杏吧原创

Understanding Intelligence by Rolf Pfeifer and Christian Scheier, MIT Press, price 拢37.50, ISBN 0262161818

IF YOU want to understand how a machine works, take it apart and put it back together. That鈥檚 not, of course, an option for understanding the brain-you鈥檇 find yourself with nothing but a sticky mess.

So why not do the next best thing: build your own simple creature, a robot, and see how that works. Surely with the power of today鈥檚 computers we can put together something that has limited intelligence, and studying that might give us insight into the workings of a living brain.

Unfortunately, when it comes to building mobile robots that can look after themselves, the classical artificial intelligence notion that 鈥渂rain equals computer鈥� is a non-starter. Sojourner, the rover that flew on the Mars Pathfinder mission, was hailed as a success, yet it did not have the native intelligence of a woodlouse. It needed constant instructions from mission control.

Today, AI experts are advocating a fresh approach, with the buzzwords of 鈥渆mbodiment鈥�, 鈥渟ituatedness鈥� and 鈥渆mergence鈥�. Embodiment means that a creature experiences the world through its body, which is no longer just a prop for the brain. So, for example, legs might become sensors that feed information back to the motors that move them. Situatedness means that a creature experiences the world through its actions. It is not just a spectator. Emergence means that sensible behaviours emerge by tuning the robot鈥檚 interactions with its environment; they cannot be imposed by decree from outside.

A whole new wave of robot building has grown from these ideas. Rolf Pfeifer of the University of Zurich is a leading advocate and, together with colleague Christian Scheier, he has written the first proper guidebook to explain them. More than a textbook, it takes the reader from first principles and introduces case studies from research labs around the world. This would be an ideal place to start for someone wanting to get to grips with this field, and the practical problems.

For those who don鈥檛 want to get their soldering irons out, computer simulations are a good way to test ideas. There are some simple simulated robots to play with on Pfeifer鈥檚 website (www.ifi.unizh.ch/~pfeifer/mitbook), which help to give a feel for how, say, changing a sensor can alter a robot鈥檚 behaviour. You don鈥檛 even need a complex control system. One lovely example of this is Tad McGeer鈥檚 passive dynamic walker. This is a pair of legs which will walk down a gentle slope with no electronic control whatsoever.

A strength of this book is its open acceptance of the difficulties facing researchers and the disagreements between them. Lack of dogmatism makes the authors鈥� case all the more convincing. It looks like behaviour-based robotics is coming of age. Human-level intelligence is still far over the horizon, but woodlice had better mind their backs.