Functional Cycle Components

Abstract

This chapter aims to build upon the brief, simplified description of an LCS functional cycle outlined in Section 1.3. Previously, we discussed how all LCSs include a form of discovery and learning components, and Figure 1.3 specifically illustrated many of the common LCS algorithm components in step-wise order. Here we will discuss these algorithmic components in greater detail, introduce some new ones, consider key adaptations to problem domains beyond the multiplexer example, and begin to discuss methodological differences between supervised and reinforcement learning, all within the purview of Michigan-style LCS architectures (see Section 4.3.3). This chapter will emphasise how the functional cycle seeks to learn useful state-action mappings by (1) matching the input state to classifiers (and triggering covering if needed), (2) determining whether these classifiers are correct or incorrect (or returning reward if the exact output is unknown), (3) updating the associated classifiers so their worth may be evaluated, (4) discovering potentially better rules when appropriate, and finally (5) deleting the least-contributing classifiers if necessary.