ANT-WATCHING could soon become a lot less boring for researchers aiming to
unravel their intricate social behaviour. An image-processing system unveiled at
a conference on software robots in Montreal last week not only keeps tabs on 100
individual ants at a time, but can also figure out what jobs they do in the
colony.
Learning more about ant behaviour is not just useful to zoologists. It could
help computer programmers derive simple rules to allow swarms of robots to
cooperate better, and even help to improve telephone networks
(麻豆传媒, 24 January 1998, p 32).
But watching ants is a tedious and time-consuming business. Researchers (or
more likely their students) can spend weeks observing them in the field, getting
up before dawn each day. Video equipment helps, but computers aren鈥檛 much use.
Even today鈥檚 most sophisticated programs can鈥檛 reveal much about the ants鈥�
behaviour鈥攚hether they are foragers seeking food, for example, or
patrollers protecting their colony.
Advertisement
Tucker Balch, a computer scientist at Carnegie Mellon University in
Pittsburgh, Pennsylvania, is working on movement recognition among groups of
robots. He wondered whether he could extend his research into situations where
hundreds of objects鈥攁nts鈥攁re monitored simultaneously. With
colleagues Zia Khan and Manuela Veloso, Balch has created a video-analysis
system that not only tracks the ants鈥� movements but also classifies each ant
according to three behavioural types: foraging, patrolling or
home-maintenance.
A video camera films the colony from above, feeding 10 images a second into
the computer. The software searches for colour changes between each frame, and
then identifies any movement. It can keep track of up to 100 ants
simultaneously.
The program then has to work out what each ant is doing. To do this, Balch
and his colleagues programmed their computer with everything that鈥檚 already
known about ant movement. For example, foragers move about randomly until they
find food, then they head straight back to the nest. Patrollers, on the other
hand, don鈥檛 go home after finding food.
The team developed a software algorithm that uses this knowledge to work out
the role of each ant on the video. The program picks out the type of behaviour
that best matches the ant鈥檚 movements.
Balch now wants to design a system that can learn to tease out different
types of ant behaviour for itself, building its own knowledge. He hopes to use
his observations of insect social systems to improve robots鈥� ability to learn
each other鈥檚 behaviour鈥攑erhaps making them better at playing team games
like football, where they need to be aware of each other鈥檚 actions on the
pitch.
Deborah Gordon, an expert in ant behaviour at Stanford University in
California, says that automated recognition of ant behaviour is an important
step forward. This summer, she and Balch will test the current system in the
Arizona desert. One day it may do the job better than human observers, says
Balch. 鈥淥ur computer won鈥檛 get tired or bored and miss an important
辞产蝉别谤惫补迟颈辞苍.鈥�
