Reduction of dimensionality in choosing optimal control (Q4797964)

Consider the linear time-varying system, \[ dx/dt= F(x,u,t)= A(t)x+ B(t)u, \] where \(u\) is the vector of the control inputs \(u= (u_1,u_2,\dots, u_p)\) and \(x= (x_1,x_2,\dots, x_n)\) represents the trajectory of the above dynamical system. Suppose that a certain cost is being paid and this fact is described in terms of an integral quadratic functional based upon the quadratic function \(Q(x,u,t)\) of the form, \[ Q(x,u,t)= x^T G(t)x+ x^TM(t)u+ u^TM(t)x+ u^TH(t)u. \] Here, \(A(t)\), \(B(t)\), \(G(t)\), \(M(t)\) and \(H(t)\) are piecewise continuous functions of the variable \(t\), bounded on a finite interval \([t_0,T]\) and \(H(t)\) is also assumed to be a symmetric positive definite matrix. Denote by \(W(x,t_0)\) the minimal cost and assume that there exists an optimal input \(u^*(t)\). Under these assumptions, it holds \[ u^*= -H^{-1}(B^T Px+ M^Tx), \] where the matrix \(P(t)\) is the solution of a certain nonlinear matrix differential equation of Riccati type. In this respect, computational drawbacks arising when obtaining \(P(t)\) are to be overcome by introducing the pseudo-inverse matrix \(D^+(t)\) instead of the matrix \(D(t)\) such that \(D(t)= P^{-1}(t)\) and consequently proving that the optimal control input is now given by \[ u^*(t)= -H^{-1}(t)[M^T x+ B^T D^+(t)x]. \] From this result it is also demonstrated that the optimal control is only based on the principal components of \(x(t)\). Finally, a gradient method of searching for the close-to-optimal trajectory and control input in nonlinear control problems is also considered.