Tutorial I: Video Game Description Language (VGDL) and the challenge of creating agents for General Video Game Playing (GVGP)

Games have attracted many AI researchers over the last decades, allowing to benchmark multiple and varied simulated environments to test AI algorithms in. It is not rare that information about the game played is supplied in order to help the algorithm to achieve better results. This may pose a problem: in some cases it is debatable if the success of a given technique is due to the algorithm itself or to the heuristics employed. In the worse case scenario, one could argue that the research outcome is of little interest to AI because it is too tailored to the game used. General Game Playing (GGP; for board games) and General Video Game Playing (GVGP; for real-time games) are disciplines that try to tackle this issue by employing several games, that must be played by the same algorithm. The number and types of heuristics that can be used is limited by the nature of the problem: domain-specific knowledge can overspecialize the algorithm to one or some of the games, which becomes a problem if some of them are not known in advance. In this tutorial, we will explore the Video Game Description Language (VGDL) and the challenge of creating agents for GVGP. We will use the General Video Game AI (GVGAI) framework, employed for the GVGAI competition, to create different types of controllers, from simple to more complex solutions. These controllers will be able to play, at distinct performance levels, many different games without the need of domain specific heuristics. Furthermore, we will analyze the present and future challenges of GVGAI, as well as the different research opportunities it offers.