hybrid-aif#
hybrid-aif is a JAX library for active inference over models that mix continuous and discrete
variables. A model is declared as a System of state factors, observation modalities and control
nodes, with mixture and switching structure available, and can be inferred, planned, learned and
composed with other Systems. Planning runs as parallel policy search, vectorised with JAX over policies and over agents, and
objectives are assembled from named terms rather than fixed.
uv venv && uv pip install -e "." # from a checkout
The requirements are jax>=0.11, equinox>=0.13.8 and numpy>=2.0.
Elsewhere: the project page, the demo notebook, and the longer guides docs/capabilities.md and docs/composing-systems.md in the
repository.
Declaring a model
Composing and running
Inference, learning and planning
Objectives
Records and descriptions
Environments
Reference models