Planejamentos fatoriais fracionados para análise de sensibilidade de modelos de simulação de eventos discretos

Abstract

The manufacturing system modeling through simulation is used since the early 60’s and became one of the most popular and powerful tools for analyzing complex manufacturing systems. Through system modeling is possible perform its optimization. However, the integration between optimization and simulation did not happen fast. In fact, until the end of last millennium, optimization and simulation were kept well separated but this situation has changed and, nowadays, optimization software is a component of almost every simulation package. Optimization via simulation demands a considerable computational effort since to locate the optimum solution it is necessary to verify several parameter value settings. One way to accelerate the optimization is reducing the search space by selecting the variables which comprises it, once not all variables have the same importance with respect of their effect over the model output. This research has studied the use of fractional factorial design statistic techniques to identify the more important variables from two discrete-event simulation models aiming to reduce the optimization’s space search in order to accelerate that phase. This is an applied quantitative experimental research, with explanatory objective. The tool to perform the experiments is the discrete-event simulation. The research methodology was to optimize each model by two distinct procedures. The first procedure performs a variable sensitivity analysis of the model using fractional factorial designs. After identifying the more important variables, the model’s optimization is performed using this reduced search space. The second procedure performs a straightforward model’s optimization. No study was done in this approach to determine if all model’s variables impact the same effect to the output. Finally, for each model, the amount of runs of each procedure was compared. The result of the first application appointed a 59% reduction for the amount of runs between the planned optimization and the straightforward one. The second application did not present such reduction. The main reason for this bad result in the last application is the way the system was modeled, showing the importance of planning the system’s model correctly.