Feature Lists

Feature List for HUGIN Graphical User Interface 7.4 - HGUI 7.4

Construction

Construction of Hugin Knowledge Bases utilizing Bayesian network and influence diagram models (known as domains within the Hugin Decision Engine)

• creation/deletion of domains/nodes
• adding/removing edges
• accessing/changing the number of states of a node

Construction of hierarchical Bayesian network and influence diagram models. A hierarchical model may contain other models as submodels. The interface of a model consists of a set of input nodes and a set of output nodes. This feature can be considered as object-oriented Bayesian networks and influence diagrams without inheritance.

Class collections to store multiple OOBN definitions in the same file.

Discrete chance and decision nodes with subtypes labelled, boolean, numbered and interval.

More than one utility node (assumes an additively decompositing utility function).

Conditional Gaussian distributed variables (exact inference).

Function nodes to perform post-inference custom calculations

Direct specification of conditional probability distributions, conditional Gaussian densities and utility functions.

Probabilistic relations and utility functions can be specified by mathematical and logical descriptions using the Table Generator.

Rotation of the interface of instance nodes.

Substitution of the class of instance nodes.

Substitution of node parents.

Continuous-discrete node value transfer function.

Usage

An Expression Builder Wizard makes it easier to build an expression by dividing the process into steps that allow the user to concentrate on building one simple expression at a time while keeping an overview of the complete expression.

Table Generator supporting:

• Discrete distributions (Binomial, Negative Binomial, Poisson, Geometric, NoisyOr, Distribution)
• Continuous distributions (Normal, Beta, Gamma, Exponential, Weibull, Uniform, PERT, Triangular, LogNormal) including a truncation operator.
• Logical, conditional and comparison operators (and, or, not, if-then-else, equals, less than, greater than, not equals, less than or equals, greater than or equals)
• Standard mathematical operators (add, subtract, multiply, divide, power, negate, min, max, log, exp, sqrt, log2, log10, sin, cos, tan, cosh, sinh, tanh, abs, ceil, floor, mod)
• Tables need only be generated once, but can be generated on demand. This includes tables for individual nodes. The generated tables are stored in the Hugin Network and the Hugin KB files in order to speed up compilation of domains
• Specification of "number of samples per interval" for each table.

The structure and conditional probability distribution of a Bayesian network can be estimated from data using the Learning Wizard.

Learning Wizard supporting:

• Data acquisition (database connectivity via ODBC or JDBC, interface to Oracle databases, and plain text files)
• Data preprocessing (e.g., state relabeling and discretization of continuous variables)
• Specification of domain expert knowledge (i.e., structural constraints including support for adding constraints between a node and a subset of nodes)
• Structural learning using the PC algorithm
• Structural learning using the NPC algorithm with detection of minimal resolutions of ambiguous regions
• A data dependency slider to detect the relative strengths of links as found in the data,
• Generation of randomized conditional probability distributions (used as part of EM-learning)
• Estimation of the conditional probability distributions using the EM algorithm.

The PC or NPC algorithm for learning the graphical structure of a Bayesian network from a database of cases based on statistical test of dependence relations.

Chow-Liu and TAN algorithms for learning a graphical tree structure of a Bayesian network from a database of cases

EM-learning of a subset or all of the conditional probability distributions in Bayesian networks, Mixed Bayesian networks, or Influence Diagrams. Prior knowledge can be specified both in order to speed up the learning and to guide the learning algorithm (penalized-EM).

EM-learning in object-oriented Bayesian networks. This feature enables the user to exploit the composition of an object-oriented Bayesian network when estimating conditional probability tables from data.

Cut-and-paste facilities for nodes, network fragments and conditional probability tables.

File support for the Hugin Network Language and the Hugin Knowledge Base format.

Compilation of domains into junction forests (i.e. a set of junction trees). A number of different algorithms for constructing the junction tree of cliques: clique size, clique weight, fill in size, fill in weight, total weight (optimal), best greedy and user specified. A compilation log can be associated with domains. Node elimination orders may be saved to and loaded from file.

