Chapter 33. Learning Classifier Systems

Key Topics

• Representation of Classifiers

• Overview of the Classifier System

• Architecture

• Discussion

Learning classifier systems (LCSs) are best understood as a combination of three popular AI techniques: genetic algorithms, rule-based systems, and reinforcement learning. LCSs are capable of learning to predict the best action based on the current situation (a classification problem).

Rule-based systems require significant involvement from a human designer—the knowledge acquisition bottleneck. Instead, a system could use the creativity of genetic algorithms and the guidance of reinforcement learning to find the right behavior autonomously—as most animals do. By using information gathered from the environment and a high-level measure of performance as feedback, it's theoretically possible for a system to find the most appropriate set of rules. In practice, the adaptation can be difficult to apply at runtime in a realistic fashion, but LCSs have their advantages compared to other approaches (for instance, they're automated).

This chapter covers the following topics:

• The underlying representation of an LCS as the set of rules known individually as classifiers, consisting of the head (matching the sensory input) and the body (corresponding to the effectory outputs)

• The additional information stored by the rules, notably the prediction of their benefit and an estimate their error

• An introduction to classifier systems, including their internal functioning from an architectural point of view and each of the components

• The different aspects of LCSs, their different practical problems, and their applicability from a game developer's perspective

As well as introducing the theory behind classifier systems, this chapter shows one way genetic algorithms can be applied in practice.