快猫短视频

WATCH a troop of ants as they home in on a few drops of sugar solution.
First, a single insect takes a sip and wanders off. Then, in next to no time,
dozens begin to file in neat lines back and forth, following what seems like the
shortest route between nest and food. It looks as though some high intelligence
is at work here鈥攜et ants have only a few hundred neurons to help them work
out what to do next. In fact, they don鈥檛 plan, they just react.

The ants鈥� success as foragers stems in large part from using the world as a
prompt. When an ant stumbles across food, it doesn鈥檛 remember where it is, it
lays scent trails to and from it. To find food, other ants follow these trails.
At first, ants choose between a long and short path at random, but because more
ants travel the shorter path in a given time, the scent builds up more quickly
here. This becomes the favoured path. This method of using the world as a memory
bank is called stigmergy and design engineers in telephone companies are
becoming increasingly excited about it.

Telephone networks today are centrally controlled. Vast programs oversee
everything from how calls are routed through networks to who should be billed.
This all worked well when networks were small and traffic light. But in recent
years traffic has grown so dramatically that some networks are starting to creak
under the strain. What telecoms designers need is a way to make networks
self-sufficient, so they no longer need an all-seeing controller.

A number of these designers now believe the solution lies in artificial
鈥渁nts鈥�, tiny programs that roam around networks behaving much like their living
counterparts. They foresee networks controlled by ants that are capable of
breeding to create customised creatures for specific tasks. But these plans are
not without dangers. The designers are wrestling with problems such as how to
stop the ants running amok on other companies鈥� networks, and how to stop them
evolving into destructive 鈥渃reatures鈥�.

In Britain, BT鈥檚 network connects 24 million of the country鈥檚 telephones. The
ultimate controller of this giant web is the Customer Support System (CSS), a
vast program made up of 30 smaller control systems. As of August 1995鈥攖he
last time that BT released figures鈥攅ach of the 30 programs took up an
average of 350 gigabytes of memory. The CSS does everything from overseeing the
smooth running of the network鈥� redirecting traffic around broken cables,
for example鈥攖o dealing every day with 34 000 repairs and printing 433 000
bills.

At present, it still works well, but the centralised nature of the CSS is
rapidly becoming a liability. Much of its time is spent just checking that all
the elements of the network are working. It must also be constantly updated as
new subscribers, new services and new problems emerge. As it gets older it
becomes harder to adapt, and a failure at the centre could have potentially
disastrous effects across the whole network.

It鈥檚 also not just the high absolute level of calls that presents phone
companies with problems: surges in traffic are also bad news. It only needs a
delay at an airport, an accident on a motorway or a phone-in competition on the
TV, and people can find that the numbers they call are permanently engaged. The
first step away from centralised control of phone networks is likely to come in
dealing with surprises such as these traffic surges.

Two researchers at BT鈥檚 laboratories at Martlesham in Suffolk, Stephen
Appleby and Simon Prentice, have devised a system based on software ants that
reroutes calls away from heavily congested nodes鈥攖he computers that switch
calls from one branch of a network to another. The ants have a degree of control
over these computers and can, for example, instruct nodes to send them to other
nodes on the network.

To begin with, a large program rather like a queen ant wanders randomly from
node to node, measuring the traffic at each point. When it comes across heavy
congestion, it moves to the node that is generating most of the traffic, called
the source node, and spawns a smaller program鈥攁 worker ant.

Each node has a 鈥渞outing table鈥� that lists which of its neighbouring nodes a
call should be sent to next in order for it to reach its destination. It also
keeps a record of how much spare capacity exists on the links to its neighbours.
Starting at the source node, the worker visits each node on the network,
searching for routes with most spare capacity. As it goes, it updates each
node鈥檚 routing table so that any phone traffic to the source will go by the
least congested route. In effect, it leaves a trail for future phone calls to
follow.

Simulations on a mock-up of the digital network at the heart of BT鈥檚 national
phone system show that when a number of worker ants are launched in a congested
area, they stop nodes becoming overloaded and divert calls around the
congestion.

BT is not alone in seeing ants as the way forward. Telecoms companies around
the world are looking at distributed systems for managing their networks. At the
Bristol research laboratories of Hewlett-Packard, Ruud Schoonderwoerd and Janet
Bruten are working with Owen Holland from the University of the West of England
on a different species of ant that will inhabit a different environment.

Schoonderwoerd and colleagues have modelled their software ants more closely
on their insect cousins than BT鈥檚 creations. Their ants are created constantly
at different nodes around the network and journey to random destinations, where
they die. As they travel the network, they leave a 鈥渟cent鈥� trail鈥攐r
鈥減heromone鈥� trail鈥攂ehind them for phone calls to follow. In
Schoonderwoerd鈥檚 network, the routing tables have been replaced with pheromone
tables which list 鈥渟cores鈥� that the ants use to navigate.

The nodes are like junctions on a motorway network. A navigator wants to know
the next junction to head for in order to reach a certain destination junction.
A node鈥檚 pheromone tables do just this job. They score each neighbouring node
according to whether it should be the next node on the route to any other node
on the network
(see Diagram).
Ants always follow the highest-scoring route.

As an ant passes through a node it increases the score in the pheromone table
for its source node. The amount the ant adds to the score reduces as it gets
older, so the whole system responds more strongly to ants that find
shorter鈥攁nd so quicker鈥� routes. Also, the system delays ants that
pass through congested nodes thereby biasing the system in favour of ants that
find uncongested routes.