Conditional probability distributions, density functions and utility functions can be changed and used without performing a new compilation of the junction tree.

Insertion of both hard (instantiation) and soft (likelihood) evidence.

Evidence is not removed from the domain when the domain is uncompiled. Also, evidence can be entered and retracted when the domain is not compiled.

The Hugin inference engine exploits the cache in modern CPUs to speed up inference.

Retraction of evidence.

Limited memory influence diagrams (LIMIDs) for modeling decision problems. The LIMID representation removes two fundamental assumptions of the traditional influence diagram: A non-forgetting decision maker and a total order of the decisions. By removing these two assumptions, the LIMID representation implies a significant improvement in the support for modeling decision problems. In addition to supporting the LIMID representation, the tool supplies the user with additional information related to a decision model (e.g., show policies and both decision and chance nodes have a probability distribution and an expected utility function determined by the current strategy).

Propagation of evidence in junction tree of cliques

• Sum and max propagation in normal and fast-retraction mode with discrete variables
• Sum-propagation of evidence in normal mode with a mixture of Conditional Gaussian and discrete variables (exact computations)
• The probability of the propagated evidence is available as a result of propagation (normalization constant)

Retrieve the belief of a discrete node and the density of conditional Gaussian node.

Different ways of displaying the belief of a node e.g. monitor windows, listing of marginals, and mean and variance of conditional Gaussian distributed variables or display of the exact distribution.

Solution of sequential decision problems with multiple utility nodes and missing no-forgetting links.

Different ways of displaying the expected utility of a decision node (monitor windows or a list of nodes).

Generation of cases with missing values (MAR and MCAR).

Zero-compression.

Approximation.

Save-to-memory.

Simulation of all variables in a model.

Save the junction tree (without clique tables) and evidence entered into the junction tree in hkb-file.

Various triangulation options including optimal triangulation.

Generation of C, C# and Java code from model.

Revision

Sequential learning of a subset of the conditional probability distributions (adaptation) is also available in influence diagrams. Beliefs in conditional probability tables can be specified using experience counts, and the impact of prior knowledge can be faded away using fading factors.

Adaptation wizard.

Arc-reversal between pairs of discrete variables and pairs of continuous variables.

Node absorption.

Rearrange content of conditional probability distributions and utility functions.

Switch node class for discrete nodes (chance <=> decision).

Analysis

Parameter sensitivity analysis and constraints on the posterior probability of a hypothesis variable.

Support for computing joint probability distributions including a Monte Carlo algorithm for finding the most probable configurations.

Possible suspicious findings can be located using fast retraction of evidence propagation.

Support for data conflict analysis including tracing conflicts and conflict resolution.

Hypothesis driven data conflict analysis and hypothesis drive conflict analysis.

Support for detailed conflict analysis in the junction tree.

Value of information analysis on discrete random variables and discrete decision variables.

Sensitivity to evidence analysis on discrete random variables including identification of min-max beliefs, impact of subsets of evidence analysis, discrimination of hypotheses, what-if analysis, and impact of findings analysis.

Convolution and XOI loss model.

d-separation/connection analysis.

Data analysis.

Case manager.

ROC analysis.

Display of Markov Blanket.

Correlation analysis between any pair of discrete nodes.

Identification of requisite parents of a decision node in a LIMID.

Identification of requisite ancestors of a decision node in a LIMID - useful for turning LIMID into a traditional influence diagram.

Hidden node analysis.

Feature selection.

Development and Documentation

User defined attributes on nodes and domains.

User data on nodes and domains.

Readable network file format.

Hyperlinks in Node and Domain descriptions

Zoom factors in both the Hugin Network language and Hugin Knowledge base file formats.

Grouping of nodes.

Belief bars can be forces to render in scientific notation.

Support for unicode.

Internationalization (includes support for English and Japanese languages)

Graph layout algorithm with support for external algorithms

Possibility of changing name, label, position, font, color and size of nodes.

Usage and network log information.

Import and export of tables.

Print tables, networks, probabilities and junction trees on printer or to a file.

