Minimum description length (MDL) principle is one of the well-known solutions for overlearning problem, specifically for artificial neural networks (ANNs). Its extension is called representational MDL (RMDL) principle and takes into account that models in machine learning are always constructed within some representation. In this paper, the optimization of ANNs formalisms as information representations using the RMDL principle is considered. A novel type of ANNs is proposed by extending linear recurrent ANNs with nonlinear “synapse to synapse” connections. Most of the elementary functions are representable with these networks (in contrast to classical ANNs) and that makes them easily learnable from training datasets according to a developed method of ANN architecture optimization. Methodology for comparing quality of different representations is illustrated by applying developed method in time series prediction and robot control.