Published 31. July 2021
Until the 1990s, the academic algorithms of artificial intelligence were predominantly not considered applicable by the games industry, sometimes due to a lack of computing power on the end devices or a strongly divergent objective. The game genres that emerged during this time and are still popular today were developed in a way that cleverly circumvented this shortcoming: game characters were only ostensibly intelligent and interest in a game was always aroused by other aspects. These development patterns have been inherited and maintained in many games to this day.
Although the findings of the acadamic research branches concerned with artificial intelligence often eventually found their way into the development of games, this often happened with a long delay. In some cases, decades passed before a breakthrough in science also led to a breakthrough in game development. It is only in recent years that an acceleration of this transfer rate can be observed. Modern AI is now already being used in a variety of ways in the development of games, for example for procedural content generation or for the realistic animation of game characters. However, many developers still shy away from providing their customers with an AI that learns and changes during gameplay.
There are some frequently mentioned reasons for this. The computing power of the end devices is still too low -- the objective of academic AI makes the use in computer games impossible -- it would be difficult to predict what an AI learns in the game and therefore the quality and balance of the game could not be ensured -- academic AI would ultimately only bring advantages in production -- an interesting AI is also not a prerequisite for an interesting game ... just to name a few.
But what about a game where the unpredictability of what a character learns is part of the game? What if a game is interesting precisely because of this feature and doesn't get boring anytime soon? Learning and uncertainty can and must go hand in hand with a new game concept that breaks with tradition. In return, it opens up completely new possibilities: An almost unlimited number of possible game events, possible behaviors, interactions between game characters. And on the developer side, the effort to define diverse game content down to the last detail could be reduced. Of course, we cannot expect superhuman performance from an AI that has not been pre-trained on mainframes. But do we expect superhuman performance from our pet and turn away from it if it doesn't solve differential equations?
Our AI players exhibit three key characteristics that distinguish them from other AI: General learning, emergent behavior, and dynamic control. The potential of flexible and learning game characters is enormous. With Marbles AIĀ® we will take a first step in this direction.
Dr. Fabian Schrodt
Show articles: All Tech Games Cognitive Science General Artificial Intelligence AI Industry