Fitness Morphs and Nonlinear Projections of Agent-Case Embeddings to Characterize Fitness Landscapes

Abstract

The fitness landscape of an evolutionary computation system is a set of points defined by the representation of the potential solutions with a connectivity created by the variation operators. In real optimization, fitness landscape analysis can profitably confuse this connectivity with that provided by the usual metric structure for Euclidean space (they are actually very different). In this chapterwe examine fitness landscapes for a variety of discrete problems including finding self avoiding walks, finding features for DNA sequence classification, the Tartarus AI test problem, and locating cellular automata rules.We also examine a novel real optimization problem connected with the Mandelbrot set. We use agent-case embeddings, fitness morphs, and nonlinear projection to explore the fitness landscapes of these problems in a series of extended examples.All of these techniques transform information about discrete fitness into real-valued spaces enabling both analysis and visualization.