Artificial intelligence from 1990 to 2020

 Until the year 1990 it was unclear how to realize AI and robotics, but 30 years later in 2020 AI was available. The period in between should be described because its equal to the existence vs. the absence of artificial intelligence.

The most advanced technology until 1990 was the classical computer workstation including its connection to the internet. The typical workstation was equipped with graphics cards, sound cards, moderate amount of RAM and a hard drive. From a software side a unix compatible operating system including multitasking GUI environment was a typical application.  Such a unix workstation wasn't able to control robots and it wasn't equipped with AI, but it was an ordinary example of well engineered computer technology.

AI in the year 1990 was only available in movies and in science fiction books which are describing intelligent robots who are doing useful tasks, like driving cars, walking on two legs and do complex assembly tasks. In the reality these tasks wasn't out of reach for the engineers. Even most advanced books and papers written at Universities in 1990 doesn't contain algorithms, nor ideas how to realize robots in reality.

The single cause why engineers until 1990 have struggled with AI is because of missing problem definition. It was unclear what AI is about from a mathematical standpoint. Without such a definition it wasn't able to program it. A user oriented definition like "A robot has to walk on two legs" isn't enough to create a mathematical equation or to program an algorithm. Programming even very simple robots resulted every time into failed projects. That means simple wheeled robots in 1990 were not able to find the exit in a simple maze.

From 1990 until 2020 lots of effort was put into the AI and robotics issue. The most promising direction was to create a precise  problem space first and with this problem space, different algorithms can be combined. There are two major approaches available:

1. Create a physical problem space, e.g. invent a robot competition like Micromouse and Robocup
2. Create a dataset problem space, e.g. a dataset with motion capture recordings, or a dataset with OCR problems

After a problem space was created for example a micromouse maze, its possible to benchmark a certain robot and a certain algorithm how well it performs in the puzzle. For example, a certain robot will need 30 seconds until it found the exist. Or a neural network will recognize in a ocr dataset 50% of the pictures correctly.

During the period 1990-2020 endless amount of datasets and robotics competitions were introduced, described and compared to each other in the academic literature. Especially the second attempt "create a dataset problem space" has become an important enebling technology in AI, because it allows to discuss AI from a mathematical standpoint. Instead of asking what is AI from a philosophical standpoint, the new question was, if a certain dataset makes sense, or what the score for a certain neural network is on a dataset. These hard scientific questions can be adressed with existing tools like back propagation algorithms, statistics and diagrams. A well defined problem space realized as a machine learning dataset was the major steps from former alchemy driven AI philosophy towards a scientific defined AI.

In the published literature from 1990 until 2020 it can be shown, that over the years more advanced datasets were created which were solved with more advanced neural network algorithms. Early datasets in the 1990 were short tables which were described only in the appendix of a paper, while in later years, the entire paper described the dataset in detail because it was the main subject.

Modern AI published after the year 2020 is mostly the result of improved problem formulation. After this date, endless amount of physical problem descriptions plus thousands of datasets with additional problem descriptions are available. These datasets are very realistic, they have to do with real challenges like grasping objects, biped walking, path planning in a maze and even natural language understanding. So we conclude, that the search for Artificial intelligence is equal to the search for a problem formulation. Only if a problem was formulated in a mathematical format, a computer can be used to solve it.