A Comparative Study of Stochastic Optimizers for Fitting Neuron Models. Application to the Cerebellar Granule Cell

Abstract

This work compares different algorithms to replace the genetic optimizer used in a recent methodology for creating realistic and computationally efficient neuron models. That method focuses on single-neuron processing and has been applied to cerebellar granule cells. It relies on the adaptive-exponential integrate-and-fire (AdEx) model, which must be adjusted with experimental data. The alternatives considered are: i) a memetic extension of the original genetic method, ii) Differential Evolution, iii) Teaching-Learning-Based Optimization, and iv) a local optimizer within a multi-start procedure. All of them ultimately outperform the original method, and the last two do it in all the scenarios considered.

On the Design of a New Stochastic Meta-Heuristic for Derivative-Free Optimization

“…The multi-start SASS is a especially descriptive comparison: Since Tangram can be seen as a rule-based multi-start component linked to SASS, it should serve as a better guide for this local solver rather than simply generating random starts. Finally, TLBO has been included in the comparison due to its fame of being simple to tune and effective [4]. It is also relevant to highlight that both TLBO and MSASS achieved very good results in the model tuning application described in [4], which supports their selection.…”

“…Alternatively, if the function modeled the strength of the resulting product, the points of interest would presumably be the maxima of the corresponding function. One of the applications in which optimization stands out is model tuning, where the parameters become variables, and the objective function is the comparison between the achieved and desired output [4,12]. It allows automating processes that used to rely on experts and might be biased by them.…”

“…These strategies are known as heuristics when they are problem-specific approaches and meta-heuristics of general-purpose. Similarly, the use of randomness serves to enhance the exploration capabilities at the expense of uncertainty [4].…”

“…There exist numerous population-based meta-heuristics that can be applied to the problem defined above [1,15]. Traditional genetic algorithms, Differential Evolution [4,13], and the Universal Evolutionary Global Optimizer (UEGO) [6,11] are good examples. However, they have multiple parameters, so they require fine parameter tuning.…”

“…However, they have multiple parameters, so they require fine parameter tuning. Teaching-Learning-Based Optimization (TLBO) [4,14] avoids this problem as a population-based method that only needs the population size and the number of iterations. Regardless, this sort of method will generally need numerous function evaluations to converge due to its haphazard exploration strategy.…”