Architectures

AI architectures allow focused techniques to be assembled together to produce behaviors and capabilities. As levels of intelligence need to scale up, architectures will play an increasingly important role in the development of commercial game AI systems.

The idea behind patterns is to provide guidance for custom architectures using proven designs (reminiscent of software engineering). This book presents a variety of different AI architectures, but most are based on the same pattern. This pattern results from a modular approach that abstracts the common functionality as reusable capabilities, and provides behaviors that are convenient to experiment with.

AI Modules

At the lowest level, AI techniques are encapsulated modules. By abstracting out the implementation behind common interfaces, the techniques themselves become transparent and interchangeable. Therefore, the AI engineer can design the architecture first, and afterward decide which technique is used to provide the functionality.

Many of the AI techniques discussed in this book can in fact provide very similar functionality, such as pattern recognition, regression, or prediction. This includes neural networks, finite-state machines, or rule-based systems, each of which can provide control or problem-solving facilities interchangeably. Because of the tremendous flexibility of such technology, it can be easily reused in many parts of the architecture.

Capabilities

A common set of practical capabilities provides tremendous assistance when creating the illusion of intelligence. For example, movement, shooting, and weapon selection are capabilities. Each of them is a component built upon the AI techniques—generally only one.

There are two different ways to implement the capabilities to make them readily available for nonplayer character (NPC) development:

• Customized components provide the correct capabilities by passing specific data to the AI technique during the initialization. The capabilities are particular instances of the AI modules.

• Dependencies allow the capabilities to be implemented as independent components that provides their own interfaces, but that depends on AI modules.

The capabilities have the advantage of being conceptually independent from others, so they are very easy to develop.

Behaviors

The higher-level components provide distinctive behaviors, such as hunting, evading, and camping. Each of these behaviors can be built almost trivially when relying on the capabilities.

Because game AI is based on the appearance of intelligence, visible behaviors are the essence of the problem. Therefore, behavior-based design provides many advantages for game developers, allowing them to use techniques from nouvelle AI such as incremental development and extensive testing.

Arbitration

Finally, the top level of the architecture is a component that arbitrates between the behaviors. Its role is to decide which behavior needs to be activated based on the current situation.

Compared to lower-level tasks, the arbitration is relatively simple because all the details are abstracted out by the behavioral components. This makes them simple to handle from every point of view:

• Designers can control the arbitration using static techniques such as finite-state machines or the subsumption architecture.

• The AI can learn to control the system using adaptive techniques such as reinforcement learning or classifier systems.

The flexibility in the implementation of the arbitration is an incredible advantage for game AI developers, because they can opt for the simplest solution first—using adaptive techniques to provide interesting behaviors within the game.