Machine Learning
Machine learning (ML) is a subfield of artificial intelligence. ML enables IT systems to independently recognise patterns and regularities (algorithms) on the basis of existing data and examples of results and to develop solutions. The knowledge gained can be generalised and used for the analysis of previously unknown data. Depending on the application, different learning categories of ML are applied. These are: supervised learning, unsupervised learning and reinforcement learning. The type of ML most similar to human learning is reinforcement learning. Reinforcement learning is based on reward and punishment.
Areas of machine learning (ML) presented in an overview© Eigene Darstellung
Examples of ML learning categories
Unsupervised learning
An easy-to-understand example is cluster analysis. Clustering here means grouping data together.
The data does not have to be labelled. In contrast to classification in supervised learning, unsupervised learning finds the groups through the spatial similarity of the data points considered. This is used in customer segmentation, for example. By dividing customers into specific groups, they can be advertised to much more effectively.
Supervised learning
A typical example of supervised learning is the classification of objects. You can test object recognition on your smartphone with the app "Google Lens". Unlike unsupervised learning, this type of machine learning requires labelled data. This means, for example, that a photo of a dog must be labelled "dog". By using many labelled images of dogs, a model can be trained that can reliably recognise dogs even from new unlabelled images. A regional example of such an application can be found here:
https://siwalusoftware.com/de/
Reinforcement learning
Again, we will use an example that is easy to understand. You may know industrial robots that move similarly to a human arm. These robots are usually programmed for a work step in which a defined path is followed or something is gripped at a defined point. But how can a robot, for example, independently learn to pick the right object out of a container with randomly arranged objects (bin picking)?
Unlike the other types of ML, no special training data is required. In our example, the robot learns through reward and punishment. If a robot grasps a part correctly (target definition), it receives a "reward", if it does not grasp the part correctly, it receives a "punishment". The aim of the system is to maximise the number of correct grips. Until such a robot has learned the goal of independent grasping, a great many grasping processes have to be played through. The ML model then improves itself with use.
Establish machine learning sustainably
1. preparation of the data
Depending on whether an already pre-trained ML model is available or whether it is a user-defined ML model, different amounts of data are needed to train this model. In general, however, historical data is needed to forecast future events.
2 Identify the area of application
Start with the implementation of individual small applications.
3. create a machine learning culture
A mindset that is open to using machine learning to improve all processes in the organisation and to identify and explore the opportunities for doing so on the fly.
4. develop teams
Raise awareness of machine learning among engineers and decision makers.