The conditional expectation via a general class of nonlinear networks
Mang Zhu, JAMES A. CADZOW · 2002
A general class of nonlinear parametric multi-layered networks is introduced. This class is a generalization of the standard neural network. The dynamic behavior of members of this class are analyzed and the popular least squared error criterion is used for judgement of goodness of model fit. The output of the network is shown to be an estimator of the conditional expectation function for the desired output under condition made on the given inputs. Multi-directional search (MDS) as a new nonlinear programming technique is discussed in the paper. Examples of the simulation results are given at the end of the paper to show the exact fit of the calculated expectation function and the output of the network. The results from back propagation and MDS are compared.>