Regresar

Approximate dynamic programming methodology for data-based optimal controllers

Abstract:

In this article, we present a methodology for learning data-based approximately optimal controllers, within the context of learning and approximate dynamic programming. There are previous solutions in dynamic programming that use linear programming in discrete state space, but cannot be applied directly to continuous space. The objective of the methodology is to calculate data-based optimal controllers for continuous state space, these controllers are obtained by a lower estimation of the accumulated cost through functional approximators with linear parameterization. This is solved non-iteratively with linear programming, but it requires to provide appropriate conditions for regressor regularization and to introduce a cost of leaving the region with valid data, in order to obtain satisfactory results (avoiding unrestricted or poorly conditioned solutions).