No longer just the province of specialist sectors,
agent-based computing is changing the way systems interact and how
they are managed.
Agent-based computing has already transformed processes such as
automated financial markets trading, logistics, and industrial
robotics. Now it is moving into the mainstream commercial sector as
more complex systems with many different components are used by a
wider range of businesses.
Organisations that have successfully implemented agent
technologies include DaimlerChrysler, IBM and the Ministry of
Defence.
So what are agent technologies? In essence, they are autonomous
software systems that can decide for themselves what they need to
do. Agents are capable of operating in dynamic and open
environments and often interact with other agents - including both
people and software.
"Agents are a way to manage interactions between different kinds
of computational entities, and to get the right kind of behaviour
out of large-scale distributed systems," says Michael Luck of the
School of Electronics and Computer Science at the University of
Southampton and executive director of the EU-funded AgentLink
action co-ordination programme.
"The idea of grid computing is based on large-scale distributed
computation in support of what are called virtual organisations.
All they need to do is to be able to interact.
"We have built small-scale systems, and we are starting to build
large-scale systems, where the component software entities will
determine what to do. It is about machines joining and leaving
dynamically as they see fit and as the system allows."
Luck argues that the growing complexity of the interactions in
emerging distributed systems means new dynamic techniques need to
be introduced to provide more flexible mediation and
management.
One of the basic ideas of agent-based computing is that there
are multiple agents in the environment which talk to each other,
essentially autonomous software systems that can decide for
themselves what they need to do.
For example, laws, norms, guides for behaviour, even policing
and trust between electronic components, can all help in the
mediation and management of such computational systems.
"We can build these systems, but we have no experience of how to
manage such large-scale, open and dynamic systems," Luck says.
"Management of these systems is concerned with mediating the
interactions of components, whether they are supercomputers or
groups of low-level factory floor devices like sensors and
actuators.
"In human societies we have developed laws, norms, regulations
and systems of policing, but we do not have that in computational
systems. We need computational entities that will do what we do in
the real world.
"We need norms and rules of behaviour within systems, so that if
agents joining and leaving a system do not comply, there must be
some sort of sanction."
Some aspects of agent-based computing seek to capture human
notions such as trust, reputation, dependence, obligations,
permissions, institutions and other social structures in electronic
form.
Luck adds that computational analogues of trust and reputation
need to be developed so that they can make judgments based on past
histories of interactions.
The notion of agent-based computing has been adopted
enthusiastically in the financial trading community, where
autonomous market trading agents are said to outperform human
commodity traders by 7%.
"One example is the Zero Intelligence Plus (Zip) autonomous
adaptive trading agent algorithm developed by Dave Cliff, a
colleague of mine at Southampton University," Luck says.
"Inevitably, machines can monitor stock market movements much
more quickly than humans, and if you can encode the kinds of rules
that you want, then it is not unreasonable to imagine that
computational traders will be able to outperform humans.
"I am surprised that the figure is only 7%. This is based on
experiments we have carried out, but there are robo-trader programs
being used in the market not just to provide information, but to do
actual trading."
Cliff developed the Zip algorithm at HP Labs between 1998 and
2005. It works by calculating the best trading strategy for
continuous double auctions, the trading basis of most financial
markets. Zip traders have the ability to "learn" from their
actions, using simple machine learning rules.
In the manufacturing sector, DaimlerChrysler implemented an
agent-based system on one factory floor to allow individual work
pieces to be directed dynamically around the production area.
The intention was to implement flexible manufacturing to meet
rapidly changing operations targets. The result was claimed to be a
20% increase in productivity.
The military has also muscled in on the act. The Ministry of
Defence has used an agent-based system to model changes in human
behaviour in military environments due to factors such as heat,
fatigue and caffeine.
On the commercial front, Magenta - an Anglo-Russian software
company specialising in the commercial use of multi-agent
technology - has worked with clients in scheduling supply chains,
semantic search, text understanding and document classification,
and pattern recognition.
