I was recently referred to a paper written by Ken Perlin and Athomas Goldberg of the Media Research Laboratory at New York University. Their paper can be found at www.mrl.nyu.edu/improv/sig96-paper. This essay presents my reactions to the paper. Readers might well desire to consult the original paper so that they can draw their own conclusions.
The authors have created a program called Improv that is "a system for the creation of real-time behavior-based animated characters". Like the Erasmatron, it's got a lot of stuff in it, and so no short review such as this can do it justice.
I did find several factors in the Improv design that I thought worthy of criticism, but after some contemplation, I have come to a more sanguine assessment of the merits of this work. This is an academic paper; the values that guided its creation arise from the requirements of the academic universe. I have no credentials to judge how well it achieves its goals. I can, however, point out specific areas in which the academic goals of this work differ from the goals that I pursued with the Erasmatron. Readers must bring their own value systems to bear in judging which set of goals best suits their own interests.
For example, the Improv system boasts an impressive 3D animation system, combining some of the latest high-powered animation techniques to generate impressive output, all fully animated. This goes way beyond the simple 2D faces of the Erasmatron. However, I have my doubts as to whether this system can operate satisfactorily on a home computer. Someday, no doubt, home computers will have as many cycles and bytes as the workstations used in AI labs, but for now, this impressive work remains much closer to "research" than to "product" -- as it should, being an academic enterprise.
Similarly, their system has been carefully designed to operate well on a Wide Area Network; a clever scheme allows the software to operate satisfactorily even with fairly long network latency times. Again, this is impressive work, but its relevance to commercial application is, for the next few years, questionable. The software appears to assume a fast WAN; I shudder to think what consumers would experience with a 33KB connection.
An interesting direct comparison between the Erasmatron and Improv arises from the goal of both systems to be readily accessible to nontechnical artists. The Erasmatazz brochure boldly claims, "The big idea behind the Erasmatron is to make interactive storytelling technology directly accessible to artists. We're pushing the programmers out of the way to let the artists do their job with as much direct control as possible." The Improv authors describe "an 'english-style' scripting language and a network distribution model to enable creative experts, who are not primarily programmers, to create powerful interactive applications." It would seem that we're both moving down the same path.
However, there's a big difference between an academic standard of accessibility and my notion of accessibility. The Improv scripting language is really a set of extensions to conventional programming languages. Here's an example of one of their scripts:
[In the following example, the actor walks onstage, turns to the camera, bows, and then walks offstage again.]
define SCRIPT "Curtain Call"
{ "walk to center" }
{ continue until { my location equals center } }
{ "turn to camera" }
{ continue until { "turn to camera" is done } }
{ "bow" }
{ continue for 3 seconds }
{ "walk offstage" }
Now, this looks fairly straightforward, but there are a lot of syntactic assumptions built into this example. What happens if spacing is incorrect, or braces or quotation marks aren't balanced? How does the user know the difference between "continue until" and "continue for"? This is a ridiculously simple behavior; what would the script for a really complicated behavior look like?
In user interface terms, this is straight out of the 1960's. The user types in a bunch of text; the program compiles the text, reporting syntax errors; if it compiles properly, then it runs. Now in academic terms, there's nothing wrong with that, because this is not a research paper on user interface, it's a paper on agent behavior. I will criticize the authors at this point for overreaching themselves. When they claim that Improv will "...enable creative experts, who are not primarily programmers, to create powerful interactive applications", I think that they are going beyond the bounds of academic research. This thing has a long way to go before it can be used by civilians.
I hasten to confess that my own house is way too glassy for me to be throwing stones. The Erasmatron, after all, has plenty of techie-isms of its own. How about them Greek letters, those macros, and those equations? I could write a great many more negative things about the Erasmatron than about Improv, largely because I *know* the flaws in the Erasmatron so intimately. But I also know that I expended a huge amount of effort to make the Erasmatron understandable to writers, and I am proud of the progress that I did make. Clearly, much more effort will be required before the Erasmatron is truly accessible to non-technical people -- but at least I made a genuine start.
There is a deeper problem, though, one that I cannot easily exemplify in a short essay. It has to do with the intended audience for the paper. The authors are speaking to other AI professionals, not to writers, playwrights, or other creative professionals. Accordingly, they quite properly confine their design considerations to the intellectual universe of AI academia. In doing so, they minimize the value of their work to storytellers.
The fundamental task of a software designer is to act as a bridge between the starkly simplistic intellectual universe of the computer and the universe of the user. To be such a bridge, that designer must reach out toward the user's worldview and incorporate that worldview in the design. I can make no claims to understanding the intellectual universe of drama, but from the first time I heard Brenda Laurel talking about Aristotle, I knew that there was something important there. I've read Fields and Aristotle and a dozen others, I have shared long confusing discussions with people who do understand drama, and I've come to respect the intellectual edifice these people have built. I have tried to build many of these concepts into the Erasmatron; in many places I have failed. But in Improv I see a far less determined attempt in this direction. It is an AI researcher's impression of drama; characters can wave their hands, scratch their heads, move around, dance -- but I see no mention of motivation, conflict, plot, or resolution.
I don't want to be too hard on these folks; they're certainly making an effort. Consider, for example, this precious lump of academese:
"Because of this variability, the user's experience of an actor's personality and mood must be conveyed largely by that actor's probability of selecting one choice over another."
My memory is poor, but as I recall, Aristotle in the Poetics wrote: "Character manifests itself through action." As you can see, their hearts are in the right place even if their verbiage isn't.
The other day, WiredNews published a nice review of the Erasmatron. They did their homework and gathered the opinions of a number of independent observers. One, an AI researcher, warned that, until I publish an academic paper, my work won't have much credibility in the AI world. He's absolutely right -- but the converse is equally true: until the AI people come out of their labs, their work won't have much impact on writers.
Design Precepts for the Erasmatron