Optimal control, statistics and path planning (Q5931756)

The paper deals with the connection between spline techniques and linear dynamical systems of form \({\mathcal S}= \{\dot x= Ax+ bu,y= cx\}\), where \(x\) is a vector, \(u\), \(y\) are scalar functions and \(A\), \(b\), and \(c\) are constant matrices such that \(cb=cAb= cA^2b=\cdots= cA^{n-2}b= 0\). Note that the general solution of \({\mathcal S}\), with initial condition \(x_0= 0\), is \[ y(t)= L_t(u):= \int^T_0 g_t(s) u(s) ds,\tag{1} \] where \(g_t(s)= c\exp(A(t- s))b\) if \(t> s\), and zero otherwise. The main purpose of this work is to give a unified treatment to a series of eight problems which come from optimal control theory, statistics, and numerical analysis. Let \(J(u)= \int^T_0 u^2(s) ds\). The article starts with the problem of finding a control law \(u_0(t)\), such that \(J(u_0)= \min_u J(u)\), subject to the constraints \(y(t_i)= L_{t_i}(u)\), \(i= 1,\dots, N\), with \(u\in L_2[0, T]\). The optimal control is assumed to be in the finite-dimensional space \(\text{span}\{g_{t_i}\}\). Thus \(u(s)= \sum^N_{i= 1} \alpha_i g_{t_i}(s)\), for some real numbers \(\alpha_i\). Equation (1) yields a matrix equation \(\widehat y= G\widehat\alpha\). The condition number of the matrix \(G\) can be improved by considering smooth piecewise functions instead of the functions \(g_{t_i}\). A similar situation is presented in Problem 2, where the constraints are given in terms of inequalities \(a_i\leq L_{t_i}(u)\leq b_i\), \(i= 1,\dots, N\). The solution of this second problem can be obtained following a method inspired in the classical scheme of interpolating splines with constraints. The following six problems consider other types of functionals and constraints.