Today – 27 June 2006 – we have released new versions of the HUGIN Graphical User Interface (v6.7) and HUGIN Decision Engine (v6.5).
The main new features of this release are:
– Sensitivity to Evidence (SE) Analysis on discrete random variables in Bayesian networks and influence diagrams in the HUGIN Graphical User Interface.
– Improved conflict resolution dialog of the HUGIN Graphical User Interface.
– Improved value of information dialog of the HUGIN Graphical User Interface.
Hugin Graphical User Interface v6.7:
The HUGIN Graphical User Interface has been extended with support for sensitivity to evidence (SE) analysis on discrete random variables in Bayesian networks and influence diagrams. SE analysis includes determining minimum and maximum posterior beliefs, impact of evidence analysis, discrimination between competing hypotheses, what-if analysis, and sensitivity to findings:
– Determining minimum and maximum beliefs is useful for analysing the sensitivity of a hypothesis variable relative to an unobserved variable.
-Evidence impact analysis investigates the impact of various subsets of the evidence on a hypothesis by computing the normalized likelihoods of the hypothesis given the evidence.
-Discrimination between competing hypotheses is based on the calculation of Bayes’ factor. This analysis supports the identification of subsets of the evidence which discriminates between two hypotheses.
-What-if analysis investigates the impact of changing the value of an observed variable on the posterior distribution of a hypothesis variable.
-Sensitivity to findings analysis analyses the impact a single finding has on the posterior probability of a hypothesis.
The Value of Information analysis dialog of the HUGIN Graphical User Interface has been improved with new features:
-The dialog now gives a graphical representation of the mutual information score between each information variable and the target node.
-The mutual information score between the target and each information variable is compared to the entropy of the target node.
-The precision of the displayed mutual information score is sensitive to the selected precision.
The Conflict Resolution dialog of the HUGIN Graphical User Interface has been improved with a number of new features:
-The dialog now has the option of selecting the set of possible hypothesis variables.
-The dialog gives a graphical representation of the value of the conflict measure after resolution for each possible conflict resolution.
-The precision of the displayed conflict measure score is sensitive to the selected precision.
-It is possible to perform hypothesis driven conflict analysis. This enables the user to investigate the impact of individual findings on the posterior probability of the hypothesis.
-Support for tracing the source of a possible conflict has been improved. The user may compute partial conflicts for all subsets of a selected set of evidence.
The HUGIN Graphical User Interface has been improved with various new features. This includes:
-Simulation of chance variables in Run-Mode. This functionality allows the user to simulate an instantiation of all variables given the inserted evidence.
-The menu items under the “Network” menu have been rearranged. An “Analysis” menu item has been introduced.
-Monitors and node lists are now updated immediately after entering a value on a continuous chance node (as opposed to after the propagation).
-Functionality for reporting the beliefs of a selected node or all nodes to the Network Log has been included.
-The entropy of a discrete chance node is shown in the Usage Log when the node is selected in run-mode.
-The mutual information score between two discrete chance nodes is shown in the Usage Log (in Run Mode only) when selecting their connecting edge.
-Improved support for long menus (e.g. long menus may appear as a result of having loaded a large number of classes).
-d-separation analysis is now possible for NetworkModels in edit mode.
-Functionality for rearranging node states has been included.
-Functionality for printing monitor windows with the graph of a model has been included.
-It is now possible to include monitor windows when writing a model as BMP.
-It is possible to replace a parent node of a child node without losing the table of the child in the process.
-Finally, efforts have been put into improving the stability of the HUGIN Graphical User Interface.
Hugin Decision Engine v6.5:
The HUGIN Decision Engine has been extended with the following features:
-Functions for computing the AIC and BIC scores have been included. AIC and BIC are scores of comparing model quality taking model complexity into account.
-It is now possible to enter a data case as evidence using a single function. This is, for instance, useful for iterating over all data cases and computing posterior beliefs.
-Functions for getting the state index of a discrete chance or decision node corresponding to a value or a label have been included. This is particularly useful for inserting evidence on a node with interval subtype.
-Functionality for performing d-separation analysis has been included. This includes two functions for obtaining the nodes that are d-connected and d-separated to a set nodes, respectively, given a set of hard and a set of soft evidence.
-Functionality for cloning nodes and domains has been included.
In addition some minor revisions have been made to existing functionality of the HUGIN Decision Engine:
-The amount of case data which can be handled by the learning algorithms has been doubled (given the same amount of physical memory).
-The HUGIN API now supports simple labels without quotes in case and data files (as opposed to requiring labels always to be quoted).
-The EM algorithm reports the AIC and BIC scores to its log file after completion of parameter estimation.
-Default labels for Boolean nodes have been changed to “false” and “true” in the HUGIN C API and HUGIN ActiveX server.
-The h_domain_get_log_likelihood function (in the HUGIN C API and equivalent functions in other HUGIN APIs) now return the log-likelihood using the actual parameter values (as opposed to using the parameter values of the penultimate iteration when using EM).