AI AND A-LIFE
It's amazing what you can do with a soldering iron, a few wires and a dynamite theory about neurons. Mark Ward explores Alife in the making
PICTURE THIS. A simple machine that searches its environment, navigates round obstacles and is attracted by light. Such a beast sounds like the latest in artificial life: a robot that behaves just like an animal.
Alife it may be. But its not new and contains no silicon. It is controlled by a handful of valves and was made by a man ahead of his time, William Grey Walter.
Born in 1910 in Kansas City, Missouri, Walter was educated in England and chose to work in neurophysiology at a time when the field was starting to buzz. In Russia, Ivan Pavlov was creating a sensation with his experiments on conditioned reflexes. The results are well known: once a dog has been trained to associate the sound of a bell with food, it will salivate when the bell rings, even when no food is around. What impressed the young Walter, who met Pavlov, was the deftness with which the Russian isolated his experiment from the myriad activities of the brain, letting him study just two stimuli-the food and the bell.
In 1928, neurophysiology received another boost when a German, Hans Berger, invented the electroencephalograph and discovered brain waves. As a young man, Walter visited Berger's lab and found his device surprisingly crude. He was always an inveterate tinkerer and decided to refine the machine. Where Berger saw rhythmic fluctuations at about 10 hertz-alpha waves-Walter's more sensitive EEG identified other rhythms. He named the theta rhythms at about 5 hertz and delta rhythms down as low as 0.5 hertz.
This work established his reputation and in 1939 he moved to the Burden Neurological Institute in Bristol, where he worked until just before his death in 1977.
Walter's interest in Alife grew out of his work in neurophysiology. To unravel the complexities of the brain, he proposed building electronic models. But he also recognised the obstacle posed by the sheer number of neurons. "If the secret of the brain's elaborate performance lies there, in the number of its units, that would be indeed the only road, and that road would be closed," he wrote in his book The Living Brain. The only hope was if the number of cells was not so important as "the richness of their interconnections".
In 1948, Walter built his first robot, a three-wheeled "tortoise", covered by a plastic shell and controlled by just two neurons-a pair of interlinked amplifiers. These amplifiers connected two sensors to two motors. The first sensor, a light-sensitive cell, was fixed to the spindle that steered the single drive wheel, and faced in the same direction as the wheel. One motor steered the machine by turning the spindle, while the other drove the wheel round. The second sensor was a simple contact switch that closed whenever a tortoise's shell bumped into something. This "contact reflex" temporarily tipped one of the amplifiers into oscillation.
From these simple connections grew a wealth of complex behaviours. Normally, the photocell "scanned" round and round while the drive wheel revolved at half speed, sending the tortoise in a series of graceful curves in search of dim lights. Walter called the creature Machina speculatrix because "it explores its environment actively, persistently, systematically as most animals do".
When the machine detected a light, it stopped scanning and raced towards it. But if the light became too bright, a dazzled M. speculatrix began to scan again, turning away from the light. If it hit an object, the contact reflex would switch the machine between its normal and dazzled states, so that it repeatedly backed and turned until it had negotiated the obstacle.
All this was expected. But more interesting things happened when Walter attached lamps to the tortoise's nose. The lamp was normally on, but went off whenever it spotted another light source. When placed in front of a mirror the robot began "flickering, twittering, and jigging like a clumsy Narcissus", wrote Walter. This behaviour, he argued, if seen in an animal, "might be accepted as evidence of some degree of self-awareness".
Walter built M. speculatrix for a very specific purpose, says Owen Holland of the University of the West of England, Bristol, who has restored Walter's robots. "He wanted to prove that rich connections between a small number of brain cells produces very rich behaviour," he says.
The idea of using just a few components to generate complex behaviour has a distinctly modern feel to it. In the late 1980s, Rodney Brooks of the Massachusetts Institute of Technology used the idea to lay the foundations of a field that has since become known as behaviour-based robotics. Earlier "intelligent" robots carried large control programs that decided their every move. Typically, these robots managed only very specific tasks and were stumped if they put a foot wrong.
Brooks threw out this method of "top-down" control in favour of a "bottom-up" approach in which he delegated control to very simple elements. Each leg of his walking robot, Genghis, controlled its own actions using sensors and motors linked by a small amount of processing power. Simply by timing the activities of these processors, the robot could walk and avoid objects or clamber over them.
In formulating his approach, Brooks drew on Walter's work. As a child, Brooks had read Walter's book and built his own versions of the machines described in it. The robots designed by both men are busy, "inquisitive" machines that adapt to the world around them.
The animal-like behaviour of M. speculatrix noted by Walter is another trait singled out by adherents of behaviour-based robotics. The surprise emergence of unpredictable behaviour, such a hallmark of the natural world, is key to their claims that they are progressing towards truly lifelike artificial organisms.
Walter's next robot behaved even more like an animal. Machina docilis could be trained in much the same way as Pavlov had trained his dogs. It was actually M. speculatrix wearing on its back what Walter called the Conditioned Reflex Analogue (CORA). This created a connection between the robot's light reflex or its contact reflex and a third stimuli-a whistle. He trained the machine by blowing the whistle and then, for example, kicking it to trigger the contact reflex. "After five or six beatings, whenever the whistle was blown [M. docilis] turned and backed from an `imagined' obstacle," Walter wrote.
At the heart of CORA was a capacitor connected to both inputs-sound and contact. If a kick followed straight after a whistle, the capacitor charged up until it reached a threshold. At this point it discharged and opened an electronic gate that allowed the whistle to stimulate the same response as kicking the machine. If this conditioning wasn't reinforced, it wore off and CORA shut the gate.
Walter described how his robots helped him to study different aspects of behaviour and the brain. Ironically, while many of his theories have long since been laid to rest, the significance of his robots and the ideas that underpin them are only now being realised.
From New Scientist, 25 July 1998
© Copyright New Scientist, RBI Limited 1999