Results of a simulation experiment for Markov processes identification methods comparison

A numerical technique for the identification of discrete-state continuous-time Markov models has been developed and described in paper [6]. The given method is characterized by utilizing both initial approximations derived from the observed data and sensi-tivity estimates of the minimized criterion to small variations of identified parameters.

Proposed approach is experimentally compared with a first order gradient method with adaptive steps [1], which is traditionally applied for solving multivariate nonlinear optimization problems. In order to assess computational efficiency of suggested algo-rithms a set of models was generated, which are structurally similar to models commonly used in practice.

Essential structure of these models corresponds to the non-homogeneous birth-death process. Details of data generation stage are presented in Section 1 whereas specifications of a computing machine used in the experiment are listed in Section 2.

Results of the computational experiments with algorithm-specific settings presented in Section 3 are described in Section 4. The results obtained demonstrate that, in the considered range of numbers of identified parameters, the developed modified identifi-cation method provides substantial speedup compared to the classical gradient method.