Basing artificial emotion on process and resource management

Authors: Stefan Rank & Paolo Petta
Austrian Research Institute for Artificial Intelligence
Date: 2007-09-12

Presented at ACII 2007, Lisbon, Portugal (EU) • 2007-09-12.

████

Mother, child, dog poop

  • Hello, my name is Stefan Rank from the Austrian ...
  • The title of this talk is ...

try to open the presentation in a new unadorned window, then switch to fullscreen (F11 normally). If the slide text doesn't fit in your browser window, try decreasing the text size, try <Ctrl>+<->. Use <space bar> to advance, <Page Up/Down> & <arrow keys> to navigate. <T> toggles between slideshow and outline. Type a number followed by Return to switch to a particular slide. Slide controls are hidden in the bottom right corner.

Main Message

We propose a new approach for modelling the affective life of embodied social virtual agents:

Mother, child, dog poop

  • When creating social virtual agents, we want to consider Embodiment and Psychological theory
  • In the course of the talk we will understand what the implications of addressing these aspects are
  • and why we chose processes and resources as adequate building blocks, to improve on the status quo of modelling approaches

Disgust episode

First phenomenon to be modelled: an episode of disgust

Mother Child Dog poop

  • What we all agree we just saw:
    • Mother does not look at child
    • Child approaches dog poop
    • Mother looks at child
    • Mother looks at dog poop
    • Mother's expression changes while she retreats
    • Mother approaches child
    • Mother shows some more pronounced expression
  • As background story, I'll show a short episode of disgust (our first test case)
  • What we just saw: movement, grabbing, looking
  • storyline: ... expresses something towards the child
  • This episode is intended for the context of an interactive virtual world
  • what can we do to make that possible?

Implementation approaches

  • Agent architecture
    • Expressive behaviour triggered by specific situations
    • Modelling of internal aspects
  • BUT:
  • Scripting: that is scripting the whole spefic sequence of actions we just saw.
  • Need agents that actually interact with the world and react to the user
  • Improvement: agent architecture, trigger expressive behaviour
  • For consistency, coherence, etc: modelling of internal aspects
  • But: problem with thinking about specific situations to detect: boxology-driven approach, ie. not to be reduced to recognising boxes and mapping onto boxes; WHY:

Modelling requirements

  • Start with: general consequences of embodiment: limited control over body, over environment, continuous operation, no timeouts, several interactions at the same time.
  • On top of that emotional phenomena are fully integrated into embodied interaction and develop over time (example: mother first retracts then approaches)
  • We want richer models of what psychology tells us about emotional life
  • Scherer: several organismic subsystems are recruited to work together
  • Frijda: stresses that emotion usually overrides other activities
  • Implications for design: many processes, complete central monitoring and supervision is impossible, small-scale automatic coordination is the norm, emotion introduces coordinated larger scale changes. (?)

Implementation approaches (cont.)

Building blocks

  • Processes are concurrent strands of activity
  • Diagram: active (processing time), dormant (waiting for resource or notification), communication
  • Procs can create/start/stop other processes, and request control of
  • Resources: abstraction of limits: time, space, interactors
  • Configurations pose questions: types? hierarchy? how many? Periodically/frequency? Competition on resources?
  • Status quo: resources: processing time, interactors (sensors, actuators), communication channels; processes: concerns, world boundary, monitors
  • → general idea: get a richer, more detailed simulation of affective life; but limits of computational simulation apply:

Limits and Shortcuts

Consequence:

Examples of likely shortcuts:

  • Predicting and simulating for forming expectations as a resource
  • Acquiring relevant background knowledge
  • Interaction requirements pose temporal constraints and the implementation needs to atually run on a conventional computer
  • We need to cheat, but make it clear where and why, i.e. mark shortcuts and give reasons for excluding/simplifying
  • Example reasons: not tractable at run-time with current algorithm, or not enough developer time
  • Example: Predictions and simulating for forming expectations as a resource; detail knowledge needed for the specific episode already present

Status

  • Interactors for vision, movement are there; hearing and smell to do

Take-away Message

We propose a new approach for modelling the affective life of embodied social virtual agents:

Mother, child, dog poop

  • The quest to incorporate considerations of embodiment and psychological theory.
  • For agents in an interactive virtual world
  • Building blocks allow us to improve on the status quo and adress the implications

The End

Thank you for your attention!

Questions?

Mother, child, dog poop

  • Poster in the doctoral consortium session tomorrow.
  • (The following slides contain anticipated questions (AQ).)

AQ: not yet

Speaker's notes

  • Difference to other models/architectures
  • Implementation: why microthreads

Disclaimer and Acknowledgments

Keyboard Controls

The following apply in any supporting browser besides Opera, which uses the default Opera Show controls instead.

Action Key(s)
Go to the next slide
  • [Space bar]
  • [Return]
  • [Enter]
  • [Right arrow]
  • [Down arrow]
  • [Page down]
  • Click the left mouse button outside of the control area, Flash object, or movie
Go to the previous slide
  • [Left arrow]
  • [Up arrow]
  • [Page up]
Go to the title (first) slide [Home]
Go to the last slide [End]
Jump directly to a slide Type the slide number, then hit [Return] or [Enter]
Skip forward n slides Type the number of slides to skip, then hit any "go to next" key (except [Return] or [Enter])
Skip backward n slides Type the number of slides to skip, then hit any "go to previous" key
Switch between slideshow and outline view
  • [T]
  • Click the ø button
Show/hide slide controls
  • [C]
  • Move the mouse pointer over the control area

Further details of the S5 user interface can be found at Eric Meyer's S5 page and the s5project site.

References

[todo]no ref yet