Beneficial Preadaptation in the Evolution of a 2D Agent Control System with Genetic Programming

摘要

We examine two versions of a genetic programming (GP) system for the evolution of a control system for a simple agent in a simulated 2D physical environment. Each version involves a complex behavior-learning task for the agent. In each case the performance of the GP system with and without initial epoch(s) of preadaptation are contrasted. The preadaptation epochs involve simplification of the learning task, allowing the evolved behavior to develop in stages, with rewards for intermediate steps. Both versions show an increase in mean best-of-run fitness when preadaptation is used.