Teaching Artificial Intelligence to have fun

Max Planck researcher develops a 'Theory of Fun´

What is fun and why do we find certain things easier to do if we particularly enjoy dealing with them? Franziska Brändle at the Max Planck Institute for Biological Cybernetics in Tübingen gets to the bottom of this question. She examines key factors that make fun and motivation with a 'Theory of Fun' applicable to different everyday scenarios.

An early fundamental theory of fun, inner motivation and creativity is based on a concept developed by German computer scientist and AI researcher Jürgen Schmidhuber. According to this concept, fun results from an inner reward system. Above all, the reward system is stimulated when new, surprising patterns can be discovered and learned, which enable better orientation or evaluation of everyday situations.  

According to him, there is a reciprocal relationship between an inner motivation to have fun and its cause: finding something new and challenging oneself. For several years now, programmers in the gaming industry have been on the trail of finding ways in which computer games can adapt to the skill and learning success of their players. To them, one thing seems to be very obvious: early reward systems for solving tricky puzzles promote the fun of the game. And fun is a basic prerequisite for getting involved in a game at all.

On the search for a ´Theory of Fun´

In her research project, Franziska Brändle is now investigating new possibilities for a model. "Games can be simplified images of our real environment and thus serve as powerful learning tools for us," says Brändle. "We can use them to practice new patterns of behaviour and logic and receive immediate feedback. No other medium can do this so well. What we enjoy is the feeling of learning and mastering something by simply finding it out," she continues.

Puzzles that are too difficult or objectives that are set too high tend to lead away from performance successes - and motivation decreases. The same applies if a game is too simple. "We lose interest in it. If a game could be adjusted to the player's abilities through constant comparison, the success and fun of solving would be greater," Brändle explains.

Teaching artificial intelligence to have fun

Brändle's statements refer to processes whose principles can be used for new foundations in machine learning. According to her, software algorithms are to be developed and trained in such a way that they learn to find and optimize problem solutions in a self-motivating way.

The underlying reward system is knowledge itself. "It is some kind of a feedback signal to oneself, to be on the right track, to stay on it and keep going. This complex of different learning experiences corresponds to a certain stimulus pattern, behind which a mathematically explainable logical operation is situated. This pattern is optimised by continuous comparison with new learning experiences. It is perhaps what we would interpret as a factor of fun that software can experience," says Brändle.

New insights in virtual worlds

Despite some scepticism as to whether and how things learned in games can be transferred to our real life, games could provide a gain in knowledge comparable to books or films - you just have to try it, says the researcher.

"The next step for us is to recognize the ways in which we humans usually discover new things, for example, how we orient ourselves in unknown spaces and learn something new," explains Brändle. Various series of experiments in virtual worlds using the latest virtual reality technology will provide new results. We can be curious to see which learning models Franziska Brändle will develop on the way to a 'Theory of Fun'.