Intelligent Systems Report € August € 1996 € Volume 13 € No. 8

AAAI '96: a report from Portland

By David Blanchard
Editor

Artificial intelligence as a field of research has been around for more than 40 years, and yet it continues to fascinate the public's imagination while attracting new adherents to devote their lives to tackling its mysteries. Perhaps this comment, from one of the 1,500 attendees at the 13th National Conference on AI (AAAI '96), held earlier this month in Portland, Ore., sums it up best: "AI is a wonderful field to get into because it's both highly applied and highly theoretical."

AAAI '96 certainly illustrated both aspects of AI's appeal, highlighting money-making, real-world applications at its Innovative Applications of AI program (see article, p. 7), while providing a forum for the nuts-and-bolts R&D efforts that will go into producing the next generation of intelligent systems. Based on the evidence at hand in Portland, the AI field is in good shape as it heads into the next century.


What works, what doesn't
Frederick Hayes-Roth, CEO of Teknowledge (Palo Alto, Calif.), was chosen to give the keynote presentation on the successes and failures of the AI field, and where it is likely to succeed in the future. His company, Teknowledge, is itself a testimony to the power of persistence since it was one of the original Gang of Four expert system shell vendors which hit it big by going public back in the mid-1980s, then got gobbled up in a poorly-executed merger with American Cimflex in the late 1980s, then virtually disappeared from all radar screens as Cimflex Teknowledge in the early 1990s. Recently, all vestiges of the Cimflex connection have vanished, and the company has returned once more to its expert system roots with its M4 expert system product, a descendant of the original M1.

"The AI field should be able to achieve continuous, incremental progress," Hayes-Roth began. He then proceeded to list some of the proven intelligent technologies, such as representation (languages, domain modeling, knowledge engineering), inference (theorem-proving, heuristic reasoning, matching techniques), control (search algorithms, scheduling, demons, etc.), and problem-solving architectures (rule-based, frame-based, constraint-based, blackboard, object-oriented, etc.).

These technologies, Hayes-Roth pointed out, have led to a number of success stories for the AI field, such as speech recognition systems, autonomous and tele-operated vehicles, various domain-specific planning and scheduling systems, hundreds of case-based assistants, and literally thousands of domain-specific expert systems. However, when you take the time to examine what makes a successful intelligent system, you'll find - as Hayes-Roth did - that "AI research hasn't produced reusable components." In other words, every intelligent system almost has to reinvent the wheel every time out, since these systems typically are expensive to develop, are focused on a specific domain and usually just a few specific tasks within that domain, and are highly customized.

"The scope of knowledge incorporated to date is too small," he explained. Individual projects succeed, and succeed phenomenally well, but these individual projects cannot be combined to compose systems of wider scope and capability. "We must find a way to integrate across tasks," he urged.

The principal strategy for developing the next generation of intelligent systems, according to Hayes-Roth, should emphasize:

• context-adaptable building blocks,
• reusable knowledge,
• high-value component functions,
• composition architecture for multi-task systems,
• architecture for semantic interoperability, and
• a small number of domains, which should be attacked persistently.

Rising to the challenge
Taking its cue from Hayes-Roth's vision of the next generation, a panel of AI experts set forth their own challenge problems for the AI field as a whole to tackle.

For instance, Tom Mitchell with Carnegie Mellon University wants to see programs built that turn the World Wide Web into the world's largest knowledge base. "The challenge is to build programs that can `read' the Web and turn it into, say, a frame-based symbolic representation that mirrors the content of the Web." He also urged, "Let's build agents that exhibit life-long machine learning, rather than machine learning algorithms that learn one thing and then get rebooted." And he'd like to see the field apply machine learning to learn to understand natural language.

MIT's Rodney Brooks, one of the leading artificial life researchers, asked, "Can we build a program which can install itself and run itself on an unknown architecture? How about a program which can probe an unknown architecture from a known machine and reconfigure a version of itself to run on the unknown machine?" And ultimately, "Can we build a system by evolution that is better at a non-trivial task than anything that has been built by hand?"

Nils Nilsson, with Stanford University, asked for a robot that can perform "any task that a human might "reasonably" expect it to be able to perform given its effort/sensor suite. [And] the robot must stay on-the-job and functioning for a year without being sent back to the factory for re-programming."

