Differentiable Particle Filters: End-to-End Learning with Algorithmic Priors

We present differentiable particle filters (DPFs): a differentiable of the particle filter algorithm with learnable motion and models. Since DPFs are end-to-end differentiable, we can train their models by optimizing end-to-end state estimation, rather than proxy objectives such as model accuracy. DPFs encode structure of recursive state estimation with prediction and measurement that operate on a probability distribution over states. This structure an algorithmic prior that improves learning performance in state problems while enabling explainability of the learned model. Our on simulated and real data show substantial benefits from end-to- learning with algorithmic priors, e.g. reducing error rates by ~80%. Our also show that, unlike long short-term memory networks, DPFs learn in a policy-agnostic way and thus greatly improve generalization. code is available at://github.com/tu-rbo/differentiable-particle-filters .