OF all the topics I write and lecture about, nothing annoys audiences more than the prospect of intelligent machines smarter than us. The mere suggestion seems to be an impossibility for the majority of human minds to accept and it is guaranteed to cause consternation. To avoid causing distress I often refer to auto-pilots, engine management systems and the chess chestnut of Deep Blue versus Garry Kasparov as examples of their being better than us. But this often produces an even stronger reaction the dismissive assertion that they are just controllers and they don't do it our way and so they are not really intelligent.

Because of these problems in getting the message across I have been looking for a more concrete example of machine superiority. I did not have to look hard or far inside my own industry for a useful example. When designing integrated circuits the vertical tracks outnumber the horizontal. That is, the circuit tracks connecting the replicated blocks of transistor logic elements are dominantly vertical. So designers have developed a range of heuristics for the layout, spacing and interconnection over many generations of designs. These heuristics are culled from human experience and are embedded in computer-aided design software and go some way towards automating the layout and interconnection process.

In an unusual design exercise a system required a flip of the track dominance from vertical to horizontal. Without decades of experience in this design domain the team struggled to achieve an efficient solution. Eventually they resorted to a genetic algorithm to attempt repeatedly to find a good solution. This software attempted a design using the base heuristic developed by the human designers for the vertical tracks. After a few cycles the algorithm achieved a chip real estate reduction of 20 per cent compared with the human designers.

What happened next? The humans reverse-engineered the design process and found they could only better the machine design by less than 1 per cent. So humans learn from machine.

A second example: a genetic algorithm was used by telephone network designers to improve their designs by over 20 per cent in terms of the cable plant and terminal equipment required. The starting point was the accumulated experience of the design team embedded in a base algorithm. And the magic was the incredible speed of a computer trying more variations in a few hours than a human team could attempt in 100 years. When completed, the human team unpicked the result to learn the heuristic basis so they could apply it to future designs. But again they could only marginally improve on the machine-based result, and again, humans learn from machine.

It turns out that this approach is also being, or can be, applied to the aerodynamic design of cars and aircraft, and the hydrodynamics of boats and ships. It is becoming an accepted part of the engineering and design tool-set, seen not as a threat, but as just another aid to the extension of human education, training and ability.

Peter Cochrane is BT Head of Research. Opinions expressed in this column are his personal views and should not be taken as reflecting BT policy or intent.

Peter Cochrane holds the Collier Chair for the Public Understanding of Science & Technology at the University of Bristol. His home page is:
http://cochrane.org.uk