SeSoGEN
Development of a self-learning software for the generation of intelligent storage strategies based on neural networks
A comprehensive study by the auditing and consulting firm PricewaterhouseCoopers and the economic research institute WifOR in Darmstadt shows that labor shortages will continue to increase over the next five to ten years, while the use of new technologies (robotics, Industry 4.0, the use of data analysis) that place new demands on employees will rise. This means that the number of increasingly complex software systems will grow, while the number of employees theoretically capable of understanding and optimally operating these complex systems will decrease.
In addition, configurations in the software must become more dynamic and flexible due to the environment changing faster and faster than is manually feasible for customers. Orders from end customers are being placed more and more dynamically through a wide variety of channels and with ever-increasing time pressure. Moreover, production should be able to react to changing market conditions and the warehouse management system should be able to intercept changes in product composition and in the production process as well as in the purchasing behaviour of the produced goods. In order to ensure that goods can be delivered quickly, a rapid supply of production is important and essential. Through machine learning technology, the system can identify strategies beyond the human horizon.
The aim of the "SeSoGEN" research project is to develop a self-learning software that independently generates intelligent storage strategies on the basis of neural networks. This will logically merge storage and retrieval processes in the warehouse in such a way that a completely new approach to process optimization in logistics will be created. The current state of the art is that specific strategies for storage are selected and fixed manually in the warehouse on the basis of certain item parameters such as height, width, shelf life, etc. This assigns items a sometimes-suboptimal storage location, which can lead to inefficient utilization of the warehouse and the underlying processes. In this project, we are taking a new approach to warehouse process optimization. We achieve this by merging storage and retrieval data - e.g. which items are ordered together - in the neural network. In this way, the network will independently develop previously unrealized strategies to guarantee short paths for picking and optimal warehouse utilization.
First, analysis and evaluation of methodologies and algorithms as well as design of the intelligent storage strategies take place (Working Packages 1 & 2). Subsequently, the intelligent storage strategies are developed (WP3) and a machine learning model for these strategies is implemented (WP4). Afterwards, on the basis of a parameter study on synthetic data, the hyperparameters of the machine learning model are optimized (WP5). At this stage, the model is prepared to be trained on realistic data (WP6). After completion of the training, the artificial intelligence module is implemented in the warehouse management software of the company CIM GmbH at the warehouse of a customer of the company (WP7). Finally, the whole software is tested (WP8) and validated (WP9).