Getting started: AI in freight transportation
Freight transportation, which requires managing fleets of drivers, tractors and trailers to meet the needs of diverse shippers, represents one of the most complex instances of a dynamic resource allocation problem. It involves managing complex resources (such as truck drivers) to respond to the demands of shippers as they arise, while managing the uncertainties of shippers, driver behavior, loading and unloading, and traffic delays. Resources are spatially distributed, with loads that are known anywhere from a week or more into the future, down to last minute requests.
Truckload carriers have been trying to bring analytics to their industry since the industry was deregulated in 1980. Initial efforts used the tools of deterministic optimization, which could not handle the uncertainties and richness of this complex industry. In the early 2000s, machine learning and “artificial intelligence” captured our imagination, but these tools are designed for estimating, predicting, and classifying (not making decisions) and required very large datasets to train these models. Reinforcement learning emerged for making decisions, but is limited to playing games (such as chess and Chinese Go), controlling robots and other “single entity” problems.
In this lecture series, I am going to introduce you to the field of “sequential decision analytics,” which means making decisions over time, under uncertainty. I will illustrate the ideas entirely in the context of freight transportation and logistics, starting with the truckload industry which offers tremendous richness. The analytical framework is completely general, and applies to any sequential decision problem, whether it is picking what price to charge, or how to assign thousands of drivers to loads. In the process, we will build on the tools developed in the field of deterministic optimization, as well as machine learning.
The point of departure from these fields is that while machine learning builds functions (statistical models) to predict or estimate something, we are going to build functions that we call “policies” for making decisions. Policies can be simple (bring in more trailers when the trailer pool falls below 10), or complicated (estimate the probability that we can cover a load offered five days from now).
I have identified four classes of policies that cover every possible way of making decisions, including whatever you are doing now! The key is to find the best policy (or combination of policies) that works well over time, under different types of uncertainty. This instructional series will walk you through the four classes of policies, along with other steps required to make computers intelligent.
Optimal Dynamics is a New York City-based startup that has raised $4.4M to date in order to automate and optimize the logistics industry through the use of High-Dimensional Artificial Intelligence. To find out more please visit: www.optimaldynamics.com
