This chapter describes the application of a general regression neural network (GRNN) to control the flight of a helicopter. This GRNN is an adaptive network that provides estimates of continuous variables and is a one-pass learning algorithm with a highly parallel structure. Even with sparse data in a multidimensional measurement space, the algorithm provides smooth transitions from one observed value to another. An important reason for using the GRNN as a controller is the fast learning capability and its non-iterative process. The disadvantage of this neural network is the amount of computation required to produce an estimate, which can become large if many training instances are gathered. To overcome this problem, it is described as a clustering algorithm to produce representative exemplars from a group of training instances that are close to one another reducing the computation amount to obtain an estimate. The reduction of training data used by the GRNN can make it possible to separate the obtained representative exemplars, for example, in two data sets for the coarse and fine control. Experiments are performed to determine the degradation of the performance of the clustering algorithm with less training data. In the control flight system, data training is also reduced to obtain faster controllers, maintaining the desired performance.
Using general regression with local tuning for learning mixture models from incomplete data sets
