How AI speeds up LightsOutPlanning: 3 quick learnings

LightsOutPlanning, or LOP for short, is a supply chain 4.0 solution that we develop in co-creation with our AI-partner Genzai. During our LOP event last December, our customers Bridgestone, Tereos and Alpro shared the insights they gained after exposing their companies to LOP. Next question: how can we find out more quickly what input parameters mattered most to get the best EBITDA results? This is how we tame the beast AI.

3 quick learnings

Acceleration: AI algorithms can accelerate (from weeks down to hours) labour-intensive scenario analyses.

A critical human mindset remains vital: The measured accuracy of an algorithm can be used to determine if you have sufficient data in order to apply your learnings, or if you need to continue digging deeper before deriving conclusions.

Cascade: The cascaded approach, where man and machine take alternating turns, leads to faster results which are also recognised and trusted by our engineers. The AI algorithms have assisted the bluecrux engineers in achieving more accurate and faster manufacturing and distribution footprint results for its customers.

What input parameters matter most?

Nowadays supply chains use more than 1000 supply chain nodes with different input parameters and produces more than 5000 output parameters with EBITDA as a key output to analyze each scenario.

In order to investigate the potential of AI algorithms to accelerate the scenario analysis phase, bluecrux asked Genzai to apply artificial intelligence solutions. With these amounts of scenarios, the main objective was to find out more quickly what input parameters mattered most to get the best EBITDA results of the model.

Step 1: 5 regression models

How do the underlying data structures work? And how are the data for Machine Learning and regression modelling prepared? In Microsoft Azure Machine Learning, we tested 5 different regression models:

– Linear
– Bayesian
– Neural Network
– Decision Forest
– Boosted Decision Tree.

Using 90% of the scenarios, we trained the 5 models to predict EBITDA, and compared the accuracy of the predictions with the final 10% of scenarios.

Next to predicting the EBITDA outcome, we focused on the reliability and accuracy of the prediction: with 600 inputs and 2,000 scenarios, do we have sufficient data to achieve a high predictability of the outcome?

> Boosted Decision Tree performs best

During the first iteration, the Boosted Decision Tree algorithm resulted in the best outcome, and we could determine the ranking of 600 inputs based on their influence.

Although the ranking top 10 was immediately recognised as very relevant by the bluecrux network optimising engineers, the evaluated accuracy of the prediction was low: 0.43. We had too few scenarios, or too many inputs, to make reliable predictions. From their experience, bluecrux engineers were missing some important inputs that the model did not detect.

Step 2: Cascaded approach

To improve reliability, we decided on a cascaded approach. The 100 highest ranking inputs of the first iteration were used to create an additional 2,000 scenarios, with the idea to increase the number of scenarios and reduce the number of inputs to be investigated. The second iteration showed scenarios with less spread, indicating a denser set of inputs. The additional scenarios were fed into the trained Boosted Decision Tree algorithm.

The final accuracy of the EBITDA prediction improved to 0.84, where the model created a new ranking of most important inputs. The final 20 best inputs are now used by bluecrux engineers to complete the optimization of the supply chain network for its customer.