MixedInitiative Planning

1
Mixed-Initiative Planning

• Yolanda Gil
• USC CS 541
• Fall 2003

2
Outline

• Motivation and challenges
• Dialogue issues in mixed-initiative planning
  TRAINS Allen and Ferguson 02
• Integrating user guidance with a planning
  algorithm PASSAT Myers 97

3
Further Reading

• TRIPS (more at http//www.cs.rochester.edu/researc
  h/cisd/projects/trips)
• Allen, J. and Ferguson, G. "Human-Machine
  Collaborative Planning", to appear in Proceedings
  of the Third International NASA Workshop on
  Planning and Scheduling for Space, Houston, TX,
  October 27-29, 2002.
• PASSAT (more at http//www.ai.sri.com/passat/)
• Myers, K. Abductive Completion of Plan Sketches,
  AAAI 1997.
• Myers, K. L. and Jarvis, P. A. and Tyson, W. M.
  and Wolverton, M. J. A Mixed-initiative Framework
  for Robust Plan Sketching, ICAPS 2003.
• Myers, K. L. and Morley, D. N. Human
  Directability of Agents, K-CAP 2001.

4
Example Dialogue Evacuation Plan
5
Planning Assistant Plan Design through
Delegation

• Take the people in Delta to Exodus by
  helicopter, then by bus to Calypso. Take the
  people in Abyss to Calypso by bus. Take everyone
  in barges to mainland locations.
• May need to interact with the user for
  clarifications, inconsistencies, unfeasibility,
  etc.

6
Planning AssociateCollaborative Plan Design

• Here is the situation, what can you do for me?
• Q What do we have?
• A What do you need?
• Q What can you get?
• A What would you do if you had more?
• Q When do you need it?
• A What happens if we postpone deadline?

7
Range of Interactive Planning Tasks

• Generation of a solution plan generation
• User specifies goals and tasks for the solution
• User indicates preferences or constraints to be
  used during planning
• Assessment of a solution plan evaluation
• User indicates criteria to analyze plan features
• Tradeoff analysis in solution quality plan
  comparison
• User navigates solution space and indicates
  preferences
• Resource assignment scheduling
• User indicates resource allocations, temporal
  constraints
• Problem formulation design of the planning task
• User adds and retracts constraints on the
  planning problem
• User establishes policies for systems
  responsibilities

8
Challenges

• Interpreting user input
• Mapping into possible operations/responses
• Disambiguating requests
• Intelligent search
• Managing classes of solutions
• Tracking constraints and previously explored
  solutions
• Facilitating users cognitive task
• Grounding the discussion with a specific plan
• Acting on the users input
• Flexible planning framework that can support
  collaboration

9
TRIPS Collaborative Planning Dialogue Allen and
Ferguson 02
10
TRIPS Interpreting Users Requests1) Mapping

• Mapping request to possible responses
• An operation or command
• Eg how long will this take? -gt request for plan
  evaluation
• A modification of some aspect of the plan
• Eg what if we allow a stopover? -gt introduce
  goal
• Approach Lay out the types of problem solving
  operations and plan modifications allowed
• Each type has necessary conditions used to map
  the request to the relevant portion of the plan

11
Types of Problem Solving Operations
12
TRIPS Interpreting Users Requests 2)
Disambiguation

• System may need to disambiguate
• Eg Can we use a helicopter to get the people
  from Abyss?
• -gt extend current solution (eg if still
  missing Abyss evac)
• -gt add this goal and generate a new
  solution (eg if at Delta)
• -gt modify current solution (eg if now
  using a bus)
• -gt analyze feasibility of any of the
  above, rather than do it
• Approach Assume user wants to continue to work
  on the same plan and situation unless he/she
  indicates otherwise
• Heuristic to prefer in this order
• Extend current solution
• Modify current solution
• Introduce new goals
• Need to track dialogue history

13
TRIPS Plan Representation

• Four related views on plans
• Objectives goals and constraints
• Tasks abstract solutions (classes), their
  constraints, causal connections
• Resources objects available for use in solutions
• Situations states before and during execution
• A grounded, straw plan
• To ground the conversation, uses a concrete
  realization of the abstract plan that is
  currently under consideration

14
TRIPS Hybrid Plan Generation
15
PASSAT Mixed-Initiative Plan Authoring for HTN
Planning Myers 97

• User specifies tasks (prim or non-prim) that
  should be part of the solution a plan sketch
• System completes plan sketch
• Hypothesizes top-level goals
• Refine goals to include sketch tasks (anchors)
• Extended to handling incorrect sketches Myers
  03
• Orphaned tasks that do not map to any top-level
  goals
• Inconsistencies with template constraints

16
Sample Domain
17
Sample Sketch P,V
Possible plans V J L W Q F V L W Q
18
Anchor Chains for V
19
(Inverse) Anchor Chains for V
20
Goal-Anchor Graphs for V
21
Goal-Anchor Graphs for V
GA(VB)
GA(VA)
22
Goal-Anchor Graph for P
GA(PB)
23
Plan Skeleton for B GA(PB), GA(VB)
GA(VB)
GA(PB)
24
Slicing a Plan Skeleton for B (I)
GA(VB)
GA(PB)
25
Slicing a Plan Skeleton for B (II)
GA(VB)
GA(PB)
26
Slicing a Plan Skeleton for B (III)
GA(VB)
GA(PB)
27
Slicing a Plan Skeleton for B (IV)
GA(VB)
GA(PB)
28
Slicing a Plan Skeleton for B (V)
GA(VB)
GA(PB)
Slice 1
29
Slicing a Plan Skeleton for B (VI)
GA(VB)
GA(PB)
Slice 2
30
Generating Plans by Refinement of Slices
Slice 1
Plan from slice 1 V J L W Q Plan from slice 2 F
V L W Q
Slice 2
31
Sketch Completion Algorithm

• 1) Abduction stage
• - Generate anchor chains for tasks (anchors) in
  sketch
• - Nondeterm. select leaves from the chains
  (top-level goals)
• - Create plan skeletons for those top-level
  goals anchors
• 2) Refinement stage
• - Follow HTN refinement algorithm selecting
  task
• reductions consistent with skeletons
  through slicing
• - If no slice possible, return failure

32
Recap and Summary

• Dialogue issues in mixed-initiative planning
  TRAINS
• Interpreting user input, disambiguation
• Plan representation as goals/tasks/resources/state
  straw plan
• Hybrid planning architecture
• Integrating user guidance with a planning
  algorithm PASSAT
• Incorporating the users input into a plan
  generation algorithm