Release 8.3

February 2016

The main new feature of this release is improved support for Dynamic Bayesian networks. This includes an algorithm for approximate belief update in a Dynamic Bayesian network through approximation of the joint probability distribution over the temporal clones and a number of extensions to the Data Frame window for data processing and model evaluation involving data sequences.

HUGIN Graphical User Interface v8.3

The HUGIN Graphical User Interface has been improved with various new features. These include:

  • Inference in a Dynamic Bayesian network can now be performed using an approximate algorithm. The approximation is supported for prediction and when moving the time-window. Belief update inside the time window is exact. This means that the approximation is mainly relevant for 1.5 time-step Dynamic Bayesian networks with a large joint probability distribution over the temporal clones. It is possible to select between exact and approximate inference.

  • The Data Frame window has been improved with a number of new features to support evaluation of (mainly) Dynamic Bayesian networks on data sequences. This includes functionality to mark a column as sequence identifier when a data file contains more than one sequence. Functionality for propagating cases, manipulating data, and evaluating dynamic models are sensitive to the specification of a sequence identifier. There is, for instance, functionality to transform data sequences into single cases and to inspect the results of sequence classification in more detail.

  • It is now possible to change the value of a real-valued function node without parents in Run-mode.

  • It is now possible to run the OOBN EM parameter estimation on a Dynamic Bayesian Network. This is useful in combination with the improved support for handling data sequences in the Data Frame window.

  • A Structure Learning Wizard has been added to the list of Wizards. This makes it more easy to construct the structure of a Bayesian network without going through the parameter estimation step of the Learning Wizard.

  • The Learning Wizard and the Structure Learning Wizard have an optional step for feature selection. This step supports feature selection on discrete variables relative to a specific target variable and is useful when building classification models from data.

  • The DBN functionality has been extended with a fast forward button making it easy to advance a DBN a specified number of steps.

  • The DBN functionality has been extended with a menu item to open or close all monitors for a node across time-slices.

  • It is now possible to expand / collapse recursively a single instance node in Run-Mode.

  • The functionality for exporting beliefs to a Data Frame dialog now shows the mean and variance for CG nodes.

  • The same Java code is now used to generate both the 32-bit and the 64-bit versions of the HUGIN Graphical User Interface.

  • Toggle Data Frame Window from a frame inside the HUGIN GUI Frame to a freely floating window.

  • Other minor improvements.

Finally, work has been done to improve the performance of the HUGIN Graphical User Interface.

HUGIN Decision Engine v8.3

The HUGIN Decision Engine has been extended with the following features:

  • The Application Programming Interfaces (APIs) for the HUGIN Decision Engine has been extended with a new data type for representing a data set. This includes supporting functionality to manipulate and use the data set. This is useful for reading, storing and manipulating data that does not necessarily match the requirements of existing API functions for, for instance, parsing data from a file.

  • The HUGIN Decision Engine supports approximate inference in Dynamic Bayesian Networks. The approximation is supported for prediction and when moving the time-window. Belief update inside the time window is exact.

  • The HUGIN Web Service API has new widgets for the deployment of Dynamic Bayesian Networks.