Full factorial ablation of GEAKG layers

Conduct a full factorial ablation study that varies the GEAKG’s L0 MetaGraph topology, L1 operator pool, and L2 learned knowledge (pheromone matrix and symbolic rules) independently, including comparisons between Ant Colony Optimization (ACO) cold-start and transferred pheromone configurations on each target domain, to isolate the contribution of each layer and their interactions.

Background

GEAKG is organized into three layers: L0 (topology), L1 (executable operators), and L2 (learned pheromones and symbolic rules). The paper presents ablations that isolate operator quality and sequencing (e.g., random versus learned ordering), but it does not systematically vary all three layers independently.

A comprehensive, factorial ablation would clarify how much each layer contributes to overall performance and transfer, and whether benefits arise from learned pheromones versus L0 structural priors or L1 operator quality. This is particularly relevant for evaluating ACO cold-start behavior versus using transferred pheromones on target domains.