Applying Linear Regression and Neural Network Meta-Models for Evolutionary Algorithm Based Simulation Optimization

Abstract

The increase in computing power over the last decade has led to an increase in the use of simulation programs to model real world optimization problems as well as the complexity with which these problems can be modeled. Once a model has been built, an experimental design is often used to determine the effects certain parameters have on the problem trying to determine the good settings that optimize a set of outputs. However, these problems often have a large number of variables or parameters that can be changed with wide value ranges and as these simulation models become increasingly more complex they become computationally expensive to run. Most of these problems are non-linear and may not have a true optimal solution based on the inherent variability in real-world applications and the stochastic simulation model. Evolutionary algorithms are a class of computational optimization techniques that harness the power of the computer to solve a problem. The application of evolutionary search techniques as a simulation optimization technique has yielded reasonable results. However, the algorithm can take a long time evaluating just one set of decision variables owing to replications and computational time of one simulation run and not to mention the sheer number of different sets that have to be evaluated to find good solutions for these complex problems. Linear regression and neural-network meta-models can be used to generate a surface model of the simulation. Evaluating the meta-model is very fast as compared to the simulation model. Therefore, this thesis combines the use of evolution algorithms, simulation models and meta-models to produce a more efficient simulation optimization technique. The two types of meta-models are tested to determine their effectiveness as a meta-modeling technique and the overall effectiveness of finding the best solution.