Gradient estimators for parameter inference in discrete stochastic kinetic models
Researchers have developed new gradient estimators to improve parameter inference for discrete stochastic kinetic models, which are widely used in physics. Traditional automatic differentiation methods are not directly applicable to the Gillespie stochastic simulation algorithm due to non-differentiable operations. This paper explores three machine learning gradient estimators—Gumbel-Softmax Straight-Through, Score Function, and Alternative Path—to overcome this limitation. The study found that while the Gumbel-Softmax estimator often performs well, other estimators can provide more robust gradients in challenging scenarios, enabling effective gradient-based parameter inference with the Gillespie SSA. AI
IMPACT Introduces new computational techniques that could improve the accuracy and efficiency of scientific modeling.