Simulation optimization by genetic search

Abstract

This is an example of simulation optimization by genetic search. We compare a new technique, genetic search, to two old techniques the pattern search and the response surface method search. The pattern search uses the Hooke–Jeeves algorithm and the response surface method search uses the computer code of Dennis Smith. This research compares these three algorithms for accuracy and stability. In accuracy we look at how close the algorithm comes to the optimum. The optimum having been previously determined from exhaustive testing. We evaluate stability using the variance of the response function determined from 50 searches; the lower the variance the more stable the response. The example tested is a university time-sharing computer system with two real decision variables: quantum, the amount of time that the computer spends on a job before sending it back to the queue and overhead, that is the time that its takes to execute this routing operation. The response is the cost of operating the system determined from a cost function.