Maths boffin cracks money nut

Neural networks solve problems by mimicking the human brain. Like the brain, these complex computerised nerve centres can be taught. But the software is costly and the learning process slow. NeuroPredictor software is priced at £699. And it is fast. The developers say training can take just 20 seconds while for competitor products costing £40,000 […]

Neural networks solve problems by mimicking the human brain. Like the brain, these complex computerised nerve centres can be taught. But the software is costly and the learning process slow.

NeuroPredictor software is priced at £699. And it is fast. The developers say training can take just 20 seconds while for competitor products costing £40,000 to £400,000 it can take half a day to learn the same routine.

Software engineer Paul Weighel wrote the program using mathematical rules developed at Sheffield University. It took him 20 years to perfect.

Funding for NeuroPredictor came from Unishef, the university’s venture capital providers. The program is being sold through startup company SignalBox.

Like the brain, neural networks are a mass of neurons or nerves. Usually there are millions of these relatively simple structures grouped in layers in an architecture known as a multi-layer perceptron (MLP).

Instead of MLP, described by Weighell as ‘a steamroller to crack a money nut’, NeuroPredictor uses bigger, more sophisticated neurons arranged in a pattern called radial basis function (RBF), together with a technique called data clustering.

An RBF neural network needs fewer neurons than an MLP to learn the same new tricks. Training an RBF is faster because there are fewer neurons to teach. Speed is also affected by the way these bigger neurons are arranged.

In an MLP, complex mathematical connections made between neurons, layer by layer, makes training slow.

In contrast, an RBF network does not try to keep a physical copy of the brain, says Weighel. Connections are scalar based on whole numbers, not mathematical like an MLP, making calculations simpler and training faster.

Two Formula One teams are using NeuroPredictor in trackside trials to optimise race-car performance and cut costs. It costs the teams between £25,000 and £50,000 for each lap lasting around 90 seconds.

Within NeuroPredictor is a tool called an orthogonal non-linear regression estimator which analyses data that is time-dependent.

Dubbed the Lag Advisor, race managers are using the software to help them get better lap speeds without the attendant costs. For example, would increasing the down drag from a 2 change of angle on the rear wing make the car go faster than a 0.5 change to the front wing?

Two oil producers, a food manufacturer and a police force use NeuroPredictor for reasons they do not want to publicise.

Lag Advisor is set to become even more powerful as a separate product tailored to specific applications. And it will cost more than the parent program.