How to Use the Potentials of Machine Learning for Your Own Purpose
© putilov_denis - stock.adobe.com
Because of progressing digitization of supply chains, more and more data, which can be evaluated beneficially, is being generated and can lead to valuable findings. In this context the term artificial intelligence (AI) is also often used. As a part of AI, machine learning (ML) used within supply chain management offers high potentials and many advantages.
The possible uses of ML-procedures within supply chain management are multifaced. In almost every field of activity in SCM you can find cases of application: Starting with sales forecast and consulting area information ML is an important part of future work. Further fields of application extend from procurement, e.g. over an ML-based stock- and service level management with dynamic stock disposition, over production with subjects such as e.g. Smart Maintenance to distribution of goods and products, e.g. with consideration in route planning.
The quality of a demand forecast has a significant influence on the quality of following planning processes. As internal data as well as those factors which are possibly influencing the demands can be identified and adequately considered while using external data, there exist an enormous potential in approving supply chains compared to the status quo with help of ML-based methods.
For the most accurate forecast possible of future demand figures, you need different data in different forms and structures as well as actual forecasting methods. Those form the centrepiece of the ML-application and – like the data - need to be determined according to each application case. However, it is possible that in certain use-cases conventional, statistic and therefore non-learning algorithms are more suitable than ML-procedures. The individual requirements for each use-case should therefore always be checked – there is none general-purpose solution!
The various effects of using more or less data as well as using ML-procedures or a statistic procedure, are shown by our ML-demonstrator. We help you to increase your forecasting quality with state-of-the-art-technologies!
On the one hand the prototypical realisation of our machine learning demonstrator shows the feasibility, on the other hand it demonstrates potentials and necessities of data preparation as well as an analysis of demand influencing factors.
Like in conventional, statistical methods of demand planning, the calibration of ML-methods demands methodical knowhow and is a continuous optimization process.
To gain valid findings out of data with an algorithm, a critical amount of data in a proper quality has to be available. We advise a potential evaluation to clarify your data situation.
Only those who understand and know the content of their data can provide meaningful and complete data sets for the ML algorithm – a central requirement for the success of this methods.
There is none general-purpose-method. In fact, an ML-algorithm has to be selected reasonably. Considerable general conditions are for example the demand pattern, the necessary forecast horizon, the available data base and the accepted calculation effort.
For being able to train ML-approaches you need to prepare data with regard to the application case with high time expenditure. Experience shows, that more than half of the total effort goes into data collection, data adjustment and -organising, formation of training data sets etc.
In order to implement an ML-based algorithm successfully, a close exchange between specialists and methodologists is needed. Build an ML-Team to unite the knowledge of sales- and ML-experts.
An intense feature engineering is necessary in order to increase the efficiency of ML-algorithms. This provides a certain preparation of data, so that the algorithm is able to recognise possible correlations between influencing factors (e.g. marketing efforts) and the result quantity (e.g. sales amount). Therefore, make sure that new developments from daily businesses get the ML-teams attention!
The longer a model is trained, the more exact it is able to include characteristics of data sets.
Meanwhile there exist diverse ML-based methods and experience values especially for demand forecast. Share your experiences within this sector with other companies and include external experts.
Companies, customers and products are different from each other. Therefore, you need to analyse which method leads to the best forecast qualities. Sometimes conventional, statistic methods show better results than the latest AI-algorithms.
We provide a holistic view on supply chains and, what’s more, we also have long-term cross-sectoral experiences in implementation of demand forecast and optimization of supply chains apart from artificial intelligence.
Fraunhofer Institute for Material Flow and Logistics IML