machine learning

Intelligent and autonomous robots are an essential part of the development of industry 4.0 solutions. They will act as a direct interaction partner to humans together in teams and perform works that are much more complex than today‘s typical tasks for industrial robots. Such robots need to deal with a confusing and unpredictable environment and have to react to unforeseeable events. In order to capture this environment and to plan actions, real-time processing of complex sensor information is necessary. Conventional computer architectures appear to be insufficient for such kind of tasks. To solve this problem, hardware accelerators for robotics based on the dataflow paradigm are developed at the Robotics Innovation Center of the German Research Center for Artificial Intelligence.

To manage the increasing complexity in current hardware design processes, current systems are increasingly designed on abstract layers. While the more rapid development of prototypes is a clear advantage of this paradigm, these designs suffer from being closed up and hard to analyze. There is no simple way to extract a system’s structure from its description anymore. Nevertheless, the designers should get all the information they need during development. The computer is assisting in this process with the observation of its inner self: The simulated hardware is supervised by an artificial intelligence (AI). It learns about a system’s functions while the system itself is running. Dependencies and connections inside this system are retrieved independent from their availability, thus speeding up the development process.

Customers in online shops return at least half of the placed orders. This huge amount of return shipments results in high costs e.g. from a logistic point of view. To predict the return rate based on customer data and order information, machine learning techniques can be applied which are able to learn a powerful model for return prediction based on historical order data. This article introduces a hands-on approach for successfully applying machine learning in real world processes and shows a case study to predict the return shipment probability in an e-commerce scenario.

The self-learning search engine improves the search for relevant information in enterprises by using a profile of the searcher and a context, that has to be chosen for each search request. Case-based reasoning is used to figure out which information sources contain relevant information for the given combination of profile and context. These information sources are further preferred in enterprise search.

Robots should be able to act as flexibly as possible in different environments and contexts of action. In order to realise this goal, learning algorithms are developed which permit learning following nature’s example. If a machine that learns in this way causes damage, the question arises as to who is responsible for it. A grey area between manufacturer and owner responsibility may arise. Starting from criteria of replace ability, a firm suggestion is made as to how this grey area could be handled.