This month's column presents several views from the edge of software development. Mostly these views are problems posed or solutions suggested at a recent conference on artificial intelligence. But there's also a peek at a strange and violent world where real robots do real battle for the amusement of an evolving underculture whose tastes run to that sort of thing. And there's something about the information superhighway: obligatory, these days, like car crashes in certain kinds of movies.
In short, this column won't help you to meet last Wednesday's deadline, but it may satisfy your bloodlust.
Let's see hands of those who want to read about a car crash involving a Microsoft executive.
Uh-huh. That's about what I expected. You Microsoft employees: one hand only, please. Those Microkids are so eager.
I have to warn you: It's not much of a car crash. I don't want you to be disappointed. I'll try to work in another one later in the column to make up in quantity for what they lack in quality. The real carnage and destruction comes in the robot wars section of the column. First, though, this_.
You hold a software-development conference in Seattle, presumably, to make it convenient for Microsoft to send over a guest speaker.
It worked well at the Twelfth National Conference on Artificial Intelligence (and Sixth Innovative Applications of Artificial Intelligence Conference), henceforth AAAI-94, this August: Tuesday's Invited Talk was delivered by--Steve Ballmer??
Okay, so Microsoft's executive vice president of sales and support might not have been the program committee's first choice to address an audience of highly technical, academically oriented developers, but Steve is always a dynamic speaker. And he more than made up for any ignorance of just exactly what the attendees were doing for a living with his laser-sharp vision of what they ought to be doing.
Port your "cycle-sucking apps" to Windows, Ballmer told the crowd, citing a 44-million/year PC sales rate and a 150--200 million installed base. PCs, Ballmer said, are selling faster "than cars and small trucks."
He acknowledged that ordinary Windows is, for these AI folks, a "toy" operating system. He pointed them NT-ward, not dwelling overmuch on just what fraction of that installed base now uses or ever will use NT. But NT is the future, he said, adding, "long-term, we don't need to invest in two streams of R&D." Translation: Windows is future history. A pre-denouncement?
If the Seattle location worked well for the program committee, it didn't work so well for Steve, who dinged his car in the Convention Center parking garage. His car. If he had been at, say, COMDEX in Las Vegas, it would have been a rental.
An AI navigation system that kept him from bumping into things would be nice, he told the crowd.
At COMDEX in Las Vegas, a comment like that from a Microsoft executive would send a dozen entrepreneurs back to their suites to write business plans. Not here, though, I suspected. Not with this crowd.
Who were these people? I decided to do some objective tests to confirm what I suspected. Suitcounting the crowd at Ballmer's talk, I got a low 20 percent. Not a suit crowd, then.
Nor were they escalator striders. There are two kinds of people: those who treat escalators like elevators and those who treat them like stairs. Riders and striders. I've never known an entrepreneur who wasn't a strider. Here at AAAI-94, the strider-to-rider ratio was low, lower than at most trade shows.
My conclusion: The attendees at AAAI-94 were neither suits nor striders; neither blind followers nor tunnel-vision fanatics. This could be an interesting conference, I thought.
The list of tracks in the program reflected what the program committee took to be the active areas in AI research today: causal reasoning, spatial reasoning, nonmonotonic reasoning, model-based reasoning, uncertainty management, constraint satisfaction, knowledge bases, distributed AI (collaborating agents, for example), robotics, perception, machine learning (from reinforcement learning to induction and discovery), natural-language processing, neural nets and simulated annealing, planning, scheduling, and search.
In addition to the technical presentations, there were the student poster sessions expected at any academic conference, exhibits, a video program, a special program on advances in machine translation, a robot competition, and an AI art show.
I hope you won't be disappointed if I don't tell you all about the student papers, the video program, or the art show. I won't say much about the exhibits, either, because they were--how shall I put it--pitiful. Not the individual exhibits, just the number of them. A couple dozen companies saw fit to buy booth space. I'm not including publishers' row: Looking at the number of publishers who exhibited and the number of AI-related books they have brought out recently, you'd conclude that AI was an exciting field.
