A popular way of explaining the relationship between three big tech buzzwords—data science, machine learning, and artificial intelligence—has been coined by David Robinson, Chief Data Scientist at DataCamp: data science produces insights; machine learning produces predictions; and artificial intelligence produces actions.
In other words, data science is the human endeavor of combing through data to find patterns, and using those patterns to describe the world, or infer causal relationships, or visualize quantities. If you want to know what the average fuel efficiency of a truck in your fleet is, or the average velocity of your trains, you ask your data scientist. On the other hand, machine learning involves using a set of given data (a training set) to teach an automated algorithm how to recognize patterns on its own. You might feed the text of Shakespeare’s Hamlet and Romeo and Juliet into a training set, and then see if the machine learning algorithm can predict whether an anonymous play was authored by Shakespeare or not. To use a transportation example, you might feed a thousand videos of plastic bags flying through the air into a training set so that an autonomous vehicle could learn the difference between a plastic bag and a pedestrian.
Artificial intelligence is the oldest and hardest to define of these three terms, but one useful distinguishing mark is that AI executes, or at least recommends, actions of its own. Whereas machine learning might simply encounter a new data point and sort it into the appropriate bin, AI can play chess, perform motion planning for robots, choose an optimized route, and construct sentences in natural languages.
TMW Systems, a Trimble company, has assembled a vast data lake of information about truckloads and is working on leveraging that into a training set for predictive algorithms. FreightWaves spoke to TMW’s executive vice president and chief technology officer, Tim Leonard, about his 2017 white paper “Transportation Institution vs. Information” and TMW’s data-driven machine learning projects.
“We’re looking at the business model of full automation,” said Leonard. “I know within trucking that’s kind of far-fetched, but there are specific applications that machine learning components can perform well at. Today large enterprises have a huge set of algorithms, and what we’re looking at is the introduction of algorithms as a service, which sets us up for machine learning,” Leonard added.
Leonard explained that “people around the industry, at large enterprises, have done a great job developing customized algorithms that provide solutions for the specific fleets and lanes they’re working on. But the evolution of machine learning algorithms is different. Machine learning is the introduction of algorithms as a service—it becomes easier for the developers to swipe in, see if it works, make metadata adjustments.” In other words, the adaptability and trainability of machine learning algorithms will let load-matching, route optimization, and demand prediction algorithms be developed by outside vendors who can plug-and-play their products into a specific enterprise’s system. Then the machine algorithm would train from that company’s data and evolve into its role.
“But you can’t just set it up and leave it alone,” Leonard said. “Machine learning is great at making predictions from new data that look like the old data it trained on, but it’s bad at figuring out exceptions, paradigm shifts, or completely novel situations. I like to give the example of Bob, who drives down a certain road in Austin, Texas, every day, and stops and eats pizza on that road. He loves pizza. A device in his vehicle that tracks his movements will notice that he loves pizza, and can make recommendations for him when he’s on the road, somewhere else, looking for something to eat. But the machine algorithm didn’t realize that Bob’s something of a foodie—that he only likes pizza from non-chain restaurants. So it tries to send him to Domino’s and Little Caesar’s: completely useless recommendations for Bob. What we have to do is create a metadata mathematical exception, and go in and scrape data from Bob’s LinkedIn and Facebook feeds, which tells us that another thing that Bob really likes is locally owned businesses. Then we tell the algorithm to overlay that preference onto Bob’s pizza preference to help score the data points—the pizza joints—he encounters on his trip, and better predict which one he’ll actually enjoy eating at.”
But what’s the end game for machine learning and the trucking industry? “One of the spaces where we think machine learning can make a huge difference is last mile,” said Leonard. “We know there’s 187M cars, 52M light trucks, 8.3M straight trucks, and 2.7M Class 8 tractors on the road today. Almost every car and light duty truck has unused cargo space, about 12 cubic feet for a car, and call it 25 cubic feet for a light truck. When you add that unused capacity together, it’s the equivalent of 2.2M straight trucks or 930K tractor trailers—the capacity equivalent of 20 UPS fleets.”
“What if you could match parcel delivery with unused cargo space in personal vehicles, generating pickups and deliveries based on the routes the algorithm knew the driver already wanted to travel?” Leonard asked. “A commuter leaving downtown for the suburbs where she sleeps could generate $20 in shipping fees just by picking a few packages up from a distribution center on her way and dropping them off at her neighbors’ houses. It’s one thing for a team of professional dispatchers and supply chain professionals to route a fleet of a few thousand tractor trailers, but to really exploit all of the unused cargo space in personal vehicles—millions and millions of drivers and routes—you need automated machine learning algorithms powered by vast data resources.”