November 14, 2007
Last weekend, the AAAI Fall symposia were held in Arlington, just outside Washington, DC. Nick, Michael and I were there for the Intelligent Narrative Technologies symposium, among 60 people in attendance from the Americas, Europe, Australia and Asia. Like the 1999 symposium on Narrative Intelligence, this was a gathering of both accomplished researchers and new faces to the field. (My first foray into academia was at one of these meetings, exactly 10 years ago, at a Socially Intelligent Agents AAAI symposium at MIT, where I co-presented Petz and first met Michael and several other Oz Project members.) Unlike the 1999 NI symposium, this time around it was almost completely academic computer scientists, with almost no industry folk in attendance and few who would identify themselves primarily as artists or writers.
An important general point brought out during the symposium was the need for a standard platform for interactive character and drama research, which if adopted by multiple groups, could greatly increase sharing of research results. I couldn’t agree more; we started talking about how to make that happen.
For me, the most exciting talks were on the final day, focusing on story generation and representation. But I’ll run through the program chronologically, quickly summarizing most of the talks, and going into detail on a few. I’ll link to those pdf’s that are currently available online.
The first talk was by Michael Kriegel, a grad student from Heriot-Watt University in Edinburgh presenting a concept for a behavior authoring system in which human authors manually puppeteer NPCs, to create behaviors for them to later interact with the player. There was some discussion about how authors would be given the ability to author truly novel behaviors. … David Thue, a grad student at the University of Alberta, presented an architecture (pdf) in which the player’s preferred play style is covertly measured, via normal gameplay, at the start of a game; then the game chooses particular events to occur later that match these preferences. In the demo session later in the symposium I played a few minutes of a Neverwinter Nights mod called PaSSAGE that identified me as some who likes Tactician-style gameplay, instead of Fighter, PowerGamer or others. … Next was a paper called Emotion Discourse as Design Heuristic, which I missed.
Ian Horswill, who has recently gotten back into interactive character research, presented a position paper where he mused that it’s just as or more important to model dramatic characters’ dysfunctional behavior as anything else. He presented the example from Casablanca as case in point: if Rick hadn’t been drinking when Ilsa reappeared in his life, their dramatic scene in the bar might not have happened (“of all the gin joints…”).
Next there was a panel discussing the challenges of authoring interactive drama, moderated by Michael. I was one of three panelists, discussing my experience authoring Façade, and my current work, which includes developing authoring tools for the The Party. My co-panelists were Bryan Loyall, an original Oz Project member, co-founder of Zoesis, now at BAE Systems; and Ana Paiva, a professor at Instituto Superior Técnico in Portugal who has long been developing interactive characters and emergent narrative, her latest serious game / interactive drama project being FearNot! (pdf), a collaboration with Ruth Aylett and many others, demoed at the symposium. Michael framed the discussion by suggesting that the state of research into intelligent narrative technologies is overly comprised of theory and architecture design, with only a small amount of demonstratable technolgy being built, and an even smaller amount of playable systems released to the public; such publicly playable systems are crucial to test the claims made by the theories and architectures. Michael asked us panelists, are better authoring tools the solution to speeding up the development of interactive drama technologies? And, who are likely authors? Bryan compared it to the history of animation, also a highly technical art form. He, like the rest of us, thought authoring sophisticated interactive drama would require trained artist-programmers, even if good authoring tools existed. I suggested that if authoring tools were made for the general public, perhaps in the form of an entertainment product (like how the Sims is a kind of authoring tool), that sketchy, amateurish interactive dramas would be the result, but the authors of them would be happy because they made them and had their family and friends play them. We also agreed that personality models based only on the Big 5 traits aren’t enough to capture the richness of characters we want in interactive drama. (This issue came up later in discussions of personality-specific natural language generation.)
David Roberts of Georgia Tech presented some sophisticated techniques for modeling player preferences and satisfaction, by analyzing player behavior. This involves measuring and analyzing features values such as player motivation, plot mixing, plot homing, and so on. Read the paper (pdf) for more detail, it gets complicated. They hope to apply these results to drama management. … Next, Nick presented his thesis work on nn, a system that can alter the narration of an interactive fiction in a myriad of ways; in his presentation he focused on how events can be reordered, and how that can affect the experience for the player. I enjoyed seeing more detail about the system, I think it has a lot of promise for moving IF in a more generative direction. … Grad student Santiago Ontañón of Georgia Tech presented a drama manager that attempts to build a player model during game play. They report some increase the overall quality of the player experience from this technique, though Santiago admitted the qualitative results (positive player comments in a questionnaire) were more encouraging to him than the quantitative measurements they made.
A session on building affective, emotional characters was next. By this point in the symposium, various hallway conversations were underway that I joined in, so I only caught part of David Pizzi’s talk on creating an interactive Madame Bovary story that integrates an emotion model with behavior planning; Sandy Louchart on building characters for FearNot! (mentioned earlier) by using a “double appraisal” method, in which NPCs, when choosing what actions to do next, take into account how they would feel if the action were performed upon themselves; and I altogether missed Ana Paiva’s paper on developing interactive shadow theater (pdf), one of several projects from GAIPS.
The next morning began with Masters students Josh Tanenbaum and Angela Tomizu from Simon Fraser discussing Scarlet Skellern and the Absent Urchins, in which players can affect the mood and tone (but not the plot) of a charming Edward Gorey-like story presented in a series of comic panels (pdf). … Next Jonathan Rowe presented an ambitious project underway at NC State’s IntelliMedia group called Crystal Island, an interactive mystery built on top of Valve’s Source animation engine. Rowe is interested in identifying and measuring the factors that go into narrative presence in such experiences, such as consistency, plot coherence, affect, motivation, narrative load, and so on. … Santiago briefly talked again, this time about driving interactive drama research through building complete systems. … Chunyan Miao from Nanyang Tech in Singapore presented an architecture involving plot planning with fuzzy cognitive goal nets (pdf). … I missed a subsequent presentation on narrative learning environments.
