Our Results
It is hard to measure the capability of a model. Although NEAT's generalization is lower, it had the best map traversal. Also, the simpler models such as {7, 5} would excel during training. However, on multiple maps, the genetic algorithm would have a harder time focusing in on the best solution. More complex models such as the ones built by NEAT would perform better.
Summary
The scores of each structure is the average normalized score across all maps
The map listed on each column is the training map
The average of each row is the generalization score of that specific structure
{9, 5} and {9, 12, 5} generalized the best.
Take Away
A simple model that fits a specific problem will generally excel at that problem. However, this is not the case for generalization. More complex models are needed to generalize across a variety of maps.