Amazon announced in August 2020 that Garrett van Ryzin would be joining the company’s Supply Chain Optimization Technologies (SCOT) organization
as a distinguished scientist. SCOT is responsible for designing,
building, and operating the Amazon supply chain. SCOT systems manage
inventory for the millions of items on Amazon, compute accurate delivery
expectations for customer orders, and drive meaningful changes to
Amazon’s fulfillment center network so that customers receive their
packages in the most efficient way possible.
Prior to Amazon, van Ryzin was a professor of Operations, Technology and Information Management at Cornell Tech, and previously the Paul M. Montrone Professor of Decision, Risk, and Operations at the Columbia University Graduate School of Business. His university research work has focused on algorithmic pricing, demand modeling, and stochastic optimization.
van
Ryzin was also the head of marketplace optimization at ridesharing
companies Lyft and Uber, where he led teams that developed models for a
variety of functions, such as optimally dispatching drivers to riders,
and developing pricing models and driver pay systems that improve market
efficiency. Interestingly, van Ryzin’s paper that he wrote while
pursuing his PhD at MIT “A Stochastic and Dynamic Vehicle Routing Problem in the Euclidean Plane” imagined a world of on-demand transportation as far back as 1991.
During
his career, van Ryzin’s work on complex revenue management problems has
enabled businesses across diverse industry sectors to get the most out
of their limited capacity. To give just a few examples, van Ryzin’s
research has enabled airlines to make a series of large-scale, dynamic
and sequential decisions to determine the optimal price of a ticket at a
particular moment in time. Retail companies have used similar dynamic
optimization to manage inventory levels and prices for different
products to maximize revenue.
What I find particularly interesting are problems that
move beyond the constraints of optimizing within the system, to actually
redesigning the system itself.
Garrett van Ryzin
However, at Uber and Lyft van Ryzin tackled a new
business environment, where revenue maximization wasn’t the primary
goal. Instead, van Ryzin’s teams focused on optimizing more immediate
metrics that were vital to the very survival of their services: service
reliability, driver productivity, and growth.
For example, having a
sufficient number of idle drivers at any given time is critical to
maintaining throughput in ridesharing services. Surge pricing, a
mechanism that van Ryzin’s team at Uber optimized, maintains an
efficient level of idle drivers and encourages more drivers to get on
the street during peak hours when they are needed the most.
van
Ryzin sees technology-enabled service providers — be it at a ridesharing
company like Lyft or the Fulfilled by Amazon (FBA) service — as
transformational. Only a few decades ago, businesses like these weren’t
viable ways to organize service delivery due to high transaction costs
and lack of real-time information. However, technology has radically
improved information exchange and reduced transaction costs, which
allows independent sellers to sell their products on Amazon much more
efficiently than they could on their own.
In this interview, van
Ryzin spoke about the different facets of market optimization, the
intricacies of making automated decisions at scale, managing system
complexity using approximation and decomposition ideas, and why he
joined Amazon.
Q. What are the different elements of optimization?
I’d like to think of optimization being made up of human, technical and operational elements.
At
a human level, the understanding of behavioral economics is absolutely
critical. You have to create the right incentives for both suppliers and
buyers to drive efficiencies. This is especially important for
companies like Amazon that have many buyers and sellers participating
and a high degree of decentralized activity.
In addition to the
human considerations, you also must develop a deep understanding of the
technical elements of how these marketplaces work – the capabilities and
limitation of the technology – which in turn allows you to gain
insights into what structural changes are possible.
Finally,
building services like Amazon that provide physical goods and services
is a much more complicated endeavor than developing a service for
trading virtual entities like stocks or mutual funds. To give just one
example, at Amazon we are shipping actual, physical goods. This means
the underlying physics of the infrastructure and the different
operational elements are critical. So you must also think about your
service in terms of factors like product weight and size, labor
requirements, storage capacity, inventory levels, and lead times.
From
a scientific perspective, there are several open questions in all three
elements of market optimization. A fundamental one is determining the
best approach to take to develop models to drive efficiency.
One
approach is to develop structural models from first principles. For
example, you could make an assumption that consumers are utility
maximizers, develop a utility function and identify the parameters that
constitute this utility function.
