Evaluation of alternative discrete-event simulation experimental methods

• Main Author: Warn, Alan James
• Published: Bournemouth University 2003
• Subjects: 519; Computer Science and Informatics
• Online Access: http://ethos.bl.uk/OrderDetails.do?uin=uk.bl.ethos.274134

The aim of the research was to assist non-experts produce meaningful, non-terminating discrete event simulations studies. The exemplar used was manufacturing applications, in particular sequential production lines. The thesis addressed the selection of methods for introducing randomness, setting the length of individual simulation runs, and determining the conditions for starting measurements". Received wisdom" in these aspects of simulation experimentation was not accepted.The research made use of a Markov Chain queuing model and statistica analysis of exhaustive computer-based experimentation using test models. A specific production-line model drawn from the motor industry was used as a point of reference. A distinctive,quality control like, process of facilitating the controlled introduction of "representative randomness" from a pseudo random-number generator was developed, rather than relying on a generator's a priori performance in standard statistical tests of randomness. This approach proved to be effective and practical. Other results included: The distortion in measurements due to the initial conditions of a simulation run of a queue was only corrected by a lengthy run and not by discarding early results. Simulation experiments of the same queue, demonstrated that a single long run gave greater accuracy than having multiple runs. The choice of random number generator is less important than the choice of seed. Notably, RANDU (a "discredited"MLCG) with careful seed selection was able to outperform in tests both real random numbers, and other MLCGs if their seed were chosen randomly,99.8% of the time. Similar results were obtained for Mersenne Twister and Descriptive Sampling.Descriptive Samnpling was found to provide the best samples and was less susceptible to errorsin the forecast of the required sample size. A method of determining the run length of the simulation that would ensure the run was representative of the true condifions was proposed. An interactive computer program was created to assist in the calculation of the run length of a simulation and determine seeds so as to obtain" highly representative" samples, demonstrating the facility required in simulation software to support theses elected methods.