Abstract :
Although AI has excelled at many narrowly defined problems, it is still very far from achieving human-like performance in terms of solving problems that it was not specifically programmed for: hence the challenge of artificial general intelligence (AGI) was developed to foster more general AI research. A promising way to address this is to pose the challenge of learning to play video games without knowing any details of the games in advance. In order to study this in a systematic way the General Video Game AI (http://gvgai.net) competition series was created. This provides an excellent challenge for computational intelligence and AI methods and initial results indicate often good though somewhat patchy performance from simulation-based methods such as Monte Carlo Tree Search and Rolling Horizon Evolutionary Algorithms. Observing where these methods succeed and fail leads to the conclusion that there is still much scope for further developing algorithms that mix simulation with long-term learning. While running the competitions we have built up a large set of GVGAI players. This large pool of adaptive players leads on to very appealing potential applications in automated and semi-automated game design where the player-set can be used to evaluate novel games and new parameter settings of existing games. Initial explorations of this idea will be discussed.
