Small is Beautiful: Near Minimal Evolutionary Neurocontrollers Obtained With Self-Organizing Compressed Encoding
Shlomy Boshy, Eytan Ruppin · The MIT Press eBooks · 2002
This paper presents a novel method for evolution of artificial autonomous agents. It is based on adaptive, self-organizing compressed genotypic encoding (SOCE) of the phenotypic synaptic efficacies of the agent's neurocontroller. The SOCE encoding implements a parallal evolutionary search for neurocontroller solutions in a dynamically varying and reduced subspace of the original synaptic space. It leads to the robust emergence of compact, near minimal successful neurocontrollers starting, from arbitrarily large networks. This is important since on practice the network size needed to solve the problem is unknown beforehand. The SOCE method may also serve to estimate the network size needed to solve a given task, and to delineate the-relative importance of the neurons composing the agent's controller network.