Thursday, 13 September, 2001, 12:28 GMT 13:28 UK
Robot brains become more human
Neural networks are getting closer to real brains
By BBC News Online technology correspondent Mark Ward
Scientists have gone back to the brain in an attempt to produce smarter robots.
Researchers who use artificial neural networks - circuits that mimic brain cells - to control robots usually ignore the biochemistry of the brain.
But scientists from the University of Sussex, UK, have found that by simulating the presence of one key chemical they can enhance the performance of the neural network.
Already these novel neural networks have been used to develop control systems to make robots walk and carry out tricky recognition problems.
Signals and wires
As their name implies, neural networks are electronic circuits modelled on ideas about the way that brain functions.
Many artificial intelligence researchers use them to form control programs for robots because they can "learn" the best way to complete a task based on experience.
Typically, researchers generate hundreds of variations of a control program to see how they perform at the task they want a robot to perform. The best-performing programs are then used to create a new batch of control systems.
Crucially, small changes in the control program are introduced in the hope that, just like natural evolution, a better performer may emerge.
After many changes, a control program can emerge that is a champion at completing the task facing the robot. But a team of scientists led by Dr Phil Husbands at the University of Sussex has found that the simplified networks may be missing a vital part.
"The usual metaphor for artificial neural networks is electrical circuits and nodes connected by wires," said Dr Husbands. "The detail is abstracted away in favour of signals and wires, but that is not the whole story."
Neurophysiologists have recently discovered that Nitric Oxide (NO) plays a vital role in many brain functions, including learning and memory.
In the brain, the gas diffuses through the neural tissue and changes the way that nerves react to signals from other nerves.
"NO changes the properties of other parts of the brain's neural network at all kinds of different time scales," he said. "A lot of the subtle processes you get in biological systems are not there by accident."
Dr Husbands and his colleagues have found a way to mimic the workings of NO in the neural networks they are evolving to create control programs for robots.
Often, it takes many thousands of generations to create control programs that do a good job of completing the task given to a robot. But by mimicking NO, Dr Husbands' team has produced control programs that are much faster and far smaller than usual.
The control programs have been used to create control systems for a bipedal robot and a visual recognition system for a robot that works even when the machine is distracted by flashing lights and shadows.