Modularity and Regularity in Neural Networks Produced with Assembler Encoding
Tomasz Praczyk · Computational Methods in Science and Technology · 2013
The main focus of the paper is on the ability of the neuro-evolutionary method called Assembler Encoding to repeatedly use the information included in a genotype and to construct modular and/or regular neural networks.It reports experiments whose the main goal was to test whether the method is capable of adjusting topology of neural networks to a modular and regular problem.In the experiments, the task of Assembler Encoding was to evolve neuro-controllers responsible for balancing two or three inverted pendulums instaled on separate carts.Since both the carts and the pendulums were identical the task of neuro-controllers could be performed by means of modular/regular neural networks.