Andrew Garland, Neal Lesh, and Charles Rich - PowerPoint PPT Presentation

About This Presentation
Title:

Andrew Garland, Neal Lesh, and Charles Rich

Description:

Responding to and Recovering from Mistakes during Collaboration ... Successful collaboration requires that the human and the system each maintain an ... – PowerPoint PPT presentation

Number of Views:35
Avg rating:3.0/5.0
Slides: 10
Provided by: andrewg7
Learn more at: http://lalab.gmu.edu
Category:

less

Transcript and Presenter's Notes

Title: Andrew Garland, Neal Lesh, and Charles Rich


1
Responding to and Recovering from Mistakes during
Collaboration
  • Andrew Garland, Neal Lesh, and Charles Rich
  • Mitsubishi Electric Research Laboratories

2
Introduction
  • Successful collaboration requires that the human
    and the system each maintain an up-to-date model
    of the shared plans and goals
  • What to do if the human makes a mistake or
    interrupts the activity (called
    non-contributors)?
  • Infer intentions underlying mistakes
  • Repair shared plan by adding recovery goals
  • Generate utterances that (re-)establish mutual
    understanding about non-contributors and
    recoveries

3
Trace snippet involving recovery
User presses the engage button. (mistake) Agent
says Whoops, it was too soon to press the engage
button. Agent says Lets recover from pressing
the engage button. Agent says First, lets set
the engine speed to zero ... Agent says Weve
recovered from pressing the engage button. Agent
says Lets return to starting the engine
4
Two error-recovery criteria
  • An agent must be able to indicate when recovery
    is needed, and which actions are part of the
    recovery
  • Re-use as much of general-purpose capabilities as
    possible, e.g. when
  • selecting method to achieve recovery
  • deciding whether to take the initiative
  • generating agent responses

5
Collagen
focus stack
plan tree
communicate
observe
observe
interact
interact
6
Example current discourse state
A
B
e
d
c
B A
c
Plan tree
Focus Stack
7
Related research
  • Plan repair and adaptation (Alterman 88, Hammond
    90, Hanks Weld 95, Nebel Koehler 95)
  • Intelligent tutoring systems (Conati et al. 95,
    Rickel Johnson 99, Rickel et al. 01)
  • Probabilistic plan recognition (Bauer et al. 93,
    Charniak Goldman 93)

8
Summary
  • Non-contributors are identified by relaxing
    constraints during plan recognition (part of
    discourse interpretation)
  • Utterances are used to (re-)establish mutual
    belief about
  • When recovery is needed
  • Which actions are part of recovery
  • Implementation re-uses many components that were
    built to handle normal behavior

9
Responding to and Recovering from Mistakes during
Collaboration
  • Andrew Garland, Neal Lesh, and Charles Rich
  • Cambridge Research Lab
  • Mitsubishi Electric Research Laboratories

10
Another snippet involving recovery
  • Later, in a similar context, the activity might
    go as follows

User presses the engage button. (mistake) Agent
says Whoops, youve made this mistake
before. Agent says You take it from here. User
says What next? Agent says Lets recover from
pressing the engage button.
Write a Comment
User Comments (0)
About PowerShow.com