SIMOPTIMIZATION RESEARCH. PHASE III.

Abstract : Three techniques were developed during the Phase I study: the decentralized gradient approach (DGA), the linear response surface (LRS) approach, and the quadratic response surface (QRS) approach. The techniques are designed to be applied sequentially--each being more sensitive to parameter interactions than its predecessor but also more expensive in terms of computing time. The objective of the Phase II study was to test the general performance of the techniques with a more complex model than was used during Phase I and to determine the effectiveness of certain procedural modifications. The results of the empirical tests show that DGA converges quickly and efficiently to good solutions. The convergence of DGA appears to be somewhat slower than in Phase I; however, the tests also indicate that DGA can be permitted larger steps at each iteration. As in Phase I, LRS was able to find improvements in some of the final DGA solutions but the percentage improvements were smaller than previously experienced. The QRS procedure remains to be tested although a general integer quadratic programming algorithm, necessary for QRS implementation, has been postulated. The report contains a detailed discussion of the three techniques, the test model, the test results and their analysis, a presentation of a restricted version of the integer quadratic programming algorithm, and a summary of the Phase I study. (Author)