Garrett van Ryzin, Amazon distinguished scientist
You could also take a radically different approach and
build models based only on the underlying data – where you draw
inferences from what the data alone tells you. Here, you’re not worrying
about why something happened. Rather, you can use ideas from machine
learning to estimate and refine predictive models without trying to
understand the underlying mechanics.
What I find particularly
interesting are problems that move beyond the constraints of optimizing
within the system, to actually redesigning the system itself. The ‘Wait
and Save’ feature my group developed at Lyft is a good example. This
product allows riders to opt into waiting for ten to fifteen minutes for
a ride rather than having all rides be on-demand. In exchange for
waiting, riders get a lower price. On the technology side, what we are
doing here is actually changing the product in order to make the
marketplace more efficient. I’ve always found there’s a lot more
leverage in changing a system rather than optimizing within a fixed
system. It’s a lot trickier though because big structural changes often
mean you have to get users comfortable with entirely new products or a
completely new way of using the system.
Q. How do you account for the uncertainty and complexity inherent in large systems?
Approximation
is at the heart of optimization because you can never fully represent
the full complexity of a real-world trading system. For example, if a
consumer places an order on Amazon, you have to make several sequential
decisions with complex interactions. Which fulfillment center should I
take that order from? Should I place the items in the same box or should
I pack them in different boxes? How will fulfilling this order impact
the availability of inventory for the next order that comes in for that
product? And how will it affect the available capacity of my local
delivery assets?
You can develop approximation models by using a
rolling horizon approach. This involves taking a best guess for what the
future entails, and then updating your estimate for the future as and
when you get new information. Or you could do something that’s far more
sophisticated: build simulations of the future, and use sampling
techniques to guide your decisions. You can also utilize reinforcement
learning where you fit value functions to historical actions to arrive
at decisions that are continually refined based on data.
Decomposition
is also an important strategy for dealing with the interconnectedness
of the different elements of the system. In large systems such as
Amazon, everything is related to everything else. Supply affects costs,
which affects pricing, which in turn affects demand, which affects
dispatch, and so on. Ideally, you’d want to arrive at decisions by
taking the whole system into account. However, the size of any
real-world system makes this impossible. Any model you arrive at will be
too complex, and you’d require a large amount of time to compute
anything reasonable.
I’ve always been attracted to the idea of helping drive
innovations to get people the basic, physical necessities that are
essential to how they live.
Garrett van Ryzin
This is where decomposition comes in. You can break the
system down into individual components – such as dispatch models,
pricing models, inventory models and so on. The challenge here is to get
these different models to collaborate. You don’t want scenarios where
they are working at cross purposes with each other. For example, you
don’t want one model trying to get rid of an item and have another model
actively trying to replace it. In cases like these, you can drive
coordination between different models using an internal price or some
other mechanism that’s common to all the models.
These are just
some of the trickiest issues in optimization, and I’m excited to be at
Amazon where a lot of the innovation in these areas is taking place.
Q. Why did you decide to join Amazon?
I’ve
always admired Amazon as a company because of its incredible track
record of innovation across so many areas. I remember shopping at Amazon
when they just sold books. And today, you have Amazon Studios, AWS,
Amazon Devices, Alexa and even Project Kuiper where Amazon is putting up
over 3,000 satellites in space.
Amazon is a company that excels
at understanding economic opportunity and then building products and
services that customers value. I’ve only been here for a few months, but
I can already see how the company’s unique culture helps it be so
successful across so many areas.
I also admire the company’s
long-term perspective. Amazon doesn’t make decisions based on driving
quarter-over-quarter performance. Amazon is willing to stick with ideas
for many years. This appeals to me as a scientist as in my experience,
sticking with the right idea over the long term is essential to making
fundamental breakthroughs.
At SCOT, I’m excited to have the
opportunity to contribute across so many areas, from FBA to last-mile
delivery. Over the last few months, Amazon has helped so many people
across the world get essential items during the pandemic. I’ve always
been attracted to the idea of helping drive innovations to get people
the basic, physical necessities that are essential to how they live.
Deepfake
is basically machine-learning technology that manipulates audio and
video recording to show people things that others never did or said.
They appear ...
=============================================
Deepfake detection tool unveiled by Microsoft
Microsoft has developed a tool to spot deepfakes - computer-manipulated images in which one person's likeness has been used to replace that of another.
The software analyses photos and videos to give a confidence score about whether the material is likely to have been artificially created.
