filterpy

• added a fast inverse of a diagonal matrix function common.inv_diagonal()

• added ability to order Q matrix by dimension or derivative with the helper functions

• mahalanobis, log-likelihood, and likelihood only computed when requested; compute_log_likelihood attribute removed.

• fix for bug where attributes not saved. GitHub #136

• Many minor typos in docstring and comments fixes

• Added saving of posterior state to classes

• Added exception to IMM if filter bank dimensions do not agree. GitHub #141

• Added Mahalanobis distance to classes that didn't support it

• added flatten() to saver as an convienent way to flatten column vectors so they will display nicely in the REPL.

• Partially removed dependence on Matplotlib. Matplotlib is only imported by functions that use plotting. Github #148

• Removed ability to pass in single matrix to batch_filter functions/methods. Github #147. In general it is impossible to tell if you have a list of matrices, or an single array, depending on the shape of the arrays and the size of the input. Rather than try and obscurely fail, make the user pass the correctly dimensioned data in.

• Github #151 - added ability to specify unscented transform function for the UKF RTS smoother.

• Compute log-likelihood and mahalanobis only when requested GitHub #139. Some, but not all classes did this already; I normalized the behavior across all classes.

• added .gitattributes to force all files to use unix-style line endings

• GitHub #138. Attributes were not being set when z == None on call to update(), meaning things like log-likelihood and mahalanobis had incorrect values.

• GitHub 125: used itertools.repeat() instead of [x]*n to get iterable list of constant values

• Fixed bug using [:] to copy data instead of np.copy(), which is a deep copy

• fixed import bug in common.kinematic_kf()

Fixed build error in Python 2.7 due to using print function without importing it from future.

Added filterpy.common.Saver class, which can save all the attribute of any filtering class. Replaces KalmanFilter.Saver, which only worked for the KalmanFilter class.

Added optional parameter specifying a Saver object to be passed into all of the batch_filter() functions/methods.

Added attribute z to most of the filter classes. This is mostly so the

Changes to documentation - mostly making it more consistent.

• #49 Added tests to distribution, as they contain a lot of examples

• #113 added plotting of 3D covariance ellipsoid with plot_3d_covariance
• Fixed bug where multivariate_gaussian accepted negative covariances
• #108 used pylint symbolic names
• Got code in compliance with pylint
• Fixed #105 - this was just test code, and it turns out the code as was was correct, so I deleted the second return statement
• #88 fixed HInfinity.batch_filter to not use R

• BUG: UKF smoother had an important bug in it. DO NOT use the rts_smoother prior to this release! GitHub issue #97

• Added jerk option to the computation of Q. GitHub issue #83

• I pushed a broken version! This should fix that. Do not use 1.2.2.

• This push sucks in some minor documentation and PEP8 changes to a few files.

• changes log-likelihood and likelihood back into attributes/properties and added to several classes so the IMM code works uniformally across all

• bug fix in gh code that ignored k (submitted by Lucal Beyer)

• doc block fix by Andriy Teraz

• Added computation of mahalanobis distance in stats
• Added better default values for F and H in KalmanFilter
• Fixed bugs due to calling .dot on float64 values
• Removed all of the getter/setters as unpythonic
• added function to create kinematic filters
• changed log likelihood and likelihood to methods to avoid the slow computation it requires.
• Improved documentation of method parameters

• fixed bug in likelihood to never return 0, which can happen due to floating point underflow

• Added a Saver class to save the state of the KalmanFilter in arrays. Makes it easy to save the intermediate values while performing filtering.

• Fixed bug where likelihood() returned None

• Altered RTS_smoother algorithm to also return the predicted covariances

• Mark as first stable release
• Fix #42: For conda-build to work, correct setup.py and MANIFEST.in path to LICENSE file, ensuring they are distributed to pypi.

• Fix #53: UKF rts_smoother does not use residual functions
• Fixed #54: Comments in multivariate_multiply incorrectly called the covariance the mean.
• Added logpdf to stats. Computes logpdf - mostly a wrapper around stats.multivariate_normal.logpdf as older versions of that function do not support the allow_singular keyword. But it also flattens out the vectors for you so you do not have to do anything special with column vectors.

• Added Cubature Kalman filter.
• Bug in Q_continuous_white_noise(). The first term in the matrix should be (dt3)/3, not (dt4)/3.
• Added log-likelihood computation to UKF.
• Added simplex points for UKF
• fixed bug in KF matrix size check

• Github issue #37. rts_smoother uses the wrong index for F and Q: they should use k+1, not k. This caused poor smoothing performance when either F or Q are time varying.

• Github issue #40. Fixed behavior of multivariate_gaussian to accept list as the covariance matrix.

Modified the update() and predict() functions to work in the univariate case. You can pass int/floats into the equations and get floats back.

• Added discrete_bayes module, which supports discrete Bayesian filtering.

• Brought docstrings (mostly) into compliance with NumPy documentation style. This requires installation of numpy doc with pip install numpydoc

docs\conf.py has been modified to use numpydoc.

Move to minor version numbering doesn't mean anything other than it got absurd to be using 3 digits for version numbers. We are far past alpha here. I will be moving to 1.0.0 soon, probably after I finish the book and flesh out a few points.

