Best viewed on the website.
- Hot fixes for :py:mod:`~pomdp_py.representations.distribution.particles.WeightedParticles`.
- :py:mod:`~pomdp_py.framework.basics.Agent` can be created with a name.
- Bug fixes for release script and GHA workflow.
- Refactored
_search
in :py:mod:`~pomdp_py.algorithms.po_uct.POUCT` to be more modular - More fields made accessible for :py:mod:`~pomdp_py.algorithms.po_uct.POUCT`
- Allows initializing :py:mod:`~pomdp_py.framework.basics.Agent` and :py:mod:`~pomdp_py.framework.basics.Environment` without all models ready, and provide those models later (or update them) by
set_models
- Minor function renaming for :py:mod:`~pomdp_py.framework.basics.Option`
- Added :py:mod:`~pomdp_py.framework.oopomdp.DictState`
- Improved modularity and documentation for the RockSample example
- :py:mod:`~pomdp_py.representations.distribution.particles.WeightedParticles` now hashable if frozen.
- Changed
det_dict_hash
to not use sorting - Added a Github Action workflow for automatically building, releasing wheels and uploading them to PyPI
- Default branch renamed to
main
, anddev-latest
is the development branch. Before a release,dev-<version>
should be created based off ofdev-latest
.
- Modernize repo without introducing new features.
- Migrate to
pyproject.toml
to specify dependencies, with minimalsetup.py
to build Cython extensions. - Separate core and optional dependencies. Optional dependencies are needed only for pomdp_py.problems.
- Added CI testing to repo.
- Added pre-commit checks to repo.
- Bumped minimum Python requirement from 3.7 to 3.8 due to EOL of 3.7.
- Fix
cpdef -> cdef
to avoid installation failure after Cython 3.0.0 release (pomdp-py#30). - Added float_precision argument to to_pomdp_file (pomdp-py#29)
- Add
__init__
signature for Environment in comments to be visible in docs - Fix
s -> sp
in :py:mod:`~pomdp_py.algorithms.value_iteration.ValueIteration` (pomdp-py#20) - Allow updating rollout policy for :py:mod:`~pomdp_py.algorithms.po_uct.POUCT` and :py:mod:`~pomdp_py.algorithms.pomcp.POMCP`
- Fix in
setup.py
so that wheel builds properly. - Change set to list in
pomdp_problems.tiger.tiger_problem.py
to tame error regardingrandom.sample
in Python 3.11. - Minor bug fixes and documentation.
- Fix in :py:mod:`~pomdp_py.representations.distribution.histogram` (in
__str__
): Print all of histogram as is instead of printing top 5 to avoid confusion. - Improve documentation for the tiger tutorial; Specifically, clarified :py:mod:`~pomdp_py.framework.basics.PolicyModel` and gave a reference to :py:mod:`~pomdp_py.algorithms.po_uct.ActionPrior`.
- Built documentation with new Sphinx version (4.5.0)
- Bug fix in :py:mod:`~pomdp_py.representations.distribution.particles` (in
add
) - Added classes for tabular models :py:mod:`~pomdp_py.utils.templates.TabularTransitionModel`, :py:mod:`~pomdp_py.utils.templates.TabularObservationModel`, :py:mod:`~pomdp_py.utils.templates.TabularRewardModel`. See an example in this Github Gist for defining and solving the CryingBaby domain, a small POMDP.
- Removed dependency on
pygraphviz
; - Added
utils.debugging.Treedebugger
, which makes it easier to inspect the search tree. See :py:mod:`~pomdp_py.utils.debugging`. - Added
WeightedParticles
; RefactoredParticles
. (:py:mod:`~pomdp_py.representations.distribution.particles`) - Optionally show progress bar while simulating in POUCT/POMCP.
- Added a CLI interface to simplify running example domains, e.g.
python -m pomdp_py -r tiger
runs Tiger. - Can initialize
Environment
withBlackboxModel
. - For the
OOBelief
class inoopomdp.pyx
, nowmpe
andrandom
can take an argumentreturn_oostate
(default True), which returns a sampled state as typeOOState
. This can be useful if you would like to inheritOOBelief
and return a state of your own type when implementing itsmpe
andrandom
functions. - Added
__ne__
methods to framework classes. - Reorganized
util
by breaking it into different modules. - Code refactoring: Remove unnecessary
*args, **kwargs
in interface signature. Backwards compatible. - Bug fix regarding hashing and pickling.
- Verified installation on Windows (TODO)
- Fix
setup.py
so thatpip install -e .
works.
- Edit
setup.py
file so that Cython modules inpomdp-py
can be :code:`cimport`ed.
(skipped versions due to attempting pypi release)
- Fix documentation (external solver examples).
- Update
tiger_problem.py
to match documentation
pomdp_py/algorithms/value_function.py
:- zero-probability observation should be skipped.
- refactored so that
value()
can take either a dict or a sequence of arguments.
- Available on PyPI
.value
field of VNode is instead changed to be a property, computed by finding the maximum value of the children Q-Nodes.
- Bug fix in
solvers.py
andconversion.py
(18fc58e0, cfc88e8d8) - Bug fix in the MOS domain's observation model (719c2edf5)
- Linked docker image in documentation issue #13.
- Updated documentations
- Updated documentation for external library interfacing. Added citation.
- Resolved issue #10. Set value in V-Node to be the max among its children Q-Nodes. Initial V-Node value set to negative infinity.
- Avoid search tree building during rollout (thanks Jason)
- Documentation clarification about
.sample
and.argmax
functions in theRewardModel
. - Small pomdps (with enumerable state, action, observation spaces)
defined in
pomdp_py
can be converted to.pomdp
and.pomdpx
file formats. - Added interfacing with pomdp_solve and tested.
- Added interfacing with sarsop and tested.
- Added
utils/templates.py
that contains some convenient implementations of the POMDP interface. - Bug fixes (in
histogram.pyx
)
- Fixed preferred rollout and action prior implementation; Previously the initial visits and values were not applied.
- Fixed UCB1 value calculation when number of visits is 0; Previously a divide by zero error will be thrown. But it should have infinite value.
- Fixed another potential math domain error due to log(0) in UCB1 value calculation when initial number of visit set to 0.
- Fixed bug in particle belief update (minor type error)
- Simplified the Tiger example code, updated the Tiger example in documentation.
- Fixed bug in ValueIteration and verified in Tiger that it's able to differentiate differentiate between listen/stay actions when horizon = 3, but not so when horizon = 1 or 2. The same behavior is observed using the pomdp solver by POMDP.org.
- Added an exact value function in
pomdp_py.algorithms.value_function
. It is a simpler exact value iteration algorithm. - Added Load/Unload domain Pull request #9
- Pull request #11
- Added
num_sims
parameter to POMCP/POUCT that allows specifying the number of simulations per planning step (Previously onlyplanning_time
was available. - Added cythonized versions of tiger and rocksample domains which are much faster.
- Added continuous light-dark domain. A solver (BLQR) is attempted but not ready yet.
- Bug fix in 2D MOS domain rollout; action step size changeable
- Added Tag domain, tested with POUCT random rollout
- Documentation
- Rewritten with cleaner interfaces and Cython integration
- Refactored POMCP: Now it extends POUCT which does not assume particle belief representation.
- Included Tiger, RockSample, and a 2D multi-object search (MOS) domain.
- Using Sphinx Documentation
- Implementation of POMCP and OO-POMDP