Reducing Service Part Costs with a Machine Learning Pooling Model

BrandPost By Babak Farmanesh, Gentry Pate, Mario Cornejo
Nov 26, 2019
AnalyticsBig DataHadoop

Pooling optimization dramatically changes stocking strategy to save $20M

Credit: Dell EMC

Dell Technologies’ Global Service Parts organization (GSP) routinely plans for parts that may need to be replaced if a system fails. Forecasting and planning for rare parts failures is like insurance. We determine the probability of failure and — based on these probabilities and several other factors — we determine where to stock parts in order to avoid impacting customers, while also controlling our costs.

If a customer has bought a same-business-day warranty, we are committed to getting the part to the customer within just hours. So, it is important to make sure we have the right parts in the right location at the right time. Since we have a large network of parts inventories, predicting where the next demand will happen is very challenging.

As an example, in the United States alone, we have over 120 warehouses (Figure 1), with tens of millions of dollars’ worth of inventory. If we were to stock at least one piece of each part across all these warehouses, the cost would be prohibitive.

figure 14 Dell EMC

Figure 1: U.S. same-business-day warehouses zip code coverage

Pooling strategy

In order to be certain that we will have the necessary parts where and when they are needed, we took our existing analytics process a step further and developed a pooling model based on optimization and machine learning techniques. This model helps us to more effectively and efficiently meet the challenge presented by expensive parts with very low failure probability. Instead of stocking each part at each warehouse across the network, the model allows us to place them at a pooled location.

Pooling locations allows us to leverage a larger geographical area with a high density of flights or an enhanced ground transportation network. We can expedite a shipment by considering the best and most feasible transportation mode between customer location and the pooled warehouse (Figures 2 and 3). Essentially, by identifying the right pooling locations, we can balance the expedited transportation cost versus the part investment. The model also considers the risk the business is willing to take.

figure 23 Dell EMC

Figure 2: Modeling best and feasible transportation modes for next flight out (NFO) vs. ground transportation.

figure 32 Dell EMC

Figure 3: Distances from a set of pooled locations to customer zip codes


Taking the next step

At GSP, we established pooling more than five years ago, and it has dramatically changed our stocking strategy, reducing costs in the continental United States by $18M. Based this success, we went on to consider further improvements to the model. By applying additional advanced pooling strategies, we have been able to reduce inventory cost even further.

In the model we have in production today, we have fixed pooling locations. However, a new dynamic pooling model is capable of handling even more information and can make better decisions at a more granular level. As a result, we will be moving from eight fixed pooling locations to dozens of micro-pooling areas that will make us nimbler and — based on our preliminary analysis — we expect to further reduce our stocking strategy by $2M, with no customer impact.

There are approximately  (the number of grains of sand on earth is ~) possible pooling strategies for each part in the U.S. network, and dynamic pooling smartly examines all these scenarios to find the “optimal” strategy within minutes! The main methodology used in dynamic pooling is stochastic integer programming, which finds the optimal pooling locations and inventory levels based on set of probabilistic parameters. As a result, our dynamic pooling model gives us the flexibility to add or remove locations on a weekly basis based on three primary factors:

  1. Price of the parts
  2. Expedited shipment cost
  3. Probability of failure of each part at each location. The probability of failures is an output of a supervised machine learning model that uses regression and classification techniques.

As an example of the influence of price, Figure 4 shows different strategies, assuming a part has the same low demand but different prices. As you can see, when a part is relatively inexpensive —$20 — it is not worthwhile to pay the expedited price. So, we plan locally at each warehouse. On the other hand, when the part cost is $2,000, the model recommends only seven locations from where we can expedite transportation either by ground or next flight out.

figure 42 Dell EMC

Figure 4: Example of the influence of price on number of pooled locations.


The application of optimization, together with machine learning models, helped Dell Technologies’ Global Service Parts organization to reduce inventory without impacting our customers, with a pooling model that has now been in place for more than five years. Constantly looking for opportunities to improve this model has led to even further cost reductions through a new dynamic pooling model capable of handling more information and making better decisions at a more granular level. The next phase will be to extend the model into other regions around the world, which presents a new set of challenges. Fortunately, we will be working with an excellent global partner ecosystem, with expertise in putting AI models such as this into operation.

To learn more