Evolutionary algorithms based on machine learning accelerate mathematical function optimization but not neural net evolution

For a decade, the second author has been dreaming of and working towards building artificial brains that consist of tens of thousands of evolved neural net circuit modules that are assembled according to the designs of human brain architects (BAs). The bottleneck with this approach is the slow evolution time of the modules (using software techniques in PCs). However, using Michalski\´s machine learning based evolutionary algorithms, such as "LEM (learnable evolution model)", the usual evolution time (for certain categories of applications, e.g. mathematical function optimization) can be reduced by a factor of hundreds (Michalski, 2000). The authors hoped that this breakthrough would allow neural net modules to be evolved far more quickly. Unfortunately, it appears that the LEM approach does not work well with the evolution of dynamic neural nets. This may be due to a combinatorial explosion of attribute-variable pairs arising during the machine-learning mode that poses a problem during the evolution of dynamic signals.