Curriculum learning with a hierarchical cellular encoding
G. Vassallo, Eleni Nisioti, Joachim Winther Pedersen, Erwan Plantec, Milton Llera Montero, Sebastian Risi · Proceedings of the Genetic and Evolutionary Computation Conference Companion · 2025
Evolving increasingly complex skills requires remembering and efficiently re-using past ones. This can be enabled by appropriately designing the genotype to phenotype (GP) map of an evolving system: skills can be encapsulated into modules that can be hierarchically combined during development. Here we study cellular encodings (CE), an early family of GP maps that can create such modular phenotypes by employing a genomic representation specifically designed to support hierarchy and modularity: grammar trees. Like functions in a computer program which can be called multiple times during execution, grammar trees can be executed at various points during development. Here, we apply CE in a curriculum learning setup and show that it can quickly progress across tasks of increasing difficulty by re-using its past solutions. In contrast, an algorithm that searches directly in the phenotypic space (NEAT) does not benefit from a curriculum while another GP map (HyperNEAT) fails when the curriculum is present. We show that enforcing hierarchy in the CE, through the use of a nested set of grammar trees, is necessary to observe these benefits. We believe that future work should scale such approaches, either by scaling CEs directly or integrating these ideas into more powerful approaches.