I have read a few articles
recently about algorithms. I think someone has written a book about them, and
usually the articles have been scare stories, talking about how dangerous
algorithms are and how they are taking over our world. Well, I disagree. I love
them.
What is an algorithm? The
free online dictionary defines it as a step-by-step problem-solving procedure,
especially an established, recursive computational procedure for solving a
problem in a finite number of steps. I
think we had algorithms before computers, but the word has gained currency in
the last half century as computers have become more prevalent.
Essentially, an algorithm is
a way of modeling complexity. Any problem that has a lot of available data, and
where a model or set of rules can help to make sense of the data can benefit
from an algorithm. As computers have gained power, so algorithms have been able
to become more complex and process more data.
I see a clear parallel
between a computer algorithm and the models, assumptions and stereotypes we all
use to simplify our lives. Should I cross the road here, or there? Our brain
goes through lots of steps very quickly. How tired am I? How far would the
diversion be? Can I trust the signal “walk” signal? How courteous are drivers
around here? Am I wearing visible clothes? Do I have kids with me? How much of
a hurry am I in? Are there any shady looking characters about? The brain will
process all these questions and make a decision for us. In “Rain Man”, the
autistic Dustin Hoffman found this sort of complex processing of uncertainty
too difficult, and made a mess of the challenge. But most of us can judge the
risks and subtle choices quite well. The human brain is immensely powerful.
What has happened is that
computers are gradually catching up with our brains, through sheer force of
data. This is something to celebrate, not fear.
When I look back at my
career, algorithms have played a major part. Working at the boundary of
strategy, marketing and analytics, I have faced many challenges where a
slightly better model of complexity could lead to competitive advantage and
business value. I was lucky to do this at a time when we were facing many of
these challenges from first principles, and I found them very interesting.
My first paid job was for a
company trying to develop software to help logistics companies plan their
depots and truck routes. It was called Vanplan, but our software did not really
deal well with the vast data and complex variables involved so it did not sell
well, and wags in the company rechristened it Truckfuck. Inside Vanplan was a
database of all the main junctions and roads in the UK and their average
speeds. We had to build this from scratch using maps, and paid part-timers to
create this. There was an elegant algorithm, that I tried to improve and run
faster, since the computer power at the time was a major constraint. Then
Vanplan had other related algorithms, which also needed tweaking to make them
reliable and faster and to deal with real-world variables like rush hours and
driver habits. It was fascinating.
It is hard to credit that
this was only thirty years ago. Now Google have coded the same data globally
with far more accuracy, and the equivalent of Vanplan is offered online as a
cheap application, computing in a second what used to take us hours. I call
that progress.
I can quote many other
examples. I developed or interpreted or simply applied many algorithms. I
allocated stations to sales reps, defined sales rep territories, optimized depot
structures, and sought locations for new stations. I tried to model petrol station
pricing, and optimize the location inside stores for categories and items. I
built models of costs and incomes, for my own firm and competitors. I tried to
model an insurance scheme whereby we would guarantee an engine cold-start
performance, based on weather patterns and where trucks typically were parked
overnight. I tried to model the financial impact of alternative advertising
media and campaigns.
All of these required
algorithms. They were imperfect, having to model data and assumptions, and were
limited by the computing power available at the time. It was fun. I was
blessed, in the right place at the right time with the right skills and
experience and motivation.
The only thing that has
happened since is that data has become more widely available and computers have
become immensely more powerful. This means the algorithms are better, more
reliable, and more usable in more areas. Good.
Actuaries always tried to
assess pension risks and returns. Now they can do it more accurately. Banks
always tried to assess loan risks. Years ago it was based on the biases of the
bank manager, now they have better systems and data. The police always used
models to try to catch criminals, choosing Friday nights to hang about outside
pubs. Now they have wonderful profiling and contextual algorithms. Intelligence
agencies always tried to find terrorists and spies, with comical ineptitude.
Now they are rather better, even if still as dishonest. Brokers always tried to
pick stocks, based on dubious knowledge and bias. Now they have data to spare.
Doctors always guessed ailments based on symptoms and test results. They still,
do, but have become much better at it. When I was interviewed for Shell in
1982, I got in based on a combination of science, prejudice and luck. The same
three factors would apply today, but the science would weigh far more heavily.
Finally, traffic planners always tried to position lights and lanes to keep
traffic moving. But recently, on one miraculous morning, I drove from Forest
Hills to the Queensborough bridge through a total of 62 lights, and stop at
just one of them.
This all can make our lives
better. We have GPS, we have wonderful knowledge sources, we can shop around to
get good coverage and prices based on our own habits and characteristics. If we
have nothing to hide, we can feel safer with the smarter policing. We can
expect more medical breakthroughs. We can still use our brains to make even
better personal models and use our own computing power to test them. Algorithms
are good.
So why all the fuss?
Partly we are scared of
things we don’t understand. Partly it is journalists and authors looking for a
story. Partly we are trained by sci-fi to be scared of computers and big
brother. And there are one or two consequences of algorithms that do require
control.
First, the finance world
needs some brake on velocity of trading. If everyone has similar algorithms,
you risk events like the flash crash of 2010. My solution is simple: charge a
Tobin tax on transactions. Built into algorithms, that on its own would reduce
velocity of trading, as well as creating many other benefits. Problem solved.
Controlling big brother is
nothing to do with algorithms, and everything to do with law and transparency.
If the spooks had to own up to what they were doing, they might still do it,
but they would need stronger justifications, and we could exercise some choices
too, such as not using US technology companies.
The final issue that I can
see is excess discrimination by segment. It is good that groups that offer
better risks should have cheaper car or medical insurance, but somehow everyone
must have affordable insurance. It is good that the police can target
neighbourhoods and certain demographics, but it should not be cost-free, and
people need an escape route from a disadvantaged segment.
Most of this is about smart
welfare. Obamacare helps, but is not enough. Imagine someone with a criminal
record and former addiction: their route to health and honesty is hard enough
already, without imposing all the added costs that algorithms would calculate.
So let the providers use the algorithms and charge the higher costs (though
don’t let them refuse coverage) and then make sure the overall welfare package
offers a fair chance and great incentives and enough support, through clinics
for example. Unfortunately, this line of thinking is hardly mainstream in the
USA, even though it would certainly pay for itself.
What about black youths whose
lives are made a misery from police harassment? Well how about the police
having to pay a token amount per search when no suspicious behavior is
discovered? The payments could be dependent on other socially desirable
actions, such as looking for work, no truancy by kids, and up-to-date
maintenance payments by single fathers. That way everyone gets an incentive to
behave better, police included, while the power of the algorithm can still be utilized.
Hopefully, succeeding generations of algorithms will be so smart that more and
more innocent sub-groups can be eliminated from target populations.
So, let us not be scared of
algorithms, let us embrace them. The world would be a less efficient place
without them. But, in a few specific cases, let us think about the possible
consequences, and put humane remedies in place.
No comments:
Post a Comment