Multilevel dynamic programming for general multiple linear-quadratic control in discrete-time systems (Q1327182): Difference between revisions

This paper extends the reach of dynamic programming to general multiple linear-quadratic control problems in discrete-time systems, where the overall objective \(J\) is a nonlinear function of multiple quadratic performance indices. The problem is non-separable in the sense of dynamic programming and is not in the form ready to be handled by the discrete maximal principle, if \(J\) is not of additive form with respect to all performance indices \(J_ i\) \((i= 1,2,\dots,k)\). The author shows that a control law which is linear with respect to the current state vector can be obtained through solving an auxiliary Lagrangian problem of a parametric linear-quadratic structure. To this end, a multiple dynamic programming solution procedure is proposed. At the lower level dynamic programming, an analytical solution of the auxiliary Lagrangian problem is generated. The upper level in turn adjust iteratively the value of the weighting vector \(\lambda\) for the Lagrangian problem. This two level process repeats until an optimal stop condition is satisfied. The overall objective functional \(J\) becomes a function of \(\lambda\) on substituting \(J_ i(\lambda)\) into \(J\). An alternative condition given by the first author for the classical Kuhn-Tucker condition is used to solve the minimization problem. A numerical example is given in detail for illustration.