IT鈥橲 OFTEN dark when they come out. 鈥淔irst tasters鈥� scour far and wide in
search of something new, 鈥渂lue-collar hunters鈥� operate closer to home, often
lured by bright lights, while 鈥減remium wanderers鈥� hunt tirelessly for a mate.
These exotic groups have kept determined researchers up many a long night
studying the vagaries of their behaviour. Curiously, the one thing they have in
common is that their distinctive habits stop as soon as the landlord calls
鈥渢ime鈥�. For these and other groups make up the great British drinking
public.
There was a time when Britons would simply walk to the local pub when they
felt thirsty. But not any longer; today, people are happy to travel farther
afield in search not only of their favourite tipple, but also the right evening
ambience. To cater for these different tastes, brewers have created a range of
branded pubs鈥攆rom Irish bars to student dives, from carveries and country
pubs to trendy singles bars.
This diversification has given the brewers a rare opportunity. Site the right
brand in the right location and the money will roll in. But opportunity never
comes without risk. Their big gamble is to decide the best place for a
particular brand of pub. Or, if they already have premises, they must choose
which brand to turn it into.
Advertisement
Solving these riddles turns out to be so tough that one British
brewer鈥擝ass Taverns 鈥攈as turned for answers to the latest in
artificial intelligence. Bass has a lot riding on choosing the right spot for
its establishments because it owns not just one brand but a range of them,
including Harvester restaurants, Vintage Inns, O鈥橬eill鈥檚 Irish bars and Edwards
肠补蹿茅-产补谤蝉.
Brewers know well enough who visits their different brands of pub, and use a
shorthand for sets of variables that groups of drinkers have in common. These
variables include the earnings of the drinkers, spending habits, their favourite
libations and how far they鈥檙e prepared to travel to taste them. So first
tasters, for example, turn out to be adventurous young, affluent types who
explore country and town for new pubs. They like strong lager and stout.
Blue-collar hunters, on the other hand, are unskilled manual workers who spend
much of their income in a pub with arcade games and a juke box. They prefer
ordinary lager and draught cider. Premium wanderers drink bottled lagers and
stout, tend to be single or newly divorced and are out to enjoy the single
life.
Add to these groups the pint-and-pension brigade, student drinkers, quality
diners, cards and dominoes players and night-clubbers, and you begin to see how
the brewers have carved up their clientele. Once a brewer has a profile of the
groups who frequent a branded pub, the problem of siting a new establishment
boils down to finding places where enough of those groups live.
These days it鈥檚 a cinch to obtain data about people鈥檚 earnings, spending and
domestic situation broken down by geographical area. Every time people join a
club, register to vote, buy something by mail order or sign up for a supermarket
loyalty card, they leave details about themselves all over some organisation鈥檚
database. Data companies buy this information and try to categorise geographical
areas according to common attributes of the inhabitants. These areas can be as
small as the 150 households that make up census enumeration districts or even
postcode units of 25 to 30 households.
Bass buys a range of these demographic data. And the first use it puts them
to is to flesh out the group descriptions of its customers. By asking people in
their pubs for their addresses, for example, the company鈥檚 analysts can search
the data to find out the likely earnings, spending habits and domestic set-up of
those customers.
But the main value of the demographic information is that Bass can search it
area by area for pockets of people who fit their brand profiles. Until recently,
it relied on analysts to trawl through the data, often in the form of paper
print-outs. Scouts would then go out and look at the promising sites鈥攁ll
very slow and labour-intensive.
Smart charts
To streamline the process, Bass first needed a geographical information
system which allows it to plot national demographic data on a computerised map.
Like a paper map, a GIS uses colouring and symbols to represent landmarks, facts
and figures. But unlike a paper map, a GIS is smart. It comes with algorithms
for calculating distance, areas and heights. Data can be linked to coordinates
relating to objects on the map, such as a road, housing estate or town. With the
right algorithms, this allows relationships between data stored at different
locations to be explored taking, say, the distances between them into account.
For example, motorists can buy digitised map data for a GIS which includes
average speeds on roads. Special algorithms can then turn this information into
drive times between two points.
Bass wanted the GIS to locate catchment areas where people meet the profiles
of its brands. For obvious reasons the company won鈥檛 reveal its real
profiles but one might look like this: within a 10-minute drive of the location,
there must be a population of more than 60 000 people, with an average income of
more 拢16 000 a year. Unemployment must be below the national average, less
than 10 per cent should be pint-and-pensioners, and there should be more premium
wanderers than blue-collar hunters. More than 35 per cent of the population must
be under 34, with at least 15 per cent going to the cinema regularly and whose
annual expenditure on eating and drinking out is greater than the national
average.
When the analysts began searching with the GIS, they ran up against a snag.
The system鈥檚 search engine uses 鈥渃risp partitioning鈥濃�攖hat is, each
location studied either passed or failed the search criteria. There was no
in-between. So the researchers had no way to find out if a rejected site nearly
met the search criteria or was way off target. 鈥淲hat we needed,鈥� says Jim
Cameron, manager of strategic planning at Bass, 鈥渨as a much better way of saying
`yes, [the area] failed, but it only just failed, so we could make a judgment
about it鈥�.鈥� So Cameron put out a tender for a search system.
