Outdoor advertising, or billboards, one of the oldest advertising media, traditionally viewed itself as a collection of property men banging up posters on gable ends and sending little men out with paste and brushes to post them. It came to computerisation with trepidation, and was dragged into the advertising world, which demanded of it accountability and availability of information, kicking and screaming. But now it is using the very latest neural networking techniques as part of the second phase of a major piece of industry research known as OSCAR, Outdoor Site Classification and Audience Research, which gives vehicular and pedestrian audience information on every one of the 65,000 or so poster sites in Great Britain. Like outdoor advertising, neural networks are not a new phenomenom. They were talked about as long ago as 1943, in a paper enticingly entitled A Logical Calculus of the Ideas Imminent in Nervous Activity. Perhaps that’s why it took nearly 50 years for them really to come back in fashion. The UK Department of Trade & Industry liked the technology so much it sponsored a three-year program, ending this year (CI No 2,720), to raise awareness of neural computing, and reckons it has succeeded. So why has it been chosen for Outdoor Site Classification phase two? For the original OSCAR research project, launched in 1985, NOP Research Group Ltd, an MAI Plc company, visited every poster site owned by members of the industry’s trade body, the Outdoor Advertising Association.
Audience estimate
The company collected large amounts of physical data about the location of the site, such as type of road, number of bus routes, distance to town centre, and about the billboards themselves, such as height from ground, angle to the road, distance from which they were clearly visible. NOP then selected a random, representative sample of 600 sites, at which it carried out actual traffic counts, both for vehicles and pedestrians passing the sites. Another research company, Audits of Great Britain Ltd, wrote a mathematical model, using traditional statistical regression techniques, to take the actual data from the 600 sites, correlate the physical features of the location with the actual traffic counts, and produce an audience estimate for every other site in the country. This model produced what was known as the gross OSCAR score, the total number of people in the vicinity of the site. NOP then applied weighted visibility factors, using the physical characteristics of the billboards, to achieve a net score, or true opportunity to see the poster. NOP has continued to collect data about every new poster site in the country, and has a programme to revisit and reclassify about 10% of sites each year. However, 10 years on, given changes to roads and traffic flows, the original model is now out of date, and NOP was commissioned by the Outdoor Advertising Association to produce a new one. Former NOP employee and now part-time consultant to the company Simon Cooper had been using neural networks for his degree in cognitive science at University of London’s Birkbeck college. Cooper, who wrote application programs for the original OSCAR project, recognised that neural networking techniques could be applied to OSCAR, to estimate traffic counts for every road in the UK. NOP bought in local authority annual average daily traffic flow data, which provides traffic figures for about 10% of all roads in the UK and had key items of location data from its site classification, including distance from main shopping centre, dual carriageway within half a mile, or government-designated primary route. The task was, using this data, to estimate traffic flows for the remaining 90% of roads in the UK. But why use neural networks? Cooper explained that a traditional statistical regression model will produce an algorithm that best fits all the given data. A neural network he said, is both supple and subtle enough to in this case estimate audience by region and area, as if it were producing multiple algorithms. Simply put, the origin
al model would take a road input as a ‘trunk road’ and create an algorithm for all trunk roads in the country, be they in London or Aberdeen, whereas the neural network would learn that a trunk road in Aberdeen had far less traffic than a similar one in London, and adjust its output accordingly. In fact, the network worked out that NOP’s site inspector in Aberdeen was making a subjective judgement about a road being a trunk road, and it corrected her data by flagging to itself, if Aberdeen and trunk road, then adjust the estimate downwards. Unlike a traditional computer program, a neural network learns as it goes along. Often described as an attempt to simulate the human brain, the network has a series of neurons or nodes, which all interconnect. It takes a series of inputs, and ultimately produces an output, the answer, by constantly adjusting weights attributed to each connection of each node, until the desired answer is produced.
By Joanne Wallen
The network first has to be trained, by giving it examples of input, and the correct answers for that input. For example, the trainer tells the network that the required output is 10, and the inputs are 1.5, 2.4, 1.6, 0.7 and 2.3. The network will give each of its connections a weight, and on the first run-through of its training data, may come up with an answer of 15. The network then adjusts each of the weights, up or down, to try to achieve 10. On its next run-through it may output 5, so it now knows it has adjusted at least some weights too far. Each run-through is called an epoch. The trainer will set the network a number of epochs, say 50, and then ask it to report its output, to see how close the network is getting to the desired answer. In between the input layer and the output, a neural network has what is known as hidden layers, and it is through these hidden layers of neurons that it begins to learn. Two neural network experts, Warren Sarle of SAS Institute Inc and Scott E Fahlman of Carnegie Mellon University’s School of Computer Science, in a light-hearted interchange over the Internet, liken training a network to a kangaroo searching for the top of Mount Everest. Everest, they say, is the optimum, but any other really high mountain would be nearly as good. The initial weights of a network are usually chosen randomly, Sarle explains, and he likens this to the kangaroo starting out anywhere in Asia. However, if the trainer knows more about the scales of the inputs, he may get the kangaroo to start near the Himalayas. Sarle and Fahlman then use various kangaroo analogies to explain the different types of algorithms used in neural network calculations. For example, there is the stabilised Newton algorithm where the kangaroo has an altimeter. It tries to jump to the top of the mountain, and if the jump takes it to a lower point, it backs up to where it was and jumps again. Or there is standard back propogation, where the kangaroo is blind and has to feel around on the ground to make a guess about which way is up. If it ever gets near the peak, it may jump back and forth across the peak without ever landing on it, which could be preferable to simulated annealing, where the kangaroo is drunk and hops around randomly for a long time. Fortunately, it gradually sobers up and tends to hop up hill, says Sarle. To build the outdoor network, Cooper used the Stuttgart University network simulator, a powerful commercial network builder, which sets up the original emulation of neurons on the computer, and then had to chose the training algorithm most appropriate to the application, and to decide on the topography, the number of neurons and connections. He describes this process as a black art rather than a science, relying as much as anything on the network builder’s knowledge and experience, hence how close to the base of Everest the kangaroo will start.
Training set
The outdoor network was trained using data for 3,000 poster sites, and tested with a further 3,000. It saw both the input data, 26 items including NOP-collected location da
ta, and the required output, the local authority traffic counts, for the training set. It never saw the data for the second 3,000 sites, which was used to test what it had learned. It has two hidden layers, within which it transforms the input information, and learns such things as the difference between trunk roads in Aberdeen and London. Cooper trained the network on a 66MHz 80486 personal computer from AT&T Corp, but once trained, it will run on as little as a 16MHz 80386 machine. Although trained using UK traffic flow data, its neurons and connections are now in place, so it can be quickly re-trained on traffic data from any country, and NOP has already had interest from Spain and Germany. OSCAR 2 is due to be launched in October. Neural networking has produced audience estimates for every road in the country, and therefore for every one of around 65,000 poster sites, with a correlation or accuracy factor of 0.85 on a scale of 0 to 1, compared with the best expected from traditional regressional techniques of 0.56. So is neural networking better than traditional alternatives? Paul Harris, director and senior statistician at NOP says this is still an open argument, especially since even when they succeed, neural networks still tell us little about why or how they have done it. However he believes that in its predictions of road traffic flows, this neural network has definitely outperformed traditional techniques.
