A conditional invertible neural network-based particle filter

This study proposes a neural network-based approach to improve the efficiency of particle filters. A conditional invertible neural network (CINN) is employed to approximate the optimal proposal distribution in a data-driven manner. The CINN is trained offline to enhance the accuracy of conditional state inference, thereby enabling the generation of high-quality particles and improving the overall filtering performance. During online filtering, the CINN serves as the proposal mechanism for particle updates. The architecture is designed based on two key principles: (1) enabling efficient particle propagation and (2) allowing for closed-form computation of particle weights. The effectiveness of the proposed particle filter framework is validated through three numerical examples.