TECH TALK
For any readers not quite up to
speed with the ever-evolving world
of IT, herewith some explanations.
• CPU (Central Processing Unit)
– think of it as the main processor
in your laptop. In a laptop, each
processor may have from one to 12
(or more) cores.
• GPU (Graphic Processing Unit)
– usually found in a video card. It
has thousands of small processors
that are not as powerful as CPUs.
However, because of the sheer
volume of the processors, the
GPU can perform many more
computations at once than a CPU.
• ML – machine learning.
• Z – type of HP workstation
designed for data science.
American has said that a
year of data formed the
bedrock of this innovative
model. Going forward, will
that database be added to
every month/year?
“Indeed, we used 12
months of data for the initial
calibration. As more data
becomes available, we will
continue to incorporate this
with our model to ensure that
the decision-making process
We’re trying to
fi nd a way to say
“yes” to more
customers
Maulin Vakil, American Airlines
a process which ensures that we are able to maximise the
opportunity to move critical shipments, such as the PPE,
COVID-19 testing kids, medical supplies and pharma products,
which we are moving over our network today.”
Developing the model
“We sought to do this by developing a model which uses
historical data to enable us to better identify which shipments are
likely to change or not show up, coupled with the introduction of
a new Fair Booking Policy, which is designed to encourage
customers to provide updated booking information as soon as it
is available, rather than updating at the point of tender, for
example.
“In October 2019, American Airlines Cargo implemented
iCargo, the new end-to-end technology platform that provides us
with the tools and the data to be smarter and more agile in the
way we develop new solutions and solve old problems. This
modernisation and technology platform now enables us to move
our own business forward by making these sorts of changes. This
is a new way of thinking, but we believe we have the
responsibility to protect our customers’ interests and allow them
to utilise our capacity to the fullest extent.”
In terms of coming up with a solution, did American approach
experts outside the carrier? There were reports of in-house
development.
“The solution was internally envisioned and prototyped at the
enterprise-wide hackathon event in 2019. Every year American
Airlines holds a hackathon event called Hackwars, where 1,000
innovation-minded team members and over 50 technology
partners spend 24 hours designing and creating solutions to
various challenges. At last year’s event, we discussed this issue
and potential solutions with several outside experts, but
eventually ended up partnering with internal IT and Operations
Research & Advanced Analytics (ORAA) departments to build the
solution in-house. At the time, they were testing a graphics
processing unit (GPU) workstation from Nvidia. The model ORAA
developed uses features derived from text fi elds in the cargo
booking and the GPUs had enough computing power to calibrate
that model.”
Details of the process
How long did this changeover take? Were many people involved?
And cutting over to the new system: did that present any
challenges to American?
“It took a few weeks and a team of nine people to develop the
system that predicts whether freight will be tendered or not and
proactively contacts the customers involved. In parallel, on May
1, 2020, we implemented the Fair Booking Policy to ensure that
we are better able to meet customer needs and mitigate unused
space, especially when demand is critical.
“This policy was formed and implemented in just two months
and was a great showcase of cross-collaboration amongst teams,
all with the same vision in mind of being able to better serve our
customers in the long run.”
continues to optimise, given
the continual evolution of the
environment.
“In all, we used 20 data
fi elds such as weight, volume,
commodity, origin, destination
and so on from half a million
actual bookings to calibrate
our machine learning model
with an open-source machinelearning
package from H2O.
Using the Z-by-HP data science
workstation, which had
Nvidia’s GPUs, we had
calibration runs that were ten
times faster than calibration
on non-GPU machines.
Basically, we are using a very
powerful computer to process
millions of data fi elds and
machine learning software to
make much better decisions
which help us make better use
of our capacity.”
The concept has much to
commend it, without a doubt
– but is American following a
trend or is it trailblazing?
“We are not aware of
anyone else using this
technology. However, noshows
and late cancellations
are an issue for the entire cargo
industry. Machine learning
models are a great way to
obtain and utilise data you
need to implement changes
like these to an existing
business model.
“We expected the system to
pay for itself in a few months,
based on 2019 numbers. The
COVID-19 pandemic and the
resulting large schedule
reductions at the airline have
impacted the payback timeline
slightly. At the same time,
cargo space is now more
limited than ever and having a
system that can help us utilise
this space is critical.”
Machine learning: is this,
then, the way ahead for the
sector? It would seem so.
www.airlogisticsinternational.com August 2020 17
/www.airlogisticsinternational.com