Intelligent agents will be a vehicle for other AI-related technologies
Steve Schoepke steven_h_schoepke@fanniemae.com
The greatest potential repercussion that could
result from the emergence of intelligent agents in mainstream computing
will be the furnishing of a vehicle for other AI-related technologies to
also emerge as viable application components. The deployment of technologies
in the areas of pattern recognition, neural nets, planning algorithms,
cognition, learning and adaptation may usher in an era of "intelligent
objects" (i.e., non-autonomous agents) being embedded within production-level
applications. This possibility spawns new potential for changes in the
way that applications and their domains will be viewed, developed, and
deployed.
Current Technology: Static and Stalled
Even though there exists an ever-expanding myriad of products in computing,
the current offerings appear to be the rehashing of, and increasingly complex
variations of, currently available concepts and technologies: OOA/OOD/OOP,
internet and cyberspace, distributed processing, network computing, parallel
processing, etc. The improved features and performance of these new products
may enhance the current use of the above technologies, but none really
represent revolutionary principles. This trend may continue its course
for some time, until the laws of diminishing returns places selective pressure
on the field to pursue a change in paradigms and associated product offerings.
Current Development Paradigms: Static and Stalled
In the same manner, current development paradigms have followed the same
path as above, and are also subject to the same limitations. Whether strictly
structured or object-oriented, there exists hard-coded paths of execution
and transition states, high maintenance costs, slow turnaround time of
new versions, static behavior based on concrete and anticipated environmental
conditions, rigid and often catastrophic response to unanticipated conditions,
the constant requirement of manual intervention, etc. The presence of procedural
and if-then-else logic has traditionally provided a catalyst for the advancement
of computing, but eventually may present itself as an obstacle to progress.
Again, limitations and diminishing returns of the paradigms may usher in
something revolutionary.
Inevitability of Intelligent and Flexible Applications
From a metaphorical viewpoint, it's obvious that the evolution of organic
life has progressed from the static and reactive (single-celled organisms)
to the deliberate (complex organisms and their use of tools, ability to
be creative and to solve problems, etc.). It also seems obvious that there
is no reason why the evolution of silicon-based "life" (i.e., a loose definition
for software applications) shouldn't follow the same path. We are merely
in an early phase of this process. Accordingly, there will be a need to
build flexibility of execution and self-adjustment into an application
that can accommodate unforeseen environmental variables. The presence of:
increases in complexity and steepening of learning curves associated with
applications, shortages of talented personnel, increasing demands made
upon application software, increasing expense in time and labor required
to maintain future applications, etc. may all act as catalysts for intelligent
objects to be embedded within applications.
Future Selection Pressures for New Paradigms
As the state of diminishing returns becomes obvious, the pressure for the
deployment of embedded intelligent objects will have to accommodate the
transition from static to adaptive execution of an application. The notion
of hard-coded and anticipated functioning may have to be replaced by notions
of a "path of evolution" of behavior. This would involve the use of a "looser"
specification description that could accommodate the flexibility of adaptation
and planning: describing tendencies and future directions of behavior as
opposed to strict paths of execution. Specifications such as "Processes
backlog of 30/60/90 past due accounts using credit history data, generates
and assigns finance charges" could be replaced by "Manages the resolution
of past due accounts drawing from perceived contributing factors and choosing
the optimal resolution based on iterative experience." The language used
should probably be limiting only in the sense that the tendency or path
of evolution can be identified, but the actual behavior cannot. There will
have to be milestone "checkpoints" and behavioral constraints applied at
various point in the application's lifetime, merely to ensure the prevention
of dangerous or counter-productive behaviors. Eventually, much of the application's
behavior would be self-selecting. Controls can be embedded that delete
undesired behaviors. Along with design, analysis, and behavior patterns,
the notion of "patterns of evolution" in software applications should be
explored.
In the same vein, as the intelligent object evolves in its behavior,
its relationship with other entities in the application may change, as
well as its overall functional role. As a result, a traditionally design
tool may be forced to mutate according to this path of evolution. This
may require the periodic addressing of both the initial and the ongoing
positioning of the intelligent object in the application's design framework.
If such objects do eventually emerge, the E-R diagrams, functional specifications,
state transition diagrams, event traces, object models, etc. will have
to be viewed in a different light and altered in an ongoing manner in order
to accommodate these evolving objects. The notion of spontaneous or unanticipated
change in a current, statically designed application brings forth notions
of system crashes and resume updating. However, in the future, this may
be the norm rather than the exception.
There may exist a parallel change that's required in the mind-set of
the developer and user population. Both groups will have to "step back"
and not only understand the concept of an intelligent object, but also
the implications for the future. Incorporation of BDI, behavior patterns,
and other AI-related capabilities offered by intelligent agents and objects
may have to be rephrased in more mainstream-like terminology: Metaphorical
comparisons can be drawn between beliefs and object states, intentions
and execution paths, etc. Those areas that cannot be framed in such a manner
should be described using as many familiar terms as is possible. This will
enhance the possibilities of acceptance and understanding from an audience
that may not be familiar with AI concepts.
Conclusion
The evolution of organic life itself should be a baseline of how AI-related
technologies should be viewed, integrated into mainstream applications,
accounted for and described in design and analysis documents, and maintained
in a production environment. In other words, how we as humans (with capabilities
of learning and adaptation) would be designed using current and future
techniques and tools.
Position Paper. International
Workshop on Agent-Oriented Information Systems (AOIS'99). See http://www.AOIS.org