Agent-Case Embeddings for the Analysis of Evolved Systems

Abstract

This paper introduces agent-case embeddings, a general purpose tool for detecting a variety of solutions produced by an evolutionary algorithm. They can also be used to explore the geometry of the space of problems that agents attempt to solve. Agent-case embeddings permit the comparison of solutions evolved with different representations by directly comparing phenotypes. Use of agent-case embeddings requires that multiple instances of the problems solved by the agent be available or contrivable. Three examples of agent-case embeddings are derived for apoptotic cellular automata, agents playing the iterated prisoner's dilemma, and simple virtual robots performing the Tartarus task. The use of agent-case embeddings is shown to permit visualization of the diversity of evolved agents, demonstrates the impact of changing algorithm parameters, and explores the impact of different representations on evolutionary search. The algorithm parameters explored include population sizes, elite fraction, and choice of variation operators. Agent-case embeddings are used to demonstrate that a novel technique called single-parent crossover can localize evolutionary search in a small part of the adaptive landscape in a controlled manner.