Evolving ants

Telephone calls also follow routes with the highest scores, so they tend to
follow the short, traffic-free routes created by the ants. Simulations, again on
the model of BT鈥檚 digital network, show that Schoonderwoerd鈥檚 ants avoid some of
the problems that befall BT鈥檚 ants, which occasionally set up circular
routes.

Yet another species of ant has been designed by Marco Dorigo at the Free
University of Brussels. His ants are designed to work on data networks and use a
trail system similar to BT鈥檚 and Schoonderwoerd鈥檚. These networks break messages
into tiny packets of data, and Dorigo鈥檚 ants are attached to the packets and
drag them across the network.

Ian Hawker, a senior engineer at BT鈥檚 research labs at Martlesham in Suffolk,
reckons that despite all the research in Europe, the first appearance of ants on
a commercial network will be in the US. Competition between American telecoms
companies is fierce, and the edge that ants are likely to bring will be more
keenly sought there, he says. In the US, MCI has extended BT鈥檚 ant research, and
is planning to use ants for much more than just controlling congestion.

Rama Nune, a senior network designer at MCI, says the ants could eventually
handle most of the day-to-day tasks of running a telephone
network鈥攅verything from managing the flow of traffic to calculating
everyone鈥檚 phone bill. 鈥淲e are trying to exploit the autonomous intelligence of
the ants and the distributed network of information they use,鈥� says Nune.

MCI also wants to let the ants evolve. All ants implement an algorithm for
routing information. But it is possible that the existing algorithms could be
improved upon. One way to find better solutions is to let, say, a network
management ant 鈥渂reed鈥� with a billing ant to produce a hybrid. In order to
breed, the ants swap small segments of their programming code. Each ant will
have a built-in method for judging how well it does its allotted task, using a
measure known as a fitness function. By killing off unfit ants and allowing the
fittest to carry on breeding, it may be possible to produce ants that do their
jobs better than any human-designed ant.

These breeding programs鈥攃alled genetic algorithms鈥攁re well known
in research and industry. But they are not to everyone鈥檚 taste. Schoonderwoerd
says that with his system there was no need for such methods because his ants
are relatively simple. 鈥淭here was no reason to evolve the ants because that
would have just complicated things,鈥� he says. It is far better to just tweak the
ants鈥� software to improve their performance, he says. Dorigo thinks evolving
entire ants could do fundamental damage, such as destroying their ability to
interact with software at the nodes. But evolving just the algorithms that the
ants use does make sense, he says.

The ultimate objective of MCI is truly far out. According to Nune, its aim is
to leave control of its network solely to ants, which would find solutions to
problems without human intervention. The implications of this are profound
because nobody would ever be able to say for certain how the network
operated.

If everything works well, this need not be a problem. But the strategy
certainly has other pitfalls. Probably the hardest step for any ant-minded
telecoms company is the very first one鈥攔eleasing the ants onto its
network. Once taken, there can be no going back. At present, if a modification
to the CSS doesn鈥檛 work, BT can remove the offending software and roll back the
mammoth program to a point where it knows that most of the network will
function. But if ants have been working on a network for even a few weeks, they
may well have rewritten routing or pheromone tables across the network.
Reverting to the earlier settings may be impossible or would take too long.

Stifling creativity

Another knotty problem is that for genetic algorithms to work well, they need
to be given free rein, and this raises its own dilemmas. In order to keep the
telephone network running efficiently, evolution of the ants would have to be
constrained by ensuring that all ants had a high fitness score. But if the
fitness functions are applied too tightly then any creativity would be stifled
and evolution would cease.

On the other hand, if ant fitness levels are reduced too much the network
could grind to a halt and the ants might evolve into nasty creatures with
unacceptable behaviours. A compounding factor, says Schoonderwoerd, is that it
would be very difficult to dream up a fitness function that reflected both the
best local strategy for an ant and what is best for the network as a whole.

Then there is the problem of how to constrain the ants to a single company鈥檚
network. 鈥淗aving these techniques going across to our competitors is not one of
the main intents,鈥� says Nune. But with the ants deciding their own destiny, MCI
may not be able to stop such an exodus. Alistair Kelman, a leading computer
lawyer based in London, thinks of ants as similar to computer viruses: after
all, they both travel around and change the internal workings of computers.
Kelman doubts that these creatures will stay benign: they will eventually damage
the network. Companies such as MCI and BT need to think carefully about this.
鈥淚f you are creating an evolving product and you are a corporation letting these
things loose,鈥� he warns, 鈥測ou are liable for the consequences.鈥�

The danger of these things getting out of control has not been missed by one
of the pioneers of artificial life, Tom Ray of the Advanced Telecommunications
Research Institute in Kyoto. His Tierra system was one of the first to evolve
novel species of computer organisms
(鈥淎 life in silicon鈥�, 快猫短视频, 15 June 1996, p 32).
He was so worried about containment that he walled up his
creations within a virtual world created inside his computer. 鈥淚magine the
problems that could arise if evolving digital organisms were to colonise the
computers connected to the major networks,鈥� he says.

Ray is now wrestling with a version of Tierra that will run on the Net, a
task that has already been delayed by complaints from the administrators of
computer systems connected to the Net. Like the original Tierra, the networked
version works in a virtual world so it should be easy to stop at any time. But
the same may not be true of ants on a real network. If people tried to stop
them, it鈥檚 possible that they would evolve mechanisms to escape the attacks.
鈥淓volution remains a self-interested process,鈥� says Ray, 鈥渁nd even the interests
of confined digital organisms may conflict with our own.鈥