On the training of a multi-resolution CMAC neural network

Several artificial neural network architectures have been proposed to solve problems in control systems and system identification. However, not all neural network structures are equally suitable for real-time adaptive situations. Lattice-based Associative Memory Networks (AMNs) have several properties that are attractive for real-time adaptive modeling and control. An example of a lattice-based AMN is the Cerebellar Model Articulation Controller (CMAC) neural network. A hierarchical multi-resolution lattice approach is proposed and investigated through experimentation as a possible approach to alleviate the main drawback of AMNs: the required storage is an exponential function of the number of inputs. Relevant issues are discussed and suggestions for future improvements are given.