Eric Horvitz from Microsoft had a challenge for his fellow software developers: Develop "comprehensive autonomous decision-making systems that are situated in dynamic environments over extended periods of time, and that are entrusted with handling varied, complex tasks."

And Thomas Dean from Brown University issued a challenge "to theorists, experimentalists and practitioners alike to raise the level of expectation for collaborative scientific research in planning."


Digital pre-assembly at Boeing
The AAAI show even ventured outside its usual parameters to highlight some advances in virtual reality R&D. Airplane manufacturer Boeing (Seattle, Wash.) has developed FlyThru, a digital pre-assembly system for checking designs. Using FlyThru, a spin-off of a Boeing advanced computing research project, engineers are able to view up to 1500 models in 3-D at high speed. This virtual reality system was first deployed a few years ago to meet the needs of the Boeing 777 aircraft for large-scale product visualization and verification.

According to Bob Abarbanel, with Boeing's Information & Support Services, the digital pre-assembly process has met with very favorable results. "The 777 has had far fewer assembly and systems problems compared to previous airplane programs," Abarbanel reported.
Today, FlyThru is installed on hundreds of workstations on almost every airplane program, according to Abarbanel, and is being used on the Space Station, F22s, AWACS, and other defense projects. "In many ways, FlyThru is a data warehouse supported by advanced tools for analysis," he explained. The system is currently being integrated with knowledge-based engineering geometry generation tools.


Vendor roundup
Since this newsletter debuted back in 1983, we've been tracking the ebbs and flows of the AI industry, and one of the barometers we used in the past was the size of the trade show at the annual AAAI conferences. Once upon a time, when the major hardware vendors participated, the AAAI exhibitions were almost legendary displays of overzealous enthusiasm combined with pure innovation, wrapped up in a slick presentation. Those days ended in the early 1990s when the hardware vendors packed up and moved on; so for the past few years, the AAAI trade show has been a much smaller affair. Nevertheless, it remains one of the most significant venues for exhibitors of proven intelligent products to display their wares.

Following are capsule summaries of the products exhibited by the major AAAI '96 vendors:

• Angoss Software International (Toronto, Ont., Canada) displayed its KnowledgeSeeker IV data mining product, a knowledge discovery and prediction tool.
  
  
• Attar Software (Harvard, Mass.) showed its intelligent data mining tools: Analyzer, which uses probabilistic rule induction, genetic algorithms and neural networks; Profiler, which exploits SQL servers and massively parallel processors; and XpertRule, which combines inductive knowledge acquisition, genetic algorithms, fuzzy logic and rule-based technology.
  
  
• Franz (Berkeley, Calif.) demonstrated the benefits of its dynamic object-oriented software technology, based on Common Lisp.
  
  
• Harlequin (Cambridge, Mass.) debuted its new Adaptive Systems Group, which will provide consulting services in the fields of intelligent agents, modern decision theory, operational research, and Bayesian inferencing.
  
  
• Intelligent Automation (Rockville, Md.) displayed a number of products, including Cybelle, an intelligent agent infrastructure; Rotoscan, a ballistic analysis system; MIDIS, a concurrent engineering tool for designing electronic circuits; and MDSP-200, a digital signal processor for adaptive/neural network control.
  
  
• Isoft (Gif sur Yvette, France) exhibited its intelligent software tools, including AC2, a decision tree-based data mining development environment; ALICE, a data mining tool with advanced reporting features; and ReCall, an integrated case-based reasoning tool.
  
  
• Nomadic Technologies (Mountain View, Calif.) previewed its next-generation robot concept, the Nomad S8, which employs advances in mobility, sensing and multiprocessor networks.
  
  
• Real World Interface/ActivMedia (Jaffrey, N.H.) develops mobile robotic platforms, and exhibited its B14, B21 and Pioneer mobile systems.
  
  
• Triodyne (Niles, Ill.) is commissioned by the U.S. Department of Energy to develop new technology for environmental clean-up, waste management and related areas.
  
  
• WizSoft (Tel Aviv, Israel) exhibited its WizWhy software, which uses data mining techniques to predict outcomes for new cases.