Global and semilocal estimation in multistage optimal control

A main issue in multistage optimal control (MOC) problems is the choice of a suitable class of models for the approximation of the cost and/or the control functions. We present a comparison between local and global approaches to estimation through the application of semilocal kernel models and neural networks, respectively. Two methods for the solution of MOC are considered, namely, the approximate dynamic programming and a method based on a direct optimization of the optimal control functions.