Today we are releasing a new version of the HUGIN software (7.3). The main new feature of this release is the introduction of a new node type called the “function” node type.
A function node represents a numerical value computed from the results of a propagation in a network. A function node allows the user to define mathematical expressions representing a value computed after a successful propagation (or simulation) in a network. The function node can have continuous nodes, discrete chance nodes, and utility nodes as parents. The expression is defined using the Table Generator.
The development of the functionality concerning the function node type (introduced in HUGIN 7.3) has been sponsored by Danish mortgage credit institution Nykredit Realkredit (www.nykredit.dk).
HUGIN Graphical User Interface 7.3
The HUGIN Graphical User Interface has been improved with various new features. This includes:
– A dialog for determining the number of states of a hidden node has been introduced. The number of states is determined by iterating over all state space sizes from a minimum to a maximum value set by the user. The user can select from a number of different scoring methods. For each iteration a parameter estimation is performed and the user has the option to set the number of states of the hidden node to the best scoring option.
– A dialog to perform feature selection has been introduced. The feature selection is performed relative to a target node. For each node represented in data, a score of the correlation between the node and the selected target node is computed. The result is a list of the nodes sorted according to the score.
– The Chow-Liu algorithm has been included in the Learning Wizard. The Chow-Liu algorithm allows the user to learn a tree structure over the discrete chance nodes represented in data. The resulting tree is a maximum-likelihood estimate of the unknown distribution generating the data (under the tree structure constraint).
– A tree-augmented naïve Bayes model (TAN) learning algorithm has been included in the Learning Wizard. The TAN learning algorithm is based on the Chow-Liu algorithm. It learns a tree structure over the attribute nodes of a TAN model given a target node.
– The support for Object-Oriented Networks (OONs) has been improved with new functionality for navigating OONs in run-mode. This includes support for opening monitor and policy windows of nodes inside an instance node and traversal of instance nodes.
– Vector Quality PDF export for network structure.
– The adaptation wizard now supports the inspection of probability distributions for nodes not subject to adaptation
– When loading a case from file the probability of the evidence in the case is reported to the status bar
– Other minor improvements.
Finally, efforts have been put into improving the performance of the HUGIN Graphical User Interface.
HUGIN Decision Engine 7.3
The HUGIN Decision Engine has been extended with the following features:
– A new node type has been introduced: The “function” node type. This node type is used to express calculations to be performed using the results of inference or simulation as input. However, function nodes are not involved in the inference process itself, so evidence cannot be specified for function nodes.
– Here is a list of changes related to the introduction of function nodes:
o The category of function nodes is “h_category_function”, and the kind is “h_kind_other” (this kind is also used for utility and instance nodes).
o A real-valued function is associated with each function node. This function is specified using a model.
o The function associated with a function node must only depend on the parents of the node. The parents can be of any type of node (except instance nodes).
o The value of a function node based on the results of inference is requested by the “h_node_get_value” function.
o The value of a function node based on the results of simulation is requested by the “h_node_get_sampled_value” function.
o The “h_domain_new_node”, “h_node_clone”, and “h_node_delete” functions do not uncompile if a function node is involved.
o The “h_node_add_parent”, “h_node_remove_parent”, and “h_node_switch_parent” functions do not uncompile if the child node is a function node.
o The HKB format has been updated to support networks with function nodes (but the old format is still used for networks without function nodes).
o The NET language has been extended: The new “function” keyword denotes the definition of a function node, and “potential” specifications are now also used to specify links and models for function nodes.
– The results of simulation are now invalidated by (implicit as well as explicit) uncompile operations.
– The new “h_node_get_sampled_utility” function requests the “sampled” utility of a utility node.