CONVERGÊNCIA DO ESTIMADOR RLS PARA ALGORITMOS DE PROGRAMAÇÃO DINÂMICA HEURÍSTICA

Abstract

The union of methodologies for optimal control and dynamics programming has stimulated the development of algorithms for realization of discrete control systems of the type linear quadratic regulator (DLQR). The methodology is based on reinforcement learning methods based on temporal differences and approximate dynamic programming. The proposed method combines the approach of the value function by method RLS (recursive least squares) and approximate policy iteration schemes heuristic dynamic programming (HDP). The approach is directed to the assessment of convergence of the solution DLQR and the heuristic weighting matrices 􀜳 and 􀜴 of the utility function associated with DLQR. The investigation of convergence properties related to consistency, persistent excitation and polarization of the RLS estimator is performed. The methodology involved in a project achievements online DLQR controllers and is evaluated in a fourth order multivariable dynamic system.