A dynamic system relationship reasoning method based on bidirectional attention

Abstract

This invention discloses a method for inferring relationships in dynamic systems based on bidirectional attention, relating to the fields of artificial intelligence, deep learning, and complex network modeling. The relationship inference model includes a graph generator and a neurodynamics surrogate model. The graph generator generates interaction structures between nodes based on structural parameters and represents them using an adjacency matrix. The neurodynamics surrogate model models the dynamic system based on node trajectory data under a given adjacency matrix. The training of the relationship inference model involves training the graph generator to generate the adjacency matrix while simultaneously training the neurodynamics surrogate model using unsupervised learning to maximize the likelihood of observed node trajectory data under the adjacency matrix. The total loss function includes a likelihood-based dynamic system prediction loss term and a regularization term applied to the inferred interaction structure. This invention is used to infer potential inter-node interaction structures in dynamic systems.

Images