By: Lena Schneider
“Artificial intelligence” is a term that’s on everyone’s lips. But what does it actually mean? What do we mean by “artificial neural networks”? How do we train them? And what benefits could they offer for optimizing safety stock?
According to SAP, artificial intelligence (AI) is an umbrella term for any application where machines perform human-like tasks.
Among the many branches of AI are artificial neural networks (ANNs), which are inspired by the human nervous system and mimic the brain’s ability to process information. These networks consist of artificial neurons, with the neurotransmitters and synapses replaced by mathematical operations. This enables the networks to process information intelligently and solve complex problems without human intervention.
An ANN starts out as an initial neural network, which contains no information. To enable this network to solve concrete problems, you first have to train it. That’s where machine learning comes in, which is a key subfield of AI.
Machine learning is the process of using existing data and algorithms to teach AI systems how to recognize certain patterns and rules and develop appropriate solutions. In short, it is a process of generating artificial knowledge from experience. The quality of the resulting neural network crucially depends on selecting and preparing the training data correctly.
Once trained, the neural network is able to apply the knowledge it has learned to previously unknown problems. This enables it to optimize processes or make reliable predictions.
Using neural networks to optimize safety stocks
How can this technology be used for supply chain management (SCM)?
The clever developers at GIB identified safety stock calculations as a potential application for AI. In the latest GIB SCX release, the beta 2.0 version of S/4 HANA for safety stock optimization with neural networks is already a reality.
The addition of artificial neural networks enables you to perform much more complex safety stock calculations in the SAP system. You can incorporate more influencing variables while also evaluating their relative weight.
But all neural networks start out as naive. To enable them to optimize safety stock, you have to train them step by step.
The first step is selecting the appropriate key figures. To do this, we look at what factors actually influence safety stock. We find that factors such as replenishment time, fluctuations, delivery reliability and demand per day play a key role.
After defining the key figures, we can now start the training. Materials that already have optimal safety stock are ideal for this training. The trainer formulates the task for the neural network and then evaluates the result it puts out. This is possible because we already know the optimal result based on historical data. The algorithm can thus recognize correlations and patterns and learn from them.
Once the learning process is complete, the training expert feeds the neural network problems whose results are not yet known. For complex tasks, the training expert can no longer trace the path that the neural network takes to achieving the optimal result. With the safety stock calculation, it’s usually possible to at least assess the result qualitatively. In principle, though, this is not always the case with AI solutions; one simply has to trust in the ability of the trained ANN.
Now that the neural network is fully trained, we can use the collected knowledge to effectively calculate the safety stock for all materials.
Where is AI headed?
Multiple steps are required until a neural network is fully trained. The prototype developed by GIB is currently being tested on ever more customer systems. The goal here is to find out how much training data is required or which influencing variables are particularly suitable, for example.
Neural networks have already achieved initial success and have shown great potential, and we’re excited to see where the journey will take us. We are constantly developing and improving the processes, with the hope that this idea will lead to sustainable benefits in safety stock calculations in the future.
For more information about AI and to view Marc Bertelmann’s detailed presentation on neural networks, please visit our new knowledge platform.