The research domain of artificial intelligence (AI) has seen an enormous rise in interest in the past years and has led to many interesting and powerful discoveries. Games are of particular interest to research in AI, as games often form highly complex, but well-defined environments with a strong resemblance to the real world, which are safe, fast and easy to tune and configure for experimentation purposes.

Research results achieved with games regularly have a strong impact on applications beyond games. Typical examples of this are the widespread applications of deep neural networks, which initially got major attention through the results achieved with AlphaGo (Go) and AlphaStar (StarCraft II). Other examples are the use of game-playing AI agents in research areas such as (bio)chemistry, physics, mathematics, and economics. Conversely, real-world knowledge often impacts game technology. This is seen, for instance, in the development of digital twins, and research into procedural content generation.

Despite the many examples of the successful transfer of game AI to and from real-world domains, there are still many scenarios in which transfer between disciplines is in need of further research. For instance, many educational games have as a main objective to inform and teach players about certain topics, but the conditions under which this knowledge transfer is effective are not clear. Another example is the transfer of modern AI methods from pure research to their practical use in the games industry. One prevalent obstacle is that it is seldom obvious how a given algorithmic problem encountered in a game industry setting can reliably be solved with AI in a way that does not interfere with the game's goals, limitations, and requirements.

Our objective with this seminar is to study how to boost or even streamline transferability of game AI to and from different domains and different applications. To apply modern AI solutions to games, we need to be able to identify which parts of an existing real-world problem are transferable to a game's problem. A similar identification has to be made to transfer AI solutions between different games, or from games to the real world.

This Dagstuhl Seminar aims to explore the transferability of game AI and find theoretical and practical answers to what features make the transferability of game AI possible, what results can be expected, and what the limitations are. While the seminar will have a main focus on the transferability of AI between games, it will also include the transferability of real-world AI to game AI, and of game AI to real-world AI. Both technical and functional aspects of transferability will be considered.

The setup of this seminar is based on working groups, rather than presentations. During the seminar participants will propose problems to explore, next to problems posed by the organizers. Small groups will be formed to work on problems that interest them. A group may simply discuss a problem to try to gain new insights, or do experimental work by building prototypes. The exploration of original ideas and approaches is encouraged. Usually specific working groups last no more than a day, and at the end of each day the groups will briefly present their results in a plenary setting.

Past seminars in this series often produced results that led to novel research lines or were the start of new collaborations. A major goal of the seminar is to integrate young researchers into the game AI research community, and help them develop their ideas.

Creative Commons BY 4.0

Duygu Cakmak, Alexander Dockhorn, Diego Perez Liebana, and Pieter Spronck