Most of the time when executives in transportation and logistics cite “machine learning,” it’s little more than a buzzword that evokes vague notions of technology continually improving itself.
That was decidedly not the case with Ryan Rusnak, chief technology officer at Airspace Technologies, a time-critical freight forwarder based in San Diego. In a presentation at FreightWaves LIVE Chicago, Rusnak explained exactly how machine learning works, the problems in time-critical logistics it is best suited to solve, how Airspace has implemented it, and the results the forwarder has achieved.
Rusnak said machine learning is “programming backwards,” that instead of a large code base of complicated if-then logic, machine learning algorithms are dumb and generalized. The algorithm Google uses to recommend apps is similar to the way Netflix recommends movies and the way Airspace routes shipments all over the world; it’s just that the algorithms are trained on very different data sets.
The two basic kinds of neural networks — classification (think of a computer deciding whether a photo depicts a cat or a dog) and prediction (think of a computer estimating house prices based on ZIP code and number of rooms) — can be combined in unique ways. One application for blended neural network functionality is autonomous driving, in which algorithms must identify objects and then predict their behavior based on their identities.
One of the issues with machine learning, naturally, is sourcing large, clean data sets that can feed algorithms.
When Google was programming robot arms to pick up objects of arbitrary sizes and shapes, it tried to accelerate data generation by employing an army of arms in a large room working simultaneously and uploading results to the same database. That was still too slow, so instead Google built a simulation that was run millions of times.
Rusnak said Airspace considered doing something similar to accelerate a routing algorithm by creating a freight forwarder game in which forwarders could select the best routes to make the machine learn faster.
Airspace’s approach to technology started with a very simple premise.
“If we just categorize every single service failure, we can walk down the list and wrap every single error with technology, and make the fastest, safest forwarder that’s ever been,” Rusnak said.
The two hardest problems, though, turned out to be quoting a route and price and then dispatching the right transportation provider (for example, choosing a specific carrier to drive a truckload of freight.)
Reliably choosing the best routes is impossible for humans. The number of two-hop air routes from San Francisco to Boston in the same day is about 10,000, Rusnak said. For more complex international moves, for example a five-hop next-day route, the number of possible routes rises to more than a quadrillion.
“Machines respond immediately,” Rusnak said. “We ask, ‘what’s the fastest way from here to London,’ and we respond in half a second with route and price.”
Dispatching is the other key challenge. Getting an algorithm to select the right driver — the one who is the most likely to make the quoted delivery time — is more complicated than it seems: Distance to pickup as well as carrier attributes like efficiency, equipment and experience at a certain facility all play a role.
Rusnak pointed out that some complex deliveries, for instance a driver delivering a live organ to a hospital, can take an hour or more if the carrier has never been to the facility and does not know which procedures to follow.
Rusnak said Airspace Technologies’ machine learning-powered platform performs all of those tasks constantly, freeing up the human employees to do what they do best — customer service.
“The takeaways are to capture more data and that all the tech exists right now,” Rusnak said. “We use machine learning every single minute, every single day, 365 days a year, and I don’t know where we’d be without it.”