• Implemented a fixed-point smoother, but it is not working all that well.

Color on this: There are various recusive equations for the fixed point filter that I have found in various book - Simon, Crassidis, and Grewal. None seem to work very well. I have code that works pretty good when R is < 0.5 or so, but then the filter diverges when R is larger. I'm not seeing much in the literature that explains this very well, nor any evidence of this smoother actually being used in practice. I will give this a bit more effort, and if I can't get something reliable I'll put it in a branch and remove from trunk. Someone will have to tackle this on a rainy day.

• KalmanFilter.batch_filter() now accepts lists of all the KF matrices

• lots of docstring corrections and additions

• Deprecated plot_gaussian in favor of plot_gaussian_pdf, which is a more descriptive name.

• Added plot_gaussian_cdf and plot_discrete_cdf.

• Added function to compute update in the presense of correlated process and measurement noise.

• Added IMM filter.

• added tests for IMM and MMAE filters

• Added display of semi-axis for covariance ellipses

• various bug fixes

• Added likelihood and log-likelihood to the KalmanFilter class.

• Added an MMAE filter bank class.

• Added function to compute NEES

Installation still messed up, this is a revert to 0.0.23 minus the folder changes. I hope.

BREAKING CHANGE

Split statistical functions in filterpy.common into filterpy.stats module. I did not add or change anything, just move functions. If you get an import error, this is probably why! Switch import from filterpy.common to filterpy.stats and everything should work.

Added monte_carlo module which contains routines for MCMC - mostly for particle filtering.

Several important bug fixes and additions for the UKF filter. It is very important to update your code to this release if you are using the UKF.

• You couldn't call update() more than once in a row or the covariance matrix would be computed incorrectly,.

• Added way to specify subtract routine in the sigma point classes.

• Fixed bug in computation of weights for the Julier sigma points.

BREAKING CHANGES!!

The unscented kalman filter code has been significantly altered. Your existing code will no longer run. Sorry, but it had to be done.

As of version 0.0.18 there were separate classes for the UKF (Julier's) original formulation, and for the scaled UKF. But they are all the same thing, basically, and there were differing levels of support - the scaled version didn't have an RTS smoother, for example.

Now the sigma point and weight generation is done with a separate class, and the UKF class just performs the algorithm. This is much more configurable at perhaps the cost of being a bit harder to read and learn. But I didn't want to keep writing batch_filter, rts_smoother, etc, for every possible sigma point filter.

The best documentation on this is the chapter on the UKF in my Kalman filter book:

https://github.com/rlabbe/Kalman-and-Bayesian-Filters-in-Python/blob/master/09_Unscented_Kalman_Filter.ipynb

• Added args parameters to Hx and HJacobian of the ExtendedKalmanFilter class so you can pass additional data to them.

• Made an exception more human readable by including the size of the matrix that caused the shape error.

• Fixed assert in UKF module that incorrectly required kappa to be >= 0.

• Added multivariate_multiply to stats module.
• IMPORTANT: bug fix in the UKF RTS smoother routine.
• various typo fixes.

A bunch of small changes and bug fixes. Documentation improvements.

Note: Dumbness with git caused me to blow away all of my tags. I had a tag for each release, but did not realize pushing the project did not push the tags. A bit of hard drive stupidity, and poof. So, no good history for past releases.

These tag numbers correspond to the version number of the project on pypi. If you do pip install filterpy you will (as of now) get version 0.0.15 - this tag.

The change to _dt was stupid in 0.0.13 . I put it back to _dt, and then added an optional dt parameter to the predict() function.

• BREAKING CHANGE: _dt in UKF is now named dt to allow users to rename. You will get an exception if you try to use _dt for now.

• fixed bug in EKF.

• Mostly a change in the pypi install so that the pip install will include the test directories, and include the changelog and license.

• a few small bug fixes.

• Modified all filters to allow a 1D array for the state vector x. That is, np.array([1,0]) is allowed, as well as np.array([[1],[0]])

This is a potentially breaking change to your scripts. I tried to test all of the possibilities, but bug may remain.

• Added some tests for dimensionality of input to functions. It is far from complete, as I don't want to go overboard running tests for every function call. On the other hand, failures are obsucre. This will be finalized in few releases.

• Added Ensemble Kalman filter
• bug fixes in UKF

Minor changes to Unscented filter, mainly naming of local variables.

Significant changes to Unscented filter. Now separate classes for the different sigma computations, and predict/update split out. Provision for supplying your own residual and unscented transform functions.

• Fixed and included the fixed lag smoother algorithm.
• name change - all Z and Zs to z and zs. They are vectors, not matrices.
• Optional H parameter in KalmanFilter.update() to override the H matrix. Useful if you have a variable number of measurements on each update.

• Tests and fixes for the ExtendedKalmanFilter
• Minor name changes for the methods that compute Q in common

• Reverted the name change of .x to .X in the various classes. I have no idea what I was thinking - x is a vector, so it should be lower case.

• Moved some code to a new /examples directory to reduce clutter. It is worth noting that the code in there does not run now - it is based on the old procedural unscented KF code, not the new OO based code. However, the test_UKF.py code basically implements this example as a test using the new code. This is more a change for the future.