Oz Project

Working Notes, by Zach Tomaszewski

for ICS 699, Fall 2005, directed by Dr. Kim Binsted

Website/Overview

From: http://www.cs.cmu.edu/afs/cs.cmu.edu/misc/mosaic/common/omega/Web/Groups/oz/

• At CMU. Ceased Dec 2002. Alumni continue related work (Bates, Weyhrauch, Mateas, Sengers, etc.) Bates has formed the company Zoesis (demos, such as Demon's Pearl) w/ many alumni.
• Extending Laurel's notion of "interactive drama". "Highly interactive"--not simply choosing from fixed choices, as per hypertext.
• Oz: computer system to allow authors to create and present interactive dramas. Includes: simulated world, several characters (inc. model of mind and model of body in world), an interactor, a theory of presentation, and a drama manager. Interactor provides his character's "mind". DM controls world, character (minds), and presentation theory. Presentation filters/controls portrayal of world to interactor through his character.
• Three foci: characters (reactive, goal-directed, emotional, etc.), presentation (text or graphics, text physical input; best description/presentation of scene), drama (representing and controlling interactive story).
• Alumni: Mateas (thesis on system); Weyhrauch (thesis on search-based drama manager--"Guiding Interactive Narrative")--systems and drama aspects, more than agents. Brian Loyall thesis on believable agents; W.S.N. Reilly thesis on emotional agents.
• Related work, including B. Hayes-Roth's Virtual Theatre project at Stanford.
• Oz comes in two primary incarnations: Lyotard (text, cat--puzzle solving based on social interaction with the cat) and The Edge of Intentions (graphics, Woggles--social behaviors, believable agents). (Also, Reilly has been done a Lyotard-like game--Playground.)

Mateas, Michael. An Oz-Centric Review of Interactive Drama and Believable Agents. Technical Report CMU-CS-97-156, School of Computer Science, Carnegie Mellon University, Pittsburgh, PA. June 1997

• Again, user interacts through Presentation to control Avatar among other Characters in a World overseen by a Drama Manager.
• Oz -> believable characters. From Illusion of Life, Chuck Amuck, Art of Dramatic Writing: rich, specific personality; emotions; self motivation/goals/drives (not just reactors); change (grow over time consistently w/ personality); social relationships; illusion of life (responsiveness and depth of character).
• Classical (specific mental facility in general) vs behavioral (correct general behavior in specific context/environment) AI. Believable agents more behavioral, but generally differences--interested in specific audience perception, not general performance; characters, not AI realism.
• Top-down behavior from general personality formulations (Virtual Theatre project). Bottom up specification of behaviors (Blumberg's Silas the dog). Mix (Hap language used by Oz)--motivations and specific behaviors that can stem from those motivations.
• Thus pull and pushes between character arts, AI (and its goals), and robotics. Related to virtual pets/artificial life; chatterboxes. Rejects the search/assumption of emergent behavior in favor of artistic control.
• Oxymoron of interactive story. Based on Laurel, shaping interaction into story.
• Oz drama manager: controlling story at level of plot points. Author creates a (linear) story of points [scenes], then an evaluation function that will order a permuatation of these in the best [true?] order. Then, during interaction, the system responds to each user move from one plot point to another by analyzing all possible user and system moves (changes to the world), and ranks all resulting permutations, and then makes a system move most likely to result in the best story. (Chess-like approach.)
• Local vs. global evalutation. Global (entire story history): Oz. Local (response to each action): chatterboxes, etc--hopes for emergent story. Mix (script-and-demon approach): Dogmatix (Galyean's thesis, Narrative Guidance of Interactivity; also Pinhanzen); linear and graphs, daemons ensuring hints and obstacles. Demon complexity most limiting--generally can consider only past and relatively local actions.
• Granularity of control. Small grain (controlling actions, where essentially world == drama manager): hypertext. Large grain (control of scenes): Oz. Mix: script-and-demon. May need overall story manager, and then drama manager for each scene.
• Generative. Fixed story (static--tho branching--story; low replay value). Generative (can surpise author with new story; not yet interactive): Universe, Tailspin, Bringsjord's upcoming Brutus. Mix (variations on a plot/theme): Oz. Have less authorial control as system becomes more generative.
• Also consider interactive video--playing clips to a user depending on interaction.

