pypomp 0.0.4

Modeling and inference using partially observed Markov process (POMP) models

pypomp

Python code for modeling and inference using partially observed Markov process (POMP) models.

Expected package users

• Scientists wanting to perform data analysis on a dynamic system via partially observed Markov processes (POMPs), also called state-space models (SSM) or hidden Markov models (HMM) in other contexts.

  • This package design and intended use is similar to the popular pomp R package. As such, many of the expected use cases and motivating examples of this package can be found on the pomp package bibliography page.

• Researchers wishing to develop novel inference methodology for POMP models.

  • Like the pomp R package, this package provides a framework for implementing computer representations of arbitrary POMP models. This ability provides an environment for researchers to develop, test, and deploy novel algorithms that are applicable to POMP models.

Key features

• Estimation, filtering, and inference for highly nonlinear, non-Gaussian state space models via the particle filter.

• New algorithms for model-fitting. Gradient descent using a new gradient estimate initialized with a warm-start allows for improved maximum-likelihood inference in even highly challenging epidemiological models, while the gradient estimate can readily be plugged into a sampler from Tensorflow Probability to facilitate more efficient Bayesian inference.

• This package leverages JAX for GPU support and just-in-time compilation, enabling a speedup of up to 16x when compared to the pomp R package.

Package Development

• The pypomp package is currently at early stages of development. All version numbers below 0.1 are pre-release.

• All contributions are welcome! Contributions should keep in mind the intended uses of this package, and its intended users.

• The pypomp package is run by the pypomp organization.