The R Package optimization: Flexible Global Optimization with Simulated-Annealing

Standard numerical optimization approaches require several restrictions. So do exact optimization methods such as the Linear Programming approach appeal for linearity and Nelder-Mead for unimodality of the loss function. One method to relax these assumptions is the Simulated Annealing approach, which reduces the risk of getting trapped in a local optimum. However, the standard implementation still requires regular parameter spaces and continuous loss functions. To address this issue, we implemented a version of the Simulated Annealing method that is able to deal with irregular and complex parameter spaces as well as with non-continuous and sophisticated loss functions. Moreover, in order to gain fast but reliable solutions, we included steps to shrink the parameter space during the iterations. All these steps are summarized in the R package optimization, which we will introduce in the following article. We also included generic and real world applications in order to test our approach.