Optimal input design for identification of compartmental models. Theory and application to a model of glucose kinetics (Q1068013)

The authors study the design of optimal inputs for the identification of compartment models: \[ \dot x(t)=A(p)x(t)+Bu(t),\quad x(0)=0,\quad y(t)=C(p)x(t), \] \[ z(t)=y(t)+e(t),\quad h[x(t),u(t),p]\geq 0, \] where x(t) is the (n\(\times 1)\) state-variable vector, u(t) is the scalar input, y(t) is the scalar measurable output, z(t) is the noise corrupted continuous-time measured output, and e(t) is the measurement noise, assumed to be additive Gaussian with statistics \(E[e(t)]=0\), \(E[e(t)e(\tau)]=r(t)\delta (t-\tau)\), where r(t) is the error variance and \(\delta\) (t-\(\tau)\) the delta Dirac function. A is the (n\(\times n)\) state compartmental matrix, whose elements are related to the transfer- rate parameters \(k_{ij}\) by \[ a_{ij}=k_{ij}\geq 0,\quad i\neq j,\quad a_{ii}=-\sum^{n}_{j=0,j\neq i}k_{ij},\quad k_{0i}\geq 0. \] B is the (n\(\times 1)\) input location vector; C is the (1\(\times n)\) output vector; p is the unknown parameter vector; h denotes all additional a priori equality- and inequality constraints available on the system. Assuming the model structure is well posed, the optimal input problem can be formulated as follows: design the form of test input u(t) so that maximum-precision estimates of the unknown model parameters are obtained from the noisy data z(t). In this paper, theory and computational methods for the scalar case of the above problem as well as the optimal input results for a glucose kinetics model [cf. the first author, \textit{G. Toffolo} and \textit{E. Ferrannini}, ibid. 72, 291-315 (1984; Zbl 0557.92007)] are presented.