Evolution of Neural Structure and Complexity in a Computational Ecology

摘要

We analyze evolutionary trends in artificial neural dynamics and network architectures specified by haploid genomes in the Polyworld computational ecology. We discover consistent trends in neural connection densities, synaptic weights and learning rates, entropy, mutual information, and an information-theoretic measure of complexity. In particular, we observe a consistent trend towards greater structural elaboration and adaptability, with a concomitant and statistically significant growth in neural complexity.