Raj Reddy of Carnegie Mellon University delivered the keynote address for the conference, on the state of AI--no, sorry--on "The Excitement of AI." It's nice when the title conveys not only the subject but also the bias of the talk.
What this field needs, Reddy told the crowd, is a few decades of quiet, sustained progress. Just keep the funding flowing and leave us alone, I guess he means, and we will show you some exiting stuff. Like AI cruise control in every new car. (Reddy gave this talk before Ballmer dinged his car.)
Most of the cruise-control work has already been done, Reddy said, and could be put into action really soon. There's been similarly impressive work in intelligent automated tutoring and in planning and scheduling, the latter an outgrowth of Desert Storm. Current applications of this war work include smarter manufacturing and disaster management, with potentially big payoffs. There are ten or twenty other AI areas, Reddy said, that have shown just as much success.
Fine, fine. So why does AI get no respect? Could it be_forty years of effort at a half billion a year with few publicly visible results? Could that be it?
Probably, Reddy opines. And then there's the academic publish-or-perish mentality that leads to a focus on short-term results, while interesting AI problems generally require a large, sustained attack, and the really interesting problems, he says, probably need a 1000 to 10,000 person-year effort. Not that much different from Microsoft's NT development.
This idea of breaking down the big problems of AI and tackling manageable pieces--not particularly novel, one would think--pervaded the invited lectures at AAAI-94.
So what to do? Reddy recommends:
Reddy laid down some challenges for laborers in the field of AI:
Whew. Barbara Grosz, in her AAAI presidential address, picked up the theme of collaborating agents. Grosz gave some helpful distinctions and definitions:
"Collaboration," as opposed to interaction, a familiar computer-science paradigm, involves processes (or people, of course) working together rather than acting on each other. Collaborative activities are characterized by:
A lot of presentations at AAAI-94 dealt with collaboration among agents. A recent issue of Communications of the ACM was also dedicated to the subject. Must be a fad, I mean, trend.
Reading back over what I've written, I detect a note of cynicism. I regret it, because AI is important and the AAAI-94 attendees are doing serious, cutting-edge computer-science work. This attitude of mine is probably due to my belief that the term "artificial intelligence," like the term "liberal," is dead, and that those who have embraced it ought to just get a new word.
I also regret there being a note of cynicism in this column up to this point because we have just arrived at the point where cynicism is truly deserved.
We have finally come to the obligatory information-superhighway mention.
Depending on the level of cynicism one brings to a discussion of the information superhighway, it is either the post-cold- war $500 toilet seat, the moral equivalent of the B-1 bomber, a runaway metaphor, the future home of mankind, or as good an excuse as any for funding computer-science research.
The information superhighway will certainly need computer-science research, said ARPA's Kirstie Bellman in her invited AAAI-94 talk.
Actually, the impressively credentialed Bellman, who appears to juggle complex ideas and humongous projects with equal ease, makes a good case for needing AI in the infrastructure of the infobahn. The information highway surely will be a huge heterogeneous network made up of diverse components, since the closest thing we've got to the information superhighway, the Internet, is that already. Whenever we are dealing with a complex system, she said, it's basically a modeling problem. We will need the power of new formalisms, but they need to be the right formalisms. Even the word formalism is often used too loosely these days, she said.
That was apparently the point at which I passed out in the rarefied atmosphere, because my notes jump abruptly to an audience member asking Bellman if there is any connection between what she's been saying and anything in the PC universe.
"The government's role in this is to get out of the way," Bellman answers. "However--"
However, the commercial market is not doing well in this whole area of infrastructure. Too many commercial companies crassly trying to own the platform. "We need to keep that from happening," she said, but didn't say how. She did offer an insight that applies to a lot more than the information infrastructure, though:
"Don't confuse standards with uniformity. An individual user wants a uniform interface, but not all users want the same interface."
