The performance of cognitive models often depends on the settings of specific model parameters, such as the rate of memory decay or the speed of motor responses. The systematic exploration of a model's parameter space can yield relevant insights into model behavior and can also be used to improve the fit of a model to human data. However, exhaustive parameter space searches quickly run into a combinatorial explosion as the number of parameters investigated increases. Taking an established instance-based learning task as example, we show how simulation using parallel computing and derivative-free optimization methods can be applied to investigate the effects of different parameter settings. We find that both global optimization methods involving genetic algorithms as well as local methods yield satisfactory results in this case. Furthermore, we show how a model implemented in a specific cognitive architecture (ACT-R) can be mathematically reformulated to prepare the application of derivative-based optimization methods which promise further efficiency gains for quantitative analysis.