0.0.7.1

Patched v0.0.7

What's Changed

• Add comma after Pillow requirement by @SWauthier in #110 which also addresses Issue #113

New Contributors

• @SWauthier made their first contribution in #110

Full Changelog: v0.0.7...v0.0.7.1

0.0.7

What's Changed

• plot_beliefs, plot_likelihood now functions in utils.py
• Bug fix in calculation of variational free energy during message passing routines (see #98)
• Allow optional distributional observations in agent.infer_states(obs) with a new optional boolean argument distrib_obs, for use in hierarchical set-ups
• added two new environments (envs.visual_foraging.SceneConstruction and envs.visual_foraging.RandomDotMotion)
• added functions for quickly initialising empty A and B arrays

Full Changelog: v0.0.6...v0.0.7

0.0.6

What's Changed

• Bug fixes (see #84, #86)
• Optional action sampling precision as input to Agent (equivalent of alpha parameter in SPM) to address Issue #81: #88
• Can sample directly from policy posterior using new sample_policy() function in control.py: #89
• Fixed failing documentation: #92

Full Changelog: v0.0.5...v0.0.6

0.0.5

Release notes

• bug fixes, notably fix in B matrix learning (pymdp.learning.update_state_likelihood_dirichlet()), see here
• more full documentation based on JOSS reviews, see here
• include archived version (using Zenodo) with 0.0.5 release (doi: 10.5281/zenodo.6484849), for simultaneous publication of JOSS paper

pymdp 0.0.4

Updates include:

• Read the Docs / Sphinx-based documentation, with references thereto in a cleaned-up/shortened README.md
• new epistemic chaining demo (on README.md and in documentation)
• renaming of many inference.py and learning.py functions to make them more understandable

BUG FIXES:

• corrected action sampling routine (in pymdp.control.sample_action()) so that when marginalizing posterior probabilities from each policy, per action, we only sum the probabilities assigned to action for the next / successive timestep, not for all timesteps. Before this fix, this can lead to maladaptive behavior in many scenarios where temporally-deep planning is required.

pymdp 0.0.3

Updates include:

• more demo colab notebooks now linked on main page
• model checks in when constructing an Agent() (e.g. normalization checks)
• D vector learning / bayesian model reduction
• can pass in E vector (prior over policies) to Agent() constructor
• time-varying prior preferences (i.e. C can now be a matrix rather than having to be a vector)
• updated dependencies in setup.py to allow forwards compatibility with newer versions of various packages