© Freepik
While searching for current algorithms for special solutions, I keep coming across the most diverse designations and, upon closer inspection, realize that I already know this algorithm, but in a different setting. Therefore, the first step here is to clarify which possible questions we can/would like to answer with AI algorithms.
Author: Axel Preis, Co-Author: Stefanie Schwaar
According to Wikipedia, an algorithm is "[...] a unique set of action rules for solving a problem or class of problems." By AI, we mean a string of different action rules. The algorithm contains thereby a repetition up to a satisfying result was reached. Here it is essential that the underlying data are of sufficiently good quality. More on this in one of the following blog post – today we're talking about AI algorithms.
To better understand how different algorithms work, let's imagine the following situation: We have an urn that contains several balls. The balls may differ in terms of various characteristics such as weight, surface texture, or color. We cannot see inside the urn, but we can reach into it. One after the other we remove objects from the urn. In this way, we successively learn about the spheres and thus answer various questions. Let's now take a look at which questions we answer with which type of algorithm.
With a classifier, it is possible for us to learn an inference to a non-observable feature or class based on features of the objects. Thus, given a new object, we can determine this class based on the learned association.
This is illustrated by the following scenario:
A blindfolded person draws balls from the urn and is told what color the ball is. At the same time the person perceives the surface structure and the weight. Because the green balls are always heavy or medium-heavy in combination with a rough surface and the red balls are light and smooth or medium-heavy and smooth, the person is able to determine the color of the drawn ball himself after only 20 balls. Now the person acts as a classifier, who has learned from history to recognize the ball color on the basis of other features.
By a classifier we understand an algorithm, which assigns an object, which we consider, to one of the given classes.
A grouper can be used to learn which objects are similar or different from each other based on features of the objects. This allows us to assign a new object to a group of similar objects.
Now consider the following situation:
A blindfolded person draws balls from the urn. Again the person feels the surface structure and the weight of the balls. Here she notices that she pulls light balls with a smooth surface, light balls with a rough surface, heavy balls with a smooth surface, and heavy balls with a rough surface. Thus, the person has learned, that there are four groups of balls and now acts as a grouper, because the person has learned to assign a newly drawn ball to one of the four groups.
Consequently, a grouper is an algorithm that assigns the object of interest to a group of similar objects.
A prognosticist allows us to learn a tendency for the temporal course of the expressions of a feature of the objects based on past expressions of that feature. In this way, it is possible for us to predict future expressions.
Consider the following situation:
Again, a person who cannot see anything draws balls from the urn. This time she perceives the weight of the balls, and wants to predict the weight of the ball she draws next. Since there are rather light balls in the upper part of the urn, while the balls at the bottom are heavier, after only a few draws the person is able to make a prediction for the weight of the next ball. The prediction turns out to be slightly higher than the perceived weight of the previously drawn balls. She now acts as a prognosticator who has learned to make predictions for the next ball weight based on past weights.
Consequently, by a prognosticator we mean an algorithm that predicts feature expressions of the object we are looking at using expressions from the past.
With a detector, we are able to learn what we consider to be a normal object based on features of objects. Thus, objects that are out of line can be detected.
Now consider the following scenario:
A blindfolded person draws balls from the urn and perceives the surface texture and weight. In addition, she is told what colors the drawn balls are. Since green balls are usually heavy or medium and have a rough surface and red balls are light or medium with a smooth surface, she detects an anomaly when she pulls a light, smooth ball and is told it is green. The person now acts as a detector that has learned from history what combinations of features are common and consequently is able to detect atypical balls.
By a detector, we therefore mean an algorithm that identifies individual atypical objects among a large number of common ones.
In this blog post, we have presented different types of AI algorithms in a very simple way. These can be used to solve a wide variety of problems. What they all have in common is that they have to learn from experience in order to provide solutions. If you want to dive deeper into this topic, you can look forward to future blog entries.
Fraunhofer Institute for Industrial Mathematics ITWM