It’s a jungle out there: exploding asteroids, fire-breathing dragons, and maniacal cultists, among other lethal threats. What and how can you teach newborn babies, somehow fitted with a hero’s body and magical weapons, to defend themselves and survive for an hour? And how do you coach the asteroids, dragons, and cultists to be convincingly menacing? And what can the young hero’s loyal companions say to sound sincerely encouraging? That is roughly the challenge addressed by the general video game AI (GVGAI) framework and competitions. Checkers, chess, and Go seem pretty tame by comparison. The issue is not just the complexity of the environment, but also the desire to find a solution that works for diverse game domains. Success in this area might bring us closer to an AI approach for real life.

The core of the GVGAI approach is the video game description language (VGDL), a hierarchical JavaScript Object Notation (JSON)-like object notation with a flexible game-relevant attribute vocabulary. Basic VGDL handles 2D arcade games, anything with discretely movable sprites representing players, enemies, and obstacles, with collisions that may be destructive. Extensions of VGDL allow for multiple players, continuous motion, and 3D environments. Having a language like this means that many agent-behavior algorithms can be applied to many individual games, and their effectiveness can be tested and compared using a common execution framework. In fact, competitions have been held annually since 2014, with participants from several universities. Submissions are tested against classical arcade games such as Missile Command and Zelda.

The GVGAI execution framework consists of an interpreter that accepts a VGDL description of a game environment and its initial state, and accepts inputs from user-controlled agents. The user agents may be either human or candidate algorithms. The framework can be either set up locally or, in the case of competitions, hosted on a common server. Setting up and interacting with the framework involves some experience with GitHub, Python, shell command lines, and maybe Java. The exercises at the end of each chapter invariably ask for interactions with various software resources. Although I am not inexperienced in these requisites, the GVGAI Gym YouTube tutorial was a real blessing.

The research aspect of this activity is the design of algorithms for agent behavior, and also generating or tuning game content. The book reports on the application and enhancement of advanced techniques for game analysis, planning, and machine learning. Success is judged by good performance in the competitions. Often the best results arise from hybrid approaches, which is discouraging from a scientific point of view but stimulating from an engineering point of view. GVGAI has become a large, active research community, and studying this book is essential to joining it.