One of these projects is for Newgistics, which offers
returned-goods management services for the retail, healthcare,
service parts, telecommunications equipment and computing
industries.
Magenta has developed Intelligent Returns Management that
controls both the package and information flow from the
point-of-order, or shipment, to the final destination and creates
visibility in the entire reverse logistics chain.
Mark Hinton, chief technology officer at Magenta, says, "When
you buy from the internet or by mail order and want to send
something back, it is a big problem for retailers and
suppliers.
"Newgistics gathers the returns, and gets them back to the
supplier. We built them an agent-based system to manage that return
supply chain, handling something in the order of 100,000 parcels a
week.
"It needs to be an agent-based solution because it is a very
dynamic situation. You do not know which items are going to be
returned on a day-by-day basis, and the numbers are very high, as
are the costs of dealing with them.
"Any small gains that you can get by having an agent-based
system that can route individual parcels back through the supply
chain, scale up to be significant savings for the retailer."
The alternative is to supply all customers with a postage-paid
return package, but that is very expensive. There is also the added
complexity of dealing with environmental considerations, for
example when returns have to go to landfill.
In traditional object-oriented systems, the software is
controlled by a central thread. "However, agents are more active
than regular objects, and the key difference is that agents are
event-driven," says Hinton.
"As something comes into the system, agents wake up and do
things according to their goals and objectives. We are also
interested in aspects of agent technology as it would apply to the
semantic web, understanding language and text, and doing smarter
searching of data.
"We are talking about recognising patterns in unstructured data
in a way that would have been done traditionally with statistical
analysis techniques, but you can get agents to self-organise and
find that information.
"It also has elements of data mining, and can find things that
are obvious, like correlations between spending on certain types of
food and alcohol, for example, or the fact that people who rent
DVDs then go and buy take-away meals, but it can also find things
that are not so obvious," Hinton says.
Case study: better performance is grass roots'
motivation
The Grass Roots Group is an organisation that helps clients
design programmes that motivate employees, customers and partners,
and has deployed ASG's Tevista Performance Manager (TPM) to monitor
network performance across its web and call centre-based
services.
The company's websites are run on Microsoft Internet Information
Server across a number of web servers, and run a mixture of SQL and
AS/400 back-end databases.
Grass Roots' incentive participants accumulate rewards before
redeeming them in an online store for goods, services or holiday.
There are also web-based childcare accounts, created by regular
contributions and used to make Bacs payments to childcarers - all
of whom are enrolled with their banking details.
TPM provides what ASG calls "Tevista Synthetic User", an
approach to testing based on agent technology that is able to
measure the response times for specific application
transactions.
This means that services are not only checked for availability,
but the performance experienced by a user is measured against
defined service level agreements.
Steve Parkinson, group IS manager at Grass Roots, says, "Before
we had an automated monitoring system, people had to do a lot of
manual checking to make sure services were up and running.
"We needed a tool that could answer questions like: 'Is the
service available? Can I log in? Can I see the database?' You might
get to a website, but if you cannot log in, or if it takes forever,
you are going to have a bad customer experience."
If a network performance metric monitored by one of the agents
exceeds a predefined threshold, network administrators are alerted
automatically. It not only notifies the support team, but also
checks whether anyone has acknowledged the problem.
"You need to layer up the tests to see if the server is
available, if the site is available, if the user can log in and how
long it takes. If it meets those and other certain criteria, then
it passes," Parkinson says.
"Otherwise, if it fails and an alert is raised, the system will
very quickly identify the failure point, and will send out alerts
by e-mail and SMS on a 24x7 basis.
"We run these monitors every two minutes, so we will know of any
issues before the customer does. The solution is designed so that
if the first engineer does not acknowledge the alarm within a
defined period, the alarm is escalated onto the next engineer, and
so on, up to chief executive level."