Abstract: By means of more detailed and faithful simulation of the heuristic process where the expert assign the regulator gains, an advanced real-time expert-like self-learning algorithm has been developed. The first problem solved is how to recognize the pattern of a transient process and extract its features from sequentially sampled regulation error data corrupted by various noises and disturbances. The second is how to map those features which reflect various properties of the essential feedback regulation system onto the regulator gains for further improvement of its performances.