Pgmpy

Added
1. `BayesianNetwork.get_random_cpds` method to randomly parameterize a network structure.
2. Faster Variable Elimination using tensor contraction.
3. `factors.factor_sum_product` method for faster sum-product operations using tensor contraction.

Fixed
1. Bug in `DynamicBayesianNetwork.initialize_initial_state`. 1564
2. Bug in `factors.factor_product`. 1565

Changed
1. Runtime improvements in `DiscreteFactor.marginalize` and `DiscreteFactor.copy` methods.

Added
1. Adds checks for arguments to `BayesianNetwork.simulate` method.

Fixed
1. Fixes TAN algorithm to use conditional information metric.
2. Speed ups for all estimation and inference methods.
3. Fix in stable variant of PC algorithm to give reproducible results.
4. Fix in `GibbsSampling` for it to work with variables with integral names.
5. `DAG.active_trail_nodes` allows tuples as variable names.
6. Fixes CPD and edge creation in `UAIReader`.

Fixed
1. Fixes `CausalInference.is_valid_backdoor_adjustment_set` to accept str arguments for `Z`.
2. Fixes `BayesianNetwork.remove_cpd` to work with integral node names.
3. Fixes `MPLP.map_query` to return the variable states instead of probability values.
4. Fixes BIFWriter to generate output in standard BIF format.

Added
1. Adds BayesianNetwork.states property to store states of all the variables.
2. Adds extra checks in check model for state names

Fixed
1. Fixes typos in BayesianModel deprecation warning
2. Bug fix in printing Linear Gaussian CPD
3. Update example notebooks to work on latest dev.

Added
1. Adds a `fit_update` method to `BayesianNetwork` for updating model using new data.
2. Adds `simulate` method to `BayesianNetwork` and `DynamicBayesianNetwork` to simulated data under different conditions.
3. Adds `DynamicBayesianNetwork.fit` method to learn model paramters from data.
4. `ApproxInference` class to do approximate inference on models using sampling.
5. Robust tests for all sampling methods.
6. Adds `BayesianNetwork.load` and `BayesianNetwork.save` to quickly read and write files.

Changed
1. `BayesianModel` and `MarkovModel` renamed to `BayesianNetwork` and `MarkovNetwork` respectively.
2. The default value of node position in `DAG.to_daft` method.
3. Documentation updated on the website.

Fixed
1. Fixes bug in `DAG.is_iequivalent` method.
2. Automatically truncate table when CPD is too large.
3. Auto-adjustment of probability values when they don't exactly sum to 1.
4. tqdm works both in notebooks and terminal.
5. Fixes bug in `CausalInference.query` method.

Added
1. Adds network pruning for inference algrithms to reduce the size of network before
running inference.
2. Adds support for latent variables in DAG and BayesianModel.
3. Parallel implementation for parameter estimation algorithms.
4. Adds `DAG.get_random` and `BayesianModel.get_random` methods to be able to generate random models.
5. Adds `CausalInference.query` method for doing do operation inference with or without adjustment sets.
6. Adds functionality to treesearch to do auto root and class node selection (1418)
7. Adds option to specify virtual evidence in bayesian network inference.
8. Adds Expectation-Maximization (EM) algorithm for parameter estimation in latent variable models.
9. Add `BDeuScore` as another option for structure score when using HillClimbSearch.
10. Adds CausalInference.get_minimal_adjustment_set` for finding adjustment sets.

Changed
1. Renames `DAG.is_active_trail` to `is_dconnected`.
2. `DAG.do` can accept multiple variables in the argument.
3. Optimizes sampling methods.
4. CI moved from travis and appveyor to github actions.
5. Drops support for python 3.6. Requires 3.7+.

Fixed
1. Example model files were not getting included in the pypi and conda packages.
2. The order of values returned by CI tests was wrong. 1403
3. Adjusted and normalized MI wasn't working properly in TreeSearch.
4. 1423: Value error in bayesian estimation.
5. Fixes bug in `DiscreteFactor.__eq__` to also consider the state names order.