Automated Parallel Optimization of Simulation Parameters using Modified Nelder-Mead Simplex Algorithm

Computational simulations used in many fields have parameters that define models that are used to evaluate simulated properties. When developing these models, the goal is to choose the parameters that best replicate a set of desired properties. Mathematical optimization methods can be used to optimize the simulation parameters by defining a function that uses simulation parameters as input and outputs a value describing how well a set of experimental properties are reproduced. Because simulated properties are often calculated using stochastic sampling methods, this optimization involves an objective function that is noisy and expensive to evaluate. Also, optimization of the simulation parameters can require running long simulations. A new method is proposed to best fit the simulation output properties that are noisy. We propose a modified parallelized Nelder-Mead (NM) simplex to allow multiple simulations to run for indefinite amounts of time until noise is adequately converged. Parallelizing the NM simplex has one more purpose: to be able to utilize a look-ahead NM simplex operation. The NM simplex was modified to include more vertices to give more information on the function surface in attempt to reduce the effect of the starting points or noise. This new method is named swarm-like-complex (SLC). The SLC optimization method can be helpful when optimizing model parameters while targeting properties with noisy sampling methods, as well as optimizing more parameters simultaneously. It also accomplishes the optimization in orders of magnitude less time than a traditional, human-guided optimization