Towards effective neuromorphic controllers

A feedforward controller based on a multilayer neural network and several control configurations is proposed. The basic weak points of the controller are discussed and several approaches to attain better performance are investigated. These include, among other aspects, faster updating algorithms, efficient use of the available knowledge about the plant, better training procedures, hybrid configurations using nonneural control techniques, efficient definition of the structure, and initialization and prelearning of the neuromorphic controller.<>