest expected utility
Outline – Introduction – Example – Preliminary Results – Plan of Work – Questions
Related Work
Example critiquing [Burke et al.]
Have
Collaborative filtering [Resnick], [Hill et
al.]
Have
the user rank related items
Similarity-based heuristics [Burke]
Look
users tweak result set
at past similar user ratings
Focusing on targeted use [Stolze]
Outline – Introduction – Example – Preliminary Results – Plan of Work – Questions
Related Work
Clustering utility functions [Chajewska]
Decision tree [Stolze and Ströbel]
Min-max regret [Boutilier]
Choose
question that reduces max regret
Auctions [Smith], [Boutilier], [Sandholm]
Outline – Introduction – Example – Preliminary Results – Plan of Work – Questions
What is different?
No bootstrapping
Continuous variables
Large number of uncertain variables
Tight integration with the optimizer
Integration of multiple approaches
Dynamic elicitation costs
Outline – Introduction – Example – Preliminary Results – Plan of Work – Questions
Example Domain
Assigning rooms to conference sessions
Rooms have properties.
Sessions have preferences, constraints,
and importance values.
Each preference is a function from a
room property to utility.
Outline – Introduction – Example – Preliminary Results – Plan of Work – Questions
Example Domain
Rooms have properties.
Room 3 has one projector : 80% chance
Room 1 can accommodate 200 people.
Room 3 has no projectors : 20% chance
Sessions have preferences, constraints,
and importance values.
Each preference is a function from a
room property to utility.
Outline – Introduction – Example – Preliminary Results – Plan of Work – Questions
Example Domain
Rooms have properties.
Room 3 has one projector : 80% chance
Room 3 has no projectors : 20% chance
Sessions have preferences, constraints,
and importance values.
Invited talk very
cannot
important
be before 2 p.m.: 40% chance
Invited talk is
moderately
more important
important
than :poster
60% chance
session.
Each preference is a function from a
room property to utility.
Outline – Introduction – Example – Preliminary Results – Plan of Work – Questions
Example Domain
Rooms have properties.
Room 3 has one projector : 80% chance
Capacity of Room 1 is 200.
Room 3 has no projectors : 20% chance
Sessions have preferences, constraints,
and importance values.
Invited talk very important
: 40% chance
Invited talk moderately important : 60% chance
Each preference is a function from a
room property to utility.
Capacity preference
preference:is150
[150,
people
200, is250]
minimum,: 40% chance
200 people
Capacity
preference
is acceptable,
is [50,250
100,
people
150] is best.: 60% chance
Outline – Introduction – Example – Preliminary Results – Plan of Work – Questions
Experiments
Evaluation of RADAR
15 room properties
88 rooms
84 sessions
2500 variables
700 uncertain values
System asked to provide 50 top questions.
Outline – Introduction – Example – Preliminary Results – Plan of Work – Questions
Incremental elicitation
0.78
0.72
Certain
Incremental
Utility
Optimizer
estimate
0.58
10
20
30
40
50
No. of Questions
Outline – Introduction – Example – Preliminary Results – Plan of Work – Questions
Completed work
Questions based on potential reduction of
uncertainty
Empirical evaluation
Integration with RADAR
Outline – Introduction – Example – Preliminary Results – Plan of Work – Questions
Contributions
Fast computation of expected impact for
potential questions
Use of the optimizer for calculating more
accurate question weights.
Use of past elicitation results to improve
the elicitation process.
Unifying different elicitation strategies.
Outline – Introduction – Example – Preliminary Results – Plan of Work – Questions
Search for optimal questions
Example: Uncertain room size
100-150:40% chance
151-200:60% chance
h=20, max utility increase = 20
160-200:50% chance
100-160:50% chance
h=10
h=10, max utility increase = 30
100-130:25% chance
h=15
130-160:25% chance
h=15, max utility
increase = 100
Best-first search with the optimizer used as the
heuristic function.
Outline – Introduction – Example – Preliminary Results – Plan of Work – Questions
Elicitation rules
Encoding of elicitation heuristics
rule Uncertain-Auditorium-Size(room)
Conditions: type room Auditorium
mean size room 10, 000
std- dev size room 5, 000
Action: elicit size room
Outline – Introduction – Example – Preliminary Results – Plan of Work – Questions
Learning of elicitation rules
Derive rules based on past elicitations
…
Session
Room
Event
Room
Mean Elicitation
Type Elicit. Importance Property Value Result
… EventType
Room
Mean
Prop.Talk Prop.
Result 110
… Imp. Invited
Auditorium
Proj.
0.5
+
… 110
…
… 115
…
… 105
…
… 90
…
… 200
…
… 150
Proj.
Posters
Size
Best Paper
Proj.
Posters
Proj.
Talk
Size
Keynote
Proj.
0.5
+
Meeting R.
250
Auditorium
0.9
+
Classroom
0.5
Auditorium
100
+
Auditorium
0.3
+
115
105
90
200
150
Size
250
-
rule Learned-Rule(room,event)
Conditions:
type
room Auditorium
Proj.
0.9
+
Proj.
mean projector room 1
0.5
-
0.3
+
importance event 100
Size elicit 100
Action:
projector room+
Proj.
Outline – Introduction – Example – Preliminary Results – Plan of Work – Questions
Dynamic question costs
Same cost for all questions
Different cost for different question types
Learning of the question costs for each
type
Learning of the question costs for each
information source
Outline – Introduction – Example – Preliminary Results – Plan of Work – Questions
Experiments
Compare different approaches:
Current system
Search for optimal questions
Hand coded elicitation rules
Learned elicitation rules
Unified system
Human elicitor
Measure utility gain after each answer; also
evaluate running time
Outline – Introduction – Example – Preliminary Results – Plan of Work – Questions
Timeline
Best-First Search
Syntax for rules
Learning of rules
Learning of costs
Unified System
Experiments
Writing
Outline – Introduction – Example – Preliminary Results – Plan of Work – Questions
Addendum
Outline – Introduction – Example – Preliminary Results – Plan of Work – Questions
Absolute change in schedule quality
0.15
0.1
0.05
0
0
10
20
30
Question number (more important to less important)
40
50