A unified framework for the application and evaluation of different methods for neural parameter optimization

Automated parameter search has become a standard method in the modeling of neural systems. These studies could potentially take advantage of recent developments in nonlinear optimization, and the availability of software packages containing high-quality implementations of algorithms that proved useful in other domains. However, a systematic comparison of the available algorithms for problems that are typical in neuroscience has not been performed. We developed a software tool for fitting the parameters of neural models, which provides intuitive, uniform access to a variety of state-of-the-art optimization algorithms implemented by four different Python packages. We also established a set of benchmark problems of different complexity that involve a variety of widely used neuronal models. We then used our optimization tool to systematically evaluate the performance of the algorithms on our set of benchmark problems. We found that several evolutionary and related algorithms consistently provided good solutions for all of our benchmarks. However, the relative performance of the different methods, both in terms of the quality of the final result and in terms of convergence speed, depended substantially on the nature of the problem. We hope that our software tool and benchmarking results will facilitate the choice and application of the best parameter-fitting methods in neuroscientific research.