On World Models

Reactive intelligence typically does not have any representation of the world; thanks to the sensors, the world acts as its own model. This proves to be a surprisingly efficient and reliable form of AI. However, AI purely based on reflexes lacks a main characteristic of human intelligence: memory. Therefore, purely reactive animats fail to reach the capabilities of biological players. Although reactive systems may be designed almost as intelligent as ones with world models, they lack realism.

Throughout the development of the animats presented in this book, the animats are given small senses of state (for example, to keep track of the enemy). This not only makes the AI more realistic, it is also much easier to develop. Without relying on advanced concepts behind internal representations of the world, this additional knowledge allows the animats to perform on a human level.

Increasing the capabilities of the AI, however, requires extending the concept of world model—providing each animat with personal knowledge of its environments. The main issue involved is one of representation, because a very flexible and expressive knowledge representation language is needed to store the short- and long-term memories of animats.

As well as providing the potential for nondeterministic behaviors, world models allow the AI engineer to create interesting human phenomena such as forgetfulness and surprise. Advanced world models, beyond those depicted in this book, are extremely rare in professional game AI. That said, there is a wealth of academic information on the subject, originating from classical AI in mid-twentieth century as well as recent research. See the web site at http://AiGameDev.com/ for some resources.