One of the biggest developing trends in the logistics technology space is the growing application of machine learning in warehousing and transportation. In fact, something of an arms race has developed among technology providers as they try to leverage machine learning to differentiate their applications.
Machine learning is a branch of artificial intelligence. "Learning" occurs when a machine takes an existing data set, observes the accuracy of the output, and updates its own model so that better outputs will occur. Any machine that does this is using machine learning. It doesn't matter if data science methods are used or not. It does not matter if neural networks or some other form of supervised or unsupervised learning technique is being used. From a user's perspective, it's not necessary to get bogged down on the specific technique.
Technology providers are already applying machine learning to many areas of the warehouse. Part of what makes warehousing a suitable application for machine learning is the fact that a warehouse operating environment is constantly in flux, especially in today's direct-to-consumer facilities. These facilities must constantly balance the competing priorities of efficiency and responsiveness. At the same time, there are numerous potential constraints on warehouse operations, and it is difficult to predict under which circumstances a given function or resource may become a constraint on throughput. Predictability becomes especially difficult when a facility dynamically introduces orders into an existing workload. Machine learning's ability to adapt to changing conditions in complex environments means that it can produce insights that would not be possible with traditional software.
For example, Manhattan Associates utilizes machine learning within the Order Streaming component of its warehouse management system (WMS) to determine the amount of time required to complete a certain task in a given set of circumstances. The machine learning algorithm reviews past data including type of task, historic duration, and item characteristics. It then identifies which conditions will affect how long it takes to complete a task. The next time that task is assigned, the system can take those conditions into account when estimating how long it will take to complete the task.
As another example, JDA Software is exploring machine learning within its Luminate Warehouse Tasking application to simulate the correlations between multiple attributes (such as congestion and increasing/decreasing demand for a particular resource) and order processing times.
A conceptual illustration of this concept can be seen in Figure 1. It may be thought that the primary factor affecting order processing time is the distance from the dispatch to the pick point. However, the first chart in Figure 1 shows that the predictive ability of that algorithm (shown by the orange line) is not accurate for some of the picks. When the picks are divided into two subsets based on weight, we can see that the accuracy of the algorithm changes. Machine learning can recognize this degradation and create a new input-output relationship that offers a more robust predictive power. Machine learning may determine that distance to dispatch is the determining factor for items under 100 pounds, but that weight is the determining factor for items over 100 pounds.
Machine learning is also currently used in support of warehouse automation. RightPick, the piece-picking solution from RightHand Robotics, encounters a wide range of items and utilizes machine learning to improve its performance based on the prior experience of its robots. RightPick captures an abundance of data from its autonomous picks such as what the robot saw (camera), what it did (including approach and pick method), and what happened (such as success, failure, or placement). This data then feeds convolutional neural networks that enable the robot to distinguish between adjacent items, which help improve picking accuracy. The solution's software intelligence, driven by machine learning, is enabling the robots to pick 50 percent faster than they did a year prior. This productivity improvement is due to having a higher pick-completion ratio and a shorter pick-attempt time. Knapp, an Austria-based warehouse automation provider, also applies machine learning to the piece-picking process. Machine learning supports Knapp's Pick-it-Easy Robot by identifying item shape and determining the best grip method and ideal grip point.
Machine learning is also becoming increasingly important in transportation management and execution systems. The most notable application is generating a more informed and up-to-date estimated time of arrival (ETA) for shipments. Machine learning is working with real-time visibility solutions to learn more about constraints (such as capacity, regulations, and hours of service) and then using that information to give a much better ETA for shipments to warehouses, stores, and the end customer.
These ETA systems are using a variety of data streams. One emerging data stream involves using Internet of Things (IoT) data from trucks to get a better understanding of driver behavior, such as typical driving speeds and times as well as how they operate in heavily congested areas. Trimble Transportation's True ETA application, for example, takes sensor data from trucks and incorporates hours of service rules to know when, where, and for how long a driver needs to stop. The application also understands that where and when the driver stops will have an impact on the ETA. This is especially true if drivers stop before a major city and will have to endure rush hour traffic once they start driving again.
Other data streams include port data; social, news, events, and weather (SNEW) data; and traffic data. Many TMS companies are partnering with data aggregators such as FourKites, project44, 10-4 Systems, and others to use this data for improved ETAs. This data helps to develop forward-looking transportation plans. JDA is an example of a TMS provider that is bringing in multiple external data sources as part of transportation planning and execution. JDA uses these data streams to better understand potential disruptions in the travel time for shipments. Using machine learning, companies can make more resilient plans that can absorb disruption without making major changes. An example is learning about the downstream effect that a late container at the port has on the overall transportation network and adjusting plans and ETAs accordingly. Most importantly, this information can help companies proactively communicate with customers when a disruption occurs.
Machine learning is playing a role in other aspects of transportation management as well. Companies buy a TMS to achieve freight savings by enabling network simulation and design, load consolidation, lower-cost mode selections, and multi-stop route optimization. Machine learning gives companies the ability to maintain high service levels while achieving these savings. Shippers can learn which carriers meet on-time service levels and which do not, which lanes typically carry more chance for delays, and whether there is an optimal number of stops before shipments become late. Machine learning can aid shippers in better understanding how to drive efficiencies without sacrificing service levels.
Supply chain software companies are in the early stages of learning how to incorporate these technologies into their solutions. The solutions available today will only continue to improve. When a shipper implements a machine-learning solution, its individual solution will improve over time as it accumulates more and more data. Additionally, some supply chain solutions are offered in a many-to-many cloud architecture. These solutions have the ability to improve based upon the data not just of one shipper, but of all the shippers that are using the solution.