At the Institute for Information Processing (TNT), methods in the broad field of Machine Learning are developed and applied to various problems. One of these application areas is the development of artificial intelligence in games. Due to their diversity and accessibility, they offer an ideal test environment for machine learning algorithms and present them with numerous interesting challenges, which form the basis for numerous research projects at TNT.
Games such as Chess and Go have a wide range of possible actions per move. This leads to enormous complexity in decision making and often requires the use of heuristic techniques to approximate the best possible move. This problem is magnified in the area of digital games, where, especially in real-time strategy games, we often need to control multiple characters simultaneously and in real-time. At the TNT, we are investigating methods to abstract (simplify) this decision space to enable efficient decision making.
Many methods of autonomous game-playing are related to a single game. In the context of general game-playing, agents are developed that not only play a single game successfully, but are able to perform well in multiple games. This can be done on the use of flexible search techniques or it can be enabled by transfer learning. In the latter, partial models of the developed solution are transferred between applications to build on previously learned knowledge and thus accelerate the re-learning process.
Especially in general game-playing, robust agent behavior should be learned. Numerous machine learning methods are based on the recognition of patterns in the form of correlations. However, these are often subject to a bias and can lead to false conclusions. Modeling such a bias and learning causalities instead of correlations is fundamental to developing more robust agent behavior. Causal learning methods can also significantly increase the interpretability of behavior by inferring relevant factors that impact the agent’s decision making.
Game development is becoming increasingly complex. More and more developers are becoming involved in the process. Large-scale productions such as Grand Theft Auto and the Elder Scrolls series spend many hours creating huge immersive worlds that players can immerse themselves in. Procedural Content Generation describes a process to reduce development effort by automating the creation of game content. The TNT uses machine learning techniques to extract patterns from existing content and combine them to generate new and interesting content.
