In general, the workshop started out quite disorganized,
basically because so many different points of view were
represented: some people were interested in robots (about 10% only),
some about multimedia diffusion (video on demand) and learning
user's preferences to select appropriate content, some interested
in mobile, general magic style agents, and others
were more in the distributed software agents area.
Since no one clear vision was presented of what an agent is,
much discussion was spent on what would be standardized and
what wouldn't.
The second day was a little more efficient, as questions were prohibited
during the presentations. We were able to see the broad
spectrum of requirements for all the different
participants. There were many very interesting presentations
condensed into a very short time. We will receive (in a few
weeks) slides from all the presentations; I can send you copies
if you like.
Of the presentations I saw, my favorites were :
Donald Steiner, Siemens : their MECCA architecture shares many
qualities with OAA. They are applying the technology to
the domains of office assistant (very similar to OAA!) and to
trip planning.
Albert D. Baker (University of Cincinnati): real world domain using
a broadcast mechanism to make 1000 (!) agents interact.
Nader Azarmi (BT Labs): very interesting negotiation-capable agents
applied to the domain of business process control.
For my brief presentation, I tried to stress the following
points:
1. In my opinion, it is important for FIPA to carefully define
what is an agent and how this differs from other technologies
such as objects. What are the requirements of an agent architecture
vs. a distributed object architecture? Can't we reuse many
common notions (such as security) that standards like CORBA already
address?
2. At SRI, we have one possible definition of what an agent should be,
focusing strongly on distributed, delegated "intelligence" and
control, and on a high-level communication language.
3. We've implemented a number of applications within the framework
(probably more than any other group I saw at FIPA) and with this
experience, we've learned some valuable lessons, one being the
utility of providing hooks for extensible tools to simplify
interfacing with and using agents.
In the third day, some progress was made in defining a list of
requirements, target applications, and standardisation
items to be considered. "Results" can be seen on the FIPA homepage,
at http://www.cselt.stet.it/ufv/leonardo/fipa.htm.
I think that given all the different viewpoints, FIPA seems to be
moving towards creating some sort of standard that will primarily
reflect a distributed software agents perspective. Everything will
become much more concrete during the next meeting in October, which is
being held in Japan: at this time, a call for proposals will be
issued. I think that, if pushed by some tangible applications which
focus the vision a little, FIPA might actually arrive at a defining
a useful set of standards that will allow agent interoperation on
a larger scale than currently available. The intent is to have the
first FIPA standards defined by the end of 1997...
One question is, what is the role of OAA in all this? I believe that
if we choose to, we are in a strong position to influence much of
what FIPA produces as a standard. I think that perhaps with the
Siemens and BT Labs, we really have the closest overall framework
and approach for what FIPA is trying to accomplish. And we are the
only group who is trying to use their framework to implement the
complete range of agent applications, from cooperating robots, to
intelligent avatars, to multimodal and multimedia interfaces,
to information retrieval and management. The release of a distributable
version of OAA-lite should only enhance our bargaining power.
So, that's the report. If you have questions, feel free to ask...
Adam.