Intimidated by Bellman's search for the right formalisms, I sought something more grounded. I found the AAAI-94 robot competition.
AAAI does this every year; last year, the robots moved large boxes around. This year, they picked up and dropped several kinds of small objects into waste baskets. They distinguished among these small objects--soda cans, Styrofoam cups, and paper wads. And they moved through a multiroom environment cluttered with chairs and tables as they performed their janitorial chores.
It would have been more impressive if all the robots had actually done all these things, but many took penalty points and virtualized some of the steps--like picking up the objects and putting them in the waste baskets. But there was still a lot of computer science going on.
Chip, a robot from Chicago, for example, used texture, size, and color information in its visual system. It used sonar to ensure that its arm didn't hit chairs and tables. It used infrared to tell if the object was in its grip and tactile feedback to keep from using too much pressure in its grip.
Stanford's entry had a distinctive search strategy, moving to the center of each room and patiently scanning the entire room for objects, then zipping from object to object to (virtually) pick up and basket the trash. Most robots searched for new trash after each pickup, although they generally remembered where the closest waste basket was.
We now jump two weeks ahead and 700 miles south to San Francisco and the first annual Robot Wars event, the brainchild of Animatronics special-effects design engineer Mark Thorpe.
The difference between the two competitions is immediately apparent. The AAAI roboticists wore team t-shirts. At Robot Wars, black leather and camouflage coloration predominate. A member of the program committee gave the laconic play-by-play in Seattle; in San Francisco, the M.C. would have been at home in Thunderdome. There are distinct robotics cultures growing up. The culture at this San Francisco event owes something to demolition derby and something to Mad Max.
Understand, Robot Wars was not an academic demonstration in robotic technology. It was pitched battle in an arena, with no quarter asked or given.
A huge poster at the door warned that attendees, by the act of entering the building, were accepting the risk of injury from flying robot parts.
The program wasn't too complicated. There were various events, but basically the (radio-controlled) robots just beat each other to staggering scrap. The judge cleared the hall for the lunch break with a jet-engine blast that must have violated a city ordinance or two and certainly wasn't healthy, and after lunch the carnage resumed.
One crowd pleaser was the robot with the chain saw, which hacked the bodies and electronics of its opponents, as well as chewing on the arena itself. Another robot was studded with spikes, apparently to intimidate the other robots. It didn't seem to work. My favorite was the house robot, supplied by the sponsors to serve as a neutral nemesis in some of the events. In the final free-for-all, it was the house 'bot that reduced the others to inert rubble on the arena floor, triumphing by use of an arm/paddle appendage with which it swatted opponents and flipped them upside down.
There wasn't much computer science going on. What did seem to be going on was a new form of entertainment. I dunno, this could be bigger than disc golf. The main drawback to robot wars as a sport, it seems to me, is that the toys get fairly wrecked in the game, but I guess the same is true of football.
Oh, yes. The second car crash. The car I was in on the way to Robot Wars was rear-ended in a four-car pile-up on the Coast Highway. No big deal. Not my car. Why I was in a rental car in the San Francisco area, where I live, while Steve Ballmer was in his own car in Seattle, where he lives, is hard to explain. Probably a karma thing.
Copyright © 1994, Dr. Dobb's JournalObligatory Car Crash
Riders versus Striders
The Exciting Part
Reddy's Recommendations
There was more talk of collaborating agents during this week in Seattle than at a dozen writers' conferences.
After that, Reddy suggests, we can get on with the real work of AI: the creation of superhuman intelligence.Can't We Just Get Along?
Intentions come in two varieties, and these have different roles in collaborative planning. There is the intention to do something, and there is the intention that something be the case. Individual plans and collaborative plans are not the same kind of thing; you don't add up the individual plans to get the collaborative plan. Individual plans require:
Collaborative plans, on the other hand, require:
Grosz also pointed out that there are problems that one robot can't solve but that two collaborating robots can.The Obligatory Information-Superhighway Mention
Janitorial Combat
Carnage and Leather