Perspectives on Agent Standards

       Munindar P. Singh, Ebrahim H. Mamdani, Afsaneh Haddadi,
		 Charles J. Petrie, Michael N. Huhns

========================================================================
Agents and multiagent systems are garnering much attention these days,
especially for applications in large, open, information-rich
environments.  Modern applications presuppose the interoperation of
agents with each other and with the underlying infrastructure.  This
leads naturally to the matter of standards.

Although most engineers agree on the importance of standards, the
nature of those standards and their creation and evolution are
often hotly debated.  Standards for agents are no exception.  The
1998 International Conference on Multiagent Systems included a
panel on standards.  Much of the debate centered around the
Foundation for Intelligent Physical Agents' efforts in
standardizing an agent communication language (ACL).  FIPA---the
only serious standardization effort for agents---is a private but
open standards body similar to the ATM forum and MPEG.  Such
bodies have become important in fast-developing technology areas,
where traditional national and international bodies are too slow
to respond.

ACLs involve high-level communication, which fundamentally
differentiates agents from other kinds of software.  For example,
agents communicate in far more sophisticated ways than objects.
For instance, in contrast to objects, agents can say "No!"
Accordingly, ACLs are among the most important of the agent
technologies to be standardized.  Unfortunately, lots of ACLs
exist.  Until FIPA, the only attempted standard ACL was the
knowledge query and manipulation language (KQML), which however
diverged into several non-interoperable dialects.

========================================================================

MAMDANI:

FIPA makes public all its deliberations and interim results.  It
welcomes outside comments and allows participation by non-members
who comment.  In fact, whether standards are de facto or de jure,
a committee would be involved.  The only difference is whether
committees work behind closed doors or in the open.  De facto
standards are thus imposed on the commercial world.  They entail
that products become visible before the specifications, whereas
with the open process, the specifications are agreed upon before
products are created.

FIPA is studying potential products (in telecommunication
management, multimedia and broadcasting, personal assistance, and
manufacturing) before abstracting their common functionalities to
standardize.  Since it takes years for concepts to lead to mature
products, leaving the standardization to when the products are
mature can lead to competing standards.  By standardizing early,
FIPA hopes to catalyze product development by reducing the risks
that member companies must bear.  FIPA promotes important
principles including (a) developing standards well ahead of
maturity, (b) on a strict schedule of releases, (c) to specify
generic tools, (d) that serve a variety of products, and (e)
support verification.

========================================================================

HADDADI:

In principle, standards committees are necessary, but it is not
clear whether such efforts are worthwhile and at what stage
committees should be constituted.  To be assured of success in
standardizing agents and leading the way towards applying agent
technology, FIPA must address two major criticisms.

First, the FIPA ACL specifications are inconsistent (e.g., their
normative and formal semantics do not agree), unnecessary (e.g.,
based on extraneous concepts, such as beliefs, desires, and
intentions), or incomplete (e.g., no specification of protocols
for verification).  This is partly because of the immaturity of
the technology for which FIPA seeks standardization, and partly
because of insufficient effort being dedicated to the
standardization.

Second, the specifications set out so far have already been
considered as standards, permitting no review or modifications.
Unfortunately, these specifications are not suitable for practical
use.  Unless the specifications are thoroughly reviewed and
modified, not only will FIPA fail, but there may be damaging
consequences to the acceptance of agent technology in industry.

========================================================================

PETRIE:

"Premature optimization is the root of all evil." - Donald Knuth

This quote fits not only KQML but also the newer FIPA.  Both were
driven by standards bodies more than grassroots applications, so
there is no consensus adherence to a specification.  Contrast
these with TCP/IP, SMTP, and HTML, which were developed by small
groups.  Their success as de facto standards derived from the
powerful applications that they supported.  Committee-based
standards are not typically so fortunate.

KQML had a good start, in that it was originally motivated by the
Stanford PACT experiment.  But the experiment was never redone to
evaluate the language, and the practical aspects were
deemphasized.  For example, nothing like a TRACE function for
updating one agent on changes by another was ever considered a
primitive.  Instead, a complicated combination of other
(ill-defined) primitives is used.  And there was never any
standardization for connecting to an agent name service.

Despite doing much better in having some application models, FIPA
has also largely ignored an analysis of the several KQML
applications.  In fact, it is unclear why one should choose FIPA
over KQML if the lessons of the latter are largely ignored.

Without attention to commonly used primitives, nor standards for
interconnecting, neither KQML nor FIPA enable any
independently-developed "foreign agents" to interoperate, a common
dream of the agents community.  Much less has either resulted in
widely-used tools to build applications, although KQML has moved
much further in that direction in the last two years.

But it is getting rather late for committee-based agent protocols.
CORBA and XML will now likely take over the Internet niches that
agents might have occupied.  The agents community needs to field
useful tools and applications immediately if it is not to be
ignored.

========================================================================

HUHNS:

As early as 1990, I organized a discussion on ACLs and urged
the adoption of a speech-act-based standard.  Since then
there has been much practical experience, much rhetoric, no
research breakthroughs, and little progress.  Why?

I believe the research community has concentrated its effort on
two projects with poorly conceived methodologies: KQML and FIPA.
KQML was an attempt by a small DARPA-funded group to start with a
"clean sheet of paper" in defining an ACL.  It paid scant
attention to a decade of earlier work and, as a result, produced a
language with imprecise semantics that did not support many of the
most important agent applications.

FIPA, to its credit, has taken a contrary approach in that it
solicits opinions and contributions widely.  Unfortunately, FIPA's
methodology is to start with an ACL designed for agent-human
communication, and then modify it incrementally via annual
releases of a specification.  At the present time, the semantics
of the language is too strong to be useful, and impossible to
verify.  The resultant specification is converging towards a
useful language, but far too slowly.

Instead, I suggest a methodology based on the development of
open-source software, such as is being used for the linux
operating system, emacs editor, Apache web server, and Netscape
browser.  Each of these software systems is being written and
debugged by thousands of developers, and code releases appear
frequently rather than annually.  As a result, the software is
very robust, and evolves rapidly towards the needs of developers
and users.  This seems clearly beneficial for the development of
an ACL.

========================================================================

Although there is universal agreement on the usefulness of standards,
not only the technical details, but also the methodology for
standardization remain controversial in the agents community.  Some of
the concerns - open source vs. otherwise - apply to all of computing;
others are specific to agents but appear inevitable when you seek
standards in a fast-moving area.

========================================================================