Michael T. Cox

1
Planning as Mixed-Initiative Goal Manipulation

• Michael T. Cox
• Computer Science Engineering Department
• Wright State University
• Dayton, OH 45435
• mcox_at_cs.wright.edu
• DAGSI/AFRL HE-WSU-99-09 AFOSR
  F49620-99-1-0244

2
State-Space Planning

• Problem consists of initial and goal states
• Domain consists of type hierarchy and action set
• Operator consists of set of variable bindings,
  preconditions and post conditions
• Plan consists of sequence of operators

3
Planning as Search
I
G
4
Planning as Goal Manipulation

• Planning is steering goals through task
• Assigning resources to goals
• Time phased priorities
• Goal change
• Focus on Objectives
• Tractable planning
• Higher quality plans

5
Steering Goals through the Planning Process
Machine
Goals
Plan
Human
6
Goal Transformations

• A Goal Transformation is a minimalmovement of
  goal in a goal space
• In resource limited worlds
• lower (or raise) goals
• rather than fail (or miss opportunity)
• Exampleimpassable (R1) ? limited (R1)

7
Example Problem
8
General Form

9
Prodigy 4.0 User Interface 2.0
10
(No Transcript)
11
GTrans User Interface 2.1
12
(No Transcript)
13
Interface Evaluation

• Objective To compare the user performance under
  the search model and the goal manipulation model
• Manipulate the model and problem difficulty
• Cluster by expertise
• Measure goal satisfaction and keystrokes

14
Model Comparisons
15
Cross Domain Comparisons
16
GTrans and Prodigy/Agent
RMI facilitates multiple versions of GTrans
Dynamic goal sub goal relationships
PRODIGYgenerated plan
Goal transformation process
Human user 2 Interface
Current goals
Human user 1 Interface
17
Mixed-Initiative Replay
18
(No Transcript)
19
Conclusion

• Mixed-initiative computation improves
  (constrains)
• Complexity limitations
• Solution quality
• Solution speed
• Planning can successfully be viewed as a
  mixed-initiative goal manipulation task