Summary

Many of the concepts presented in this chapter overlap to a great degree. For very good reasons, particular combinations of design decisions are more common than others. However, it's important to realize that these design decisions are orthogonal (that is, mostly independent from each other). Table 35.1 summarizes the reliability of the different approaches.

Table 35.1. Decisions That Need to Be Made When Designing a Learning AI, Ranked by Their Likelihood of Success

	Reliable	Higher Risk
Type	Optimization	Adaptation
Representation	Knowledge	Behaviors
Phase	Offline	Online
Technique	Batch	Incremental
Component	Supervised	Feedback
Methodology	Training	Trial and error

The more reliable design decisions made, the more control the developer has over the NPC. In addition, the animats are likely to perform better and the AI to run more efficiently. There are challenges even with these safer options, but the higher-risk technologies complicate matters by an order of magnitude. Chapter 48 discusses this issue further.

Practical Demo

A skeleton animat, known as Loony, sets up to illustrate the ideas in this chapter. The animat provides three forms of feedback: supervised, reward, and fitness. Callback functions are called when this information is available, which can be interpreted as the AI engineer sees fit.