An adaptive radius adjusting method for RBF networks considering data densities and its application to plant control technology

We previously proposed a control technology to reduce CO and NOx emissions in power generating. In this technology, an optimal control logic is obtained by Reinforcement Learning (RL) and a Radial Basis Function (RBF) network which constructs a response surface for the CO and NOx properties. An improvement of estimation accuracy of the response surface can enhance the control logic performance, so the radius of RBF network should be determined properly since it is one of the most influential factors on estimation accuracy. On the other hand, adjustment of the radius should be executed within several minutes as computational time for constructing the response surface is restricted. In this paper, we propose a new radius adjusting method for RBF networks to achieve high estimation accuracy and short computational time. This method adjusts radii based on the densities of learning data, thus it can achieve both high estimation accuracy and short computational time. The results of our evaluation showed that the proposed method had higher estimation accuracy than conventional methods within a practical computational time.