# Release Notes¶

## 3.2.1¶

released on 2021-08-12

fixed a bug that caused long model build times for deep graphs

technical adaptions of

`show()`

function for the new halerium platform

## 3.2.0¶

released on 2021-07-01

- Time Series modeling

introducing the

`TimeShift`

operator and the`TimeIndex`

, which allow the user to breach the conditional independence along the data axis to build time series graphs, where variables can depend on their own past.

- Gaussian processes

introducing the

`construct_gaussian_process()`

factory to create correlated Gaussian processes.

- Metric Gaussian Variational Inference

improved the convergence behavior of the

`MGVIModel`

and added the methods`update_lbfgs()`

and`update_ngd()`

for better manual control of model solving.

## 3.1.0¶

released on 2021-05-07

- Causal calculus

introducing the

`do_operation()`

, which transforms a`Graph`

or other scopetor by applying the do operation to one or more of its variables.introducing the

`InterventionPredictor`

objective which calculates predictions for mixes of observations and interventions.

- added features in the
`CausalStructure`

class

- more convenient objective calls - every objective now has its own method in the CausalStructure class. The new methods are:

- CausalStructure instances can be saved and loaded as json specifications, see

- Gaussian process regression (
`gaussian_process_regression()`

)

better construction of the posterior graph for Gaussian process regression

- Bugfixes:

specification json files created in python 3.8 could not be loaded in python 3.7 due to the upgrade of the pickle version. The pickle version is now set to 4 to ensure compatibility.

fixed a bug that cause false alerts in cyclic dependency checks.

fixed a bug, where the ADVI model failed for Dirac distributed Variables.

fixed a bug in the conjugate gradient implementation that affected the performance of the MGVIModel and the MAPFisher model