Residual-Based Multi-peak Sampling Algorithm in Inverse Problems of Dynamical Systems

Abstract

Stochastic differential equations can describe a wide range of dynamical systems, and obtaining the governing equations of these systems is the premise of studying the nonlinear dynamic behavior of the system. Neural networks are currently the most popular approach in the inverse problem of dynamical systems. In order to obtain accurate dynamical equations, neural networks need a large amount of trajectory data as a training set. To address this shortcoming, we propose a residual-based multi-peaks sampling algorithm. Evaluate the training results of each epoch of neural network, calculate the probability density function $P(x)$ of the residual, perform sampling where the $P(x)$ is large, and add samples to the training set to retrain the neural network. In order to prevent the neural network from falling into the trap of overfitting, we discretize the sampling points. We conduct case studies using two classical nonlinear dynamical systems and perform bifurcation and first escape probability analyzes of the fitted equations. Results show that our proposed sampling strategy requires only 20$\sim $30\% of the sample points of the original method to reconstruct the stochastic dynamical behavior of the system. Finally, the algorithm is tested by adding interference noise to the data, and the results show that the sampling strategy has better numerical robustness and stability.