The winning bid came from a London-based company, SearchSpace, working with
an American GIS company called MapInfo. They proposed using a combination of
fuzzy logic and carefully chosen search algorithms that could hunt more quickly
and intelligently through Bass鈥檚 demographic data.
Applying crisp decision-making to our notional profile, a location with a
population of 60 500 within a 10-minute drive would be accepted within the 鈥渟et鈥�
of sites that meet this criterion. A site surrounded by only 59 500 people would
fall into the rejected set even though it might well turn out to be a winner.
Fuzzy logic can deal with terms such as 鈥渏ust above鈥� and 鈥渏ust below鈥�. It blurs
the edges of the accepted and rejected sets and assigns a value that indicates a
site鈥檚 degree of membership to each.
SearchSpace鈥檚 system allows Cameron and his colleagues to control the amount
of blurring at the edges of the sets. They could, for example, stipulate that
the population should not fall below 59 000, but that the search could be more
flexible about how often the filmgoers went to the cinema. This fuzzy approach
allows Bass to identify not only catchment areas that conform to all its
criteria, but also those that, in Cameron鈥檚 view, are worth making a judgment
about.
For a brand with a broad appeal, there are likely to be many candidate
catchment areas and several possible locations within each area. Where a number
of catchment areas are close to one another, the brewer may want to site more
than one pub. But how does it choose, say, the best five sites out of a possible
hundred, taking into account the presence of rival pubs? This is where the fun
really begins.
Choosing a number of sites from a number of potential locations is a
combinatorially explosive problem. If you look for 2 sites out of 100, for
example, there are a 4950 possible combinations. But if you want 5 sites out of
100, the number of possibilities increases to more than 75 million and for 10
sites, there are more a 1000 billion possible combinations.
Greedy solution
SearchSpace worked with Cameron and his colleagues to find algorithms that
could most efficiently home in on the best sites. There are a number of these.
鈥淏ut over the range of all search problems, no one algorithm is best,鈥� says
Feldman. 鈥淚n practice, you know something about the problem so you can choose
the best.鈥�
The scheme they chose begins with a 鈥済reedy algorithm鈥�, a simple-minded
routine used for quick and dirty searches. Say Bass wanted to site pubs on five
out of a 100 possible locations, each one ideally at least 5 kilometres away
from any of the others. The greedy algorithm would begin at one location and
look for the site that best fits the brand profile that was also at least 5
kilometres away, ignoring any in between. From the second chosen site, the
algorithm would cast around again for the best site to fit the brand profile at
least 5 kilometres from the other two, and so on until it had found five
suitable locations.
In some cases, the solution to a search will be straightforward, and the
greedy algorithm will spot it. But in more complicated cases, its value is
limited. The algorithm is incapable, for example, of making trade-offs between
distances between pubs and how well they fit the brand criteria. It might end up
choosing five sites that marginally meet the brand criteria but are all at least
5 kilometres apart, but overlook five far more suitable locations that are just
under 5 kilometres apart. Despite these disadvantages, Feldman says, 鈥淚t is so
quick you might as well use it.鈥�
But for a more intelligent approach, Feldman鈥檚 team next added a genetic
algorithm, which 鈥渆volves鈥� a best solution. The GA starts with a large
population of random combinations of five sites. The GA applies a 鈥渇itness test鈥�
to each combination in turn, checking how well the locations fit the brand
profiles and how nearly they meet the criterion of being 5 kilometres apart.
Combinations that score low are discarded, but the higher scorers are allowed to
鈥渂reed鈥�, by swapping a site or two鈥攋ust as parents鈥� genes are jumbled in
their offspring. Again, just as in nature, the GA can throw random mutations
into the mix by changing a single site in a combination. This new population
then goes through the fitness test again.
This process of breeding, mutating and testing continues until a single best
solution emerges. The GA doesn鈥檛 perform a comprehensive search, but it is very
effective at looking quickly through a large amount of data for complex sets of
criteria.
Bass鈥檚 smart pub locator has dramatically changed the way the company scouts
for new sites. 鈥淪ome of my guys鈥攖he regional development managers out in
the field who went and assessed the sites鈥攕ay they won鈥檛 have to get out
of bed anymore,鈥� says Cameron. Now, if someone offers to sell Bass a pub or
chain of pubs, instead of traipsing round to check them all, the regional
managers can key the pub locations into the system and let it check their
catchment areas against any of Bass鈥檚 brand profiles. The system runs on
portable computers with 2 gigabyte disc drives to store the data鈥攃ompact
enough to use comfortably in bed. 鈥淲e avoid a lot of running around by not
looking at places that we know won鈥檛 fit,鈥� says Cameron.
Bass spent nearly 拢300 000 on its system, including the data. Small
beer in the context of the 拢300 million the brewery invests each year in
its premises. Cameron is now using it on a daily basis for strategic planning as
well as for making tactical decisions on individual sites. SearchSpace,
meanwhile, has developed a generalised version of the system, called X-locate,
which the company says can be used for many kinds of retail location analysis
and planning.
So next time you go out with your fellow blue-collar hunters, premium
wanderers or night-clubbers, don鈥檛 be surprised to find the kind of place you
like just around the corner.
The brewers knew you were coming.