Constructive Transparent Radial Basis Function Network Learning for Non-linear Control

In this work a constructive radial basis function network (RBFN) learning method is applied. This approach uses the functional equivalence principles between RBFN and fuzzy systems in order to achieve a minimal structure network. Firstly, an initial network based on linguistic descriptions is constructed. Secondly, a constrained constructive adaptation law, based on a minimal resource allocating algorithm, is applied in order to on-line adjust the structure and parameters of the RBFN, keeping the transparency property and guaranteeing the linguistic interpretation. Thus, at any instant, knowledge from the network can be easily extracted, validating its structure. Experimental results on a benchmark process show the effectiveness of the presented approach.