Brian Magerko, one of the symposium organizers, presented some thoughts on how one could go about measuring dramatic believability. Essentially he’s suggesting that believability must be measured against the expectations set by the characters; for example, realistic looking characters that speak to you, e.g. Alyx Vance in Half Life 2, suffer in their believability if you can’t talk back to them, or if you hit them over the head with a crowbar and they continue as if nothing is wrong. However, in his formulation he wants to reward approaches that don’t merely go for overly simple representations, and do more than aim for low expectations; that is, as he puts it, “a character is ‘more believable’ if it is closer to some standard of realism for that modality”. While I like the idea of rewarding ambitiously-represented characters, it’s unclear to me if that particular definition of ambition works.
Next were short presentations from natural language generation (NLG) researchers, wondering how their work could be applied to AI-based narrative. James Lester of NCState reminded us of some very interesting recent NLG work by Charles Callaway, including sophisticated narrative prose generation in the domain of Little Red Ridinghood stories; this is now high on my reading list. Anyhow, we ended up having discussion over lunch about NLG and interactive drama, which fed back into the discussion about authoring.
After lunch Maria Arinbjarnar presented her initial explorations into applying game theory to rationally choose the next sentence for an NPC to speak, from a Bayesian nets knowledge base. … Next Zach Tomaszewski, grad student from U. Hawaii advised by our good friend Kim Binsted (of JAPE fame), presented his explorations into using Propp as framework for managing interactive drama. In a nutshell he concluded that Propp is too limiting for this; “Propp describes what must occur within a function”. He’s now moving on to an Aristotelian tension-arc based approach (sounds familiar! :-) … I missed Ivo Swartjes‘ presentation on applying improv techniques to emergent story generation (darn those hallway conversations!). Here’s the pdf. … I’m glad I made it back for Beth Cardier’s stand-out talk on the poetics of writing and the mechanics of narrative. In her talk The Story Molecule: Narrative as Information, she outlined her model of of her fiction-writing process, which is more free-associative and conceptual than event-based and plot-centric. She concludes “it makes sense to construct a narrative-based knowledge system as a writer would imagine it, instead of how a theorist would write about it”. She was one of the few in the room thinking as an artist, and it was a breath of fresh air after two days of talks that tended to lose sight of the need for beauty in story.
The final half-day began with a cool project, Minotour, presented by UCSB Masters students Brent Hecht and Nicole Starosielski. Using a GPS-enabled handheld device displaying a map of your current location, knowledge about your location is mined from Wikipedia, and presented to you in the form of a narrative as you walk around. Their test case was walking in Berlin from the Brandenberg Gate to Unter den Linden; as you walk, a narrative is generated that starts and ends with specific, spatial information, and working in more general and thematic narration in the middle. They’re attempting to shape your experience to be aesthetically pleasing and coherent. I suggested they could extend the system to include narration about a fictional character who walked your same path in historical times.
Next Brian Logan presented Neil Madden‘s and his work on creating agents in MMORPGs that would observe player activity, such as quests and battles, and later report on and summarize these, to be posted to forums and blogs. It’s an interesting method to extract stories from MMORPGs. I suggested they have the agents act as true embedded reporters, within teams going into battle, asking questions and truly interviewing players, getting quotes, etc.
Rafael Pérez y Pérez, developer of MEXICA, presented some new work developing a system to generate visual daydreams. He showed a brief animation, and admitted the work was in very early stages, but presented the results of a small survey that some users classified it is as being daydream-like, versus a story or non-sense narrative.
David Elson, a grad student at Columbia and sometime commenter here on GTxA, described some fascinating ongoing work into encoding stories (pdf). He’s building a system called Scheherazade focused on “formally representing, storing, and statistically analyzing stories, as well as the manner of their tellings”. He’s building a tool for users to “symbolically encode stories using a graphical interface, and see the result rendered in prose”. I found this really really interesting, and look forward to the progress he makes.
Next up was Erik Mueller, whose mild manner betrayed his heavyweight credentials in the field. Erik is a highly accomplished researcher who has created several extremely impressive systems, including Daydreamer in the 1980s, ThoughtTreasure in the 1990s, an event calculus for commonsense reasoning, and recently a book on the topic; you can watch a recent lecture of Erik’s at MIT, introduced by Minsky. At the symposium, Erik presented his most recent work on story understanding (pdf). Although the paper is filled with obtuse-looking first-order logic, his talk was very clear and understandable, describing how his system can reason about stories, such as the following:
Willa was hungry. She grabbed the Michelin guide. She got into her car.
From that, the system can figure out all kinds of untold details that must have gone on with Willa, such as reading the guide to find a restaurant, going to restaurant and eating, and so on.
Grad student Emmett Tomai presented another ambitious project, a member of Ken Forbus’ Qualitative Reasoning Group at Northwestern. They’re attempting to extract temporal and causal networks from a corpus of Aesop’s Fables, and using them to reason about and characterize related stories. By the end of this talk my head was spinning, in a good way.
Finally, Malcolm Ryan from U. New South Wales analyzed a Peter Rabbit story, pointing out how sophisticated even these “simple” stories are. It was an entertaining talk, a good way to end the symposium, with the remark that work in this area is “very hard, but exciting”.
There was follow-on improv session led by Brenda Harger of CMU ETC, that I had to miss, sadly.