Bates, Joseph. "Virtual Reality, Art, and Entertainment." Presence: The Journal of Teleoperators and Virtual Environments. 1(1):133-138, MIT Press, Winter 1992.

• VR so far: interface/controls, modeling the world. Presenting a case for importance of style and content for VR (as well as the snazzy presentation).
• Elements of novels, film, TV, etc: characters (creatures with intelligence and emotion), in a story (structured events), presented in a emotionally powerful style.
• World needs to respond without its original creator pulling the strings, so needs artistic knowledge.
• Oz is working on:
  • Cognitive/Emotional agents: subtle--use Eliza effect, that is people will suppose human characteristics if you don't offend their disbelief; broad and shallow abilities rather than narrow specialists.
  • Presentation:new media develop new techniques for this. Focusing now on text generation w/ style, but also interested in animation, music, and potential radio drama.
  • Drama: using a chess analogy, where the computer "wins" by coming as close as possible to author's intended story. Chess performance is impressive--perhaps similar results could be ported to story control. Such an advesary search approach need precise moves (plot moves) and evalution function, which requires a theory of the nature of a good story [see Lehnert ("Plot Units and narrative summarization" 1981 Cog Sci), Lebowitz ("Storytelling as planning and learning" 1985 Poetics), Dyer (In Depth Understanding)]. At fine-grained action level, this is thousands of potential branches. Can we work at some partly abstracted layer? And how to be subtle in DM control? Based on physical environment sims w/ live actors, seems DM can be intrusive, but substantial control possible through agents.

Kelso, Margaret Thomas, and Peter Weyhrauch, and Joseph Bates. "Dramatic Presence." Technical Report CMU-CS-92-195, School of Computer Science, Carnegie Mellon University, Pittsburgh, PA, December 1992. This paper originally appeared in PRESENCE: The Journal of Teleoperators and Virtual Environments, Vol 2, No 1, MIT Press.

• [See Overview, as it is an experpt from this.] In addition:
• Live Interactive Drama experiement (simulating Oz): How does it feel to be immersed in a dramatic virtaul world? What are the requirements of the characters? What are the requirements of the story/drama manager?
• The Director: Story guided by plot graph model (to give freedom of story movement). How does he guide the actors? Can the player be guided but not manipulated?
• Plot graph: directed acyclic graph, with major story moments as nodes. Edges reveate "must precede" relations, and include hints and obstacles that can affect transition between scenes. Some traversals likely to be better than others, but assume author has constrained enough to establish basic dramatic tone.
• Hint example: bringing another character in in order to facilitate a meeting scene. Obstacles can slow down pace if it's too fast. Edges may also direct the flavor of the presentation (more sombre at discovery of a body).
• Bus station scenario. Interactor given goal--not necessary for story, but sets frame and gives him direction. Details of 2 runs--different interactors, different speeds, different endings. Second run, actors knew their roles and played off each other with less director control.
• Interactors found the interaction quite real, appreciated the goals, and would play another storyline (but not this one again). Questioned degree to which characters motivated interactor response--in one case, victim is surly, in another, pathetic. Director important to keep action moving--introducing next scene/movement when the action begins to settle.
• Interactive immersion of this sort more immediate and personal (not based on empathy for some fictional protagonist). This dramatic presence, culminating in a personal moral choice in a specific context, was very powerful--despite apparent scene and costume distraction, ignored by suspension of disbelief (felt real).
• Some blatant character inconsistencies tolerated by interactor, but not by observers.
• Interactor thought things flowed quickly. Observers watched dead space and actors "looping" while interactor made up his mind.
• In another (less reported) experiment, Tea for Two, had to push interactor towards ending to meet time limit. Interactor felt manipulated. Also, hints from a DM-like narrator where not tolerated well, whereas hints from other characters within the story were fine. Further evidence that agents may be best medium of DM control.
• [Important: Director moves action along when pace lags. What sort of drama knowledge does this require? More than just scene sequences. (Story time vs. discourse time; non-reactive hints and obstacles.) Can AI do this?]

Seek: "Guiding Interactive Drama" Peter Weyhrauch. Ph.D. thesis, Tech report CMU-CS-97-109, Carnegie Mellon University, January 1997. Describes the Oz drama manager, a search-based system for guiding an interactive story experience.

Includes Mateas (thesis 2002).