A comparison of frequency domain methodology and conventional factor screening methods (Q1904757)

The author considers a simulation model with \(P\) continuous input factors and a scalar output response defined by \[ Y_j(t)= \sum^P_{p= 1} \sum^{q_r}_{r= 0} \beta_p(r) X_{jp}(t- r)+ \varepsilon_j(t) \] for \(t= 0, 1,\dots, N- 1\), \(j= 1, 2, \dots, K\) and \(q_p< N\) for \(p= 1, 2,\dots, P\). The index \(t\) counts the number of observations generated within a simulation run and \(N\) is the total number of observations from a run. The index \(j\) counts the number of simulation runs and \(K\) is the total number of runs. The simulation output for the \(t\)th observation on the \(j\)th simulation run is represented by \(Y_j(t)\) and \(X_{jp}(t)\) represents the value of the \(p\)th input factor for the \(t\)th observation on the \(j\)th simulation run. The quantity \(\beta_p(r)\) is the \(r\)th coefficient for the \(p\)th input factor. The quantity \(q_p\) represents the lag length for the \(p\)th input factor. The term \(\varepsilon_j(t)\) represents a zero mean, covariance stationary process with autocovariance \(\gamma_\varepsilon(h)= \text{Cov}(\varepsilon_j(t), \varepsilon_j(t\pm h))\) for all \(j= 1,2,\dots, K\). Aside that, the conventional two-level experimental design is implemented across independently seeded simulation runs or across subruns (batches) which are assumed to be independent. Factor levels are fixed at one of two levels during each run and changed across runs. Factors are generally scaled so that each factor has two levels: \(+ 1\) and \(-1\). The main aim is to compare the large sample size cost versus power of the test comparison of frequency domain methodology and a fractional two- level factorial design. The comparison is based on the models described above.