Machine Learning and Route Optimization: Towards Flawless Deliveries

[REPORT]
Artificial Intelligence in Logistics: 6 applications

Machine Learning is a domain of artificial intelligence that allows algorithms to learn from a set of data in order to improve their ability to solve a problem without being directly programmed for it.

Generally, a learning phase designs a mathematical model from data and then applies the model to the data to be processed. The power of Machine Learning comes from its ability to “continuously learn” from new incoming data, which allows systems designed with such technologies to improve over time.

As you can see, one point that should not be overlooked when talking about Machine Learning is data. By nature, the performance of Machine Learning is very dependent on the amount of data available. This is also the reason why it is often related to Big Data which allows to store and make accessible large volumes of data.

In the case of the Supply Chain, technologies using Machine Learning are becoming increasingly popular. They are already used to solve problems such as demand forecasting, inventory management, robot movement within a warehouse or predictive maintenance of machines (thanks to IoT).

At Kardinal, we use Machine Learning in our route optimization technology to 1) refine our predictions using data collected in the field, 2) identify problems that arise and evaluate their consequences.

How Machine Learning can provide more reliable and efficient routes

Kardinal’s route optimization solution recalculates routes in real time to adapt them to the hazards encountered along the way. Data from the field is collected and compared with so-called explanatory variables to build predictive models of the information we are interested in. As the days go by, the accumulation of data makes it possible to build increasingly precise models and to adapt the next route calculations to make them more and more reliable.

#Example 1: A driver delivers B2B and B2C packages and the delivery times are very different depending on the delivery location:

  • Kardinal’s system collects travel times, arrival and departure times at each point via the drivers’ mobile app.
  • Then, this data is compared to variables that can have an impact on the variable to be predicted (here the time spent at each delivery location): these variables are for example the size of the package, its weight, on which floor is the delivery to be made, the area in which the address is located, the type of customer, and if it’s a recurring customer.
  • Machine Learning algorithms automatically identify patterns resulting in a predictive model of “delivery time”.
  • The algorithms can predict the “delivery time”, which allows to define a time adapted to each delivery location in the next round optimization (instead of applying an identical average value on all locations as it is the case in the standard solutions)
  • Finally, the more field data is collected, the more accurate and reliable the “delivery time” information will be, allowing for continuous improvement of routes over time.

#Example 2: A bulk carrier is experiencing loading and unloading waiting times at each facility:

  • Similarly to the previous use case, the system collects the loading, unloading and waiting times at each facility through the driver’s app.
  • Similarly to the previous use case, the system collects the loading, unloading and waiting times at each facility through the driver’s app.
  • The algorithms learn from these data to identify patterns that allow them to predict that at this or that facility, the waiting time for loading is “N” times longer on Mondays between this and that time.
  • In future route optimizations, the algorithms will take these trends into account to avoid bottlenecks, for example.

Continuous route optimization solution

How Machine Learning can identify problems in the field

Another way of exploiting these learning models is the statistical identification of “outliers”, also known as irregular data. These outliers are detected when there are significant deviations between the data predicted by the algorithm and what is actually done.

For bulk transport, this makes it possible, for example, to detect an unloading facility where trucks always take a long time to deliver. This information is of course often already known by the drivers themselves thanks to their experience in the field, but they are generally not able to:

  • quantify this information (how much extra time is spent? how often does this happen?)
  • estimate the consequences: time loss (and therefore productivity loss for the round)? How will it affect the quality of service for the next customers, etc.?

This information is stored and can be used to engage in discussions with certain customers when problems are detected on a recurring basis. If the system described above does not make it possible to identify the origin of the problems (why it always takes longer to deliver to this customer/this area), it does make it possible to identify the real pain points encountered in the field and to trigger possible reorganizations and operational adjustments.

Conclusion

Machine Learning is a technology that has been on the rise for several years and that should not be confused with mathematical optimization technologies.

Nevertheless, for last mile delivery, the two approaches are quite complementary, because route optimization, by collecting data indirectly, and Machine Learning, by exploiting it, allow for true continuous improvement of the system in the long run.

Whether it’s by improving the optimization itself, or by shedding light on problems that are difficult to identify from data, Machine Learning offers Kardinal users the ability to adapt to underlying changes in their business, in addition to the day-to-day flexibility that optimization provides.