Benchmarking Punctuated Anytime Learning for Evolving a Multi-Agent Team's Binary Controllers

Abstract

Punctuated anytime learning (PAL) is a system that can be used for evolving cooperative teams of agents. PAL applied to evolving multiple populations is a kind of cooperative revolutionary algorithm (CCEA). Here we compare PAL to a canonical genetic algorithm (GA) on a widely used GA benchmarking optimization function called the Rosenbrock function. The Rosenbrock function was chosen for experimentation because it is a highly non-linear function and difficult to optimize. Results are shown from a variety of experiments with different dimensionalities of the Rosenbrock function. These findings are discussed in the context of evolving binary controllers for multi-agent cooperative teams of robots.