RadarSpaceTime

1
RADAR/Space-Time AssistantCrisis Allocation of
Resources
2
Space-Time researchers
Faculty
Students
Chris Martens
UlasBardak
JaimeCarbonell
ScottFahlman
Research staff
Peter Jansen
Greg Jorstad
Brandon Rothrock
SteveSmith
EugeneFink
3
Outline

• Purpose and main challenges
• Demo of Space-Time Assistant
• Current and future learning

4
Purpose
Automated allocation of rooms and related
resources, in both crisis and routine situations.
5
Motivating task
Scheduling of talks at a conference, and related
allocation of rooms and equipment, in a crisis
situation.

• Initial schedule
• Unexpected major change inroom availability for
  example,closing of a building
• Continuous stream of minor changesfor example,
  schedule changes and unforeseen equipment needs

6
Main challenges

• Effective resource allocation
• Collaboration with thehuman administrator

• Use of uncertain knowledge
• Dealing with surprises
• Information elicitation

• Learning of new strategies

7
Architecture
Top-level control and learning
Processnew info
8
Outline

• Purpose and main challenges
• Demo of Space-Time Assistant
• Current and future learning

9
Outline

• Purpose and main challenges
• Demo of Space-Time Assistant
• Current and future learning

10
Learning
current work (RADAR 1.0)

• Information elicitation

11
Learning

• The system learns most of the new knowledge
  during war games
• It may learn some additional knowledge during the
  test

12
Information elicitation
The system identifies critical missing knowledge,
sends related questions to users, and improves
the world model based on their answers.
13
Information elicitation

• Input
• Uncertain information about resources,
  requirements, and user preferences
• Answers to the systems questions

• Learned knowledge
• Critical additional information about resources,
  requirements, and preferences

• Knowledge examples
• Size of the auditorium is 5000 50 square feet
• Size of the broom closet does not matter

Useful when the initial knowledge includes
significant uncertainty, and users are willing to
answer the systems questions.
14
Learning of relevant questions
The system analyzes old elicitation logs and
creates rules for static generation of useful
questions, which allow asking critical questions
before scheduling.
15
Learning of relevant questions

• Input
• Log of the information elicitation

• Learned knowledge
• Rules for question generation

• Knowledge examples
• If the size of the largest room is unknown, ask
  about its size before scheduling
• Never ask about the sizes of broom closets

Useful when the knowledge includes significant
uncertainty, users answer the systems questions,
and war games provide sufficient information
for learning appropriate rules.
16
Learning of default preferences
The system analyzes known requirements and user
preferences, creates rules for generating default
preferences, and uses them to make assumptions
about unknown preferences.
17
Learning of default preferences

• Input
• Known requirements and preferences
• Answers to the systems questions

• Learned knowledge
• Rules for generating default requirements and
  preferences

• Knowledge examples
• Regular session needs a projectorwith 99
  certainty
• When John Smith gives keynote talks,he always
  uses a microphone

Useful when war games provide sufficient
information for learning appropriate defaults.
18
Effective war games

• The systems knowledge during war games
  includes significant uncertainty
• Users can obtain additional information in
  response to the systems questions
• The world model and schedule properties during
  war games are similar to those during follow-up
  tests