Computing
Complexity
THE COUP THAT really car-
ried elevator programming
up a few floors happened in the
1970s, when reprogrammable
computers came on the scene.
If someone had a new elevator
routing strategy, they no lon-
ger needed to sell a mogul on
the idea and wait for a building
to go up. Instead, they could
test and fine-tune their ideas
in software simulations.
A f lurry of new algorithms
hit the shafts. One strategy,
still popular today, is called
“estimated time of arrival con-
trol.” Basically, the computer
considers all cars moving
toward a call and assigns the
one it thinks will get there
the fastest. Another favorite
was to always hand the most
urgent call to the car predicted
to create the best outcome for
that passenger: minimize
journey time, use the least
energy, or whatever else the
designers prioritized.
The apex of the comput-
erized control is destination
dispatch, which you can expe-
rience if you visit skyscrapers
built or modernized since the
1990s. In these buildings,
rather than simply pressing up
or down, you enter what f loor
you want to go to, and it tells
you which elevator will come
to take you there.Because they know exactly
where you’re going, these sys-
tems edge closer to perfect
efficiency. People headed to
the same f loor are bunched
together, turning each eleva-
tor into an express train. As
such, you might then have to
wait longer for a lift, so desti-
nation dispatch systems often
shift their priorities depend-
ing on the time of day. During
the morning rush, when net
capacity is key, you’ll suffer
more of a wait so the system
can reduce overall trip times.
In the afternoon, when fewer
people are riding at the same
time, they can afford to collect
you sooner to reduce the men-
tal anguish of waiting.
With all of these strat-
egy options, engineers are
faced with a new problem of
choosing the best algorithm.
One of the most successful
approaches has been to let
a computer decide. Using
machine learning techniques,
engineers can specify what
success looks like, then let
the elevator controller experi-ment on its own in simulation.
At each moment, the sys-
tem inspects the state of each
simulated elevator and the
parameters of each outstand-
ing request, decides what to
do, and measures the results.
The software eventually learns
a policy for each combination
of factors. With these more
sophisticated policies, even
the people who built the soft-
ware often don’t know why it’s
doing what it’s doing.▲For this destination dispatch system,
enter your floor on the keypad outside
the elevator bank, and it will direct you
to the fastest lift.OPPOSITE PAGE: GETTY IMAGES; JOSHUA WOLFF
September 2019 69