Automatic generation of HTML network documentation

Vector quality PDF export of network structure

Some results in Analysis Wizard can be exported as a CSV file.

Many help facilities in the form of html-code including drag'N'drop.

The Hugin GUI can be customized to meet the preference of the user.

The Hugin GUI is available for a wide range of platforms including Microsoft Windows platforms, Sun Solaris (sparc and x86) and Linux Red Hat. The Hugin GUI comes as either a 32-bit or 64-bit application.

Feature List for HUGIN Decision Engine 7.4 - HDE 7.4

Construction

Construction of Bayesian network, chain graph (only available through the Hugin Network File Format) and influence diagram models (known as domains within the Hugin Decision Engine).

• Creating/cloning/deleting of domains/nodes
• Adding/removing edges
• Accessing/changing the number of states of a node

Discrete chance and decision nodes with subtypes labelled, boolean, numbered and interval.

More than one utility node (assumes an additively decompositing utility function).

Conditional Gaussian distributed variables.

Direct specification of conditional probability distributions, conditional Gaussian densities and utility functions.

Probabilistic relations and utility functions can be specified by mathematical and logical descriptions using the Table Generator.

Function nodes to perform post-inference custom calculations

Substitution of the class of instance nodes.

Substitution of node parents.

Table Generator supporting

• Discrete distributions (Binomial, Negative Binomial, Poisson, Geometric, NoisyOr, Distribution)
• Continuous distributions (Normal, Beta, Gamma, Exponential, Weibull, Uniform, PERT, Triangular, LogNormal) including a truncation operator
• Logical, conditional, and comparison operators (and, or, not, if-then-else, equals, less than, greater than, not equals, less than or equals, greater than or equals)
• Standard mathematical operators (add, subtract, multiply, divide, power, negate, min, max, log, exp, sqrt, log2, log10, sin, cos, tan, cosh, sinh, tanh, abs, ceil, floor, and mod)
• Tables only need to be generated once, but can be generated on demand
• The number of values for each interval parent node used when generating the table of a child can be specified through the Application Programming Interface to the Hugin inference engine.

The PC algorithm for learning the graphical structure of a Bayesian network from a database of cases based on statistical test of dependence relations.

Specification of domain expert knowledge when performing structural learning.

EM-learning of a subset or all of the conditional probability distributions in Bayesian networks with discrete variables. Prior knowledge can be specified both in order to speed up the learning and to guide the learning algorithm (penalized-EM).

EM-learning in object-oriented Bayesian networks. This feature enables the user to exploit the composition of an object-oriented Bayesian network when estimating conditional probability tables from data.

File support for the Hugin Network Language, the Hugin Knowledge Base format and the compressed Hugin Knowledge Base format.

Support for passwords on Hugin Knowledge Base files.

Compilation of domains into junction forests (i.e. a set of junction trees). A number of different algorithms for constructing the junction tree of cliques: clique size, clique weight, fill in size, fill in weight, total weight (optimal), best greedy and user specified. A compilation log can be associated with domains. Node elimination orders may be saved to and loaded from file.

Conditional probability distributions, density functions and utility functions can be changed and used without performing a new compilation of the junction tree.

Limited memory influence diagrams (LIMIDs) replace influence diagrams as the tool for modeling decision problems.

Usage

Insertion of both hard (instantiation) and soft (likelihood) evidence.

The Hugin inference engine exploits the cache in modern CPUs to speed up inference.

Retraction of evidence.

Evidence is not removed from the domain when the domain is uncompiled. Also, evidence can be entered and retracted when the domain is not compiled.

Retrieve state index from value or label.

Retrieve table index from a configuration of its nodes and vice versa.

Propagation of evidence in junction tree of cliques

• Sum and max propagation in normal and fast-retraction mode with discrete variables
• Sum-propagation of evidence in normal mode with a mixture of Conditional Gaussian and discrete variables (exact computations)
• The probability of the propagated evidence is available as a result of propagation (normalization constant)

Computation of the joint probability distributions of a set of discrete variables, the joint density of continuous variables, and a mixture.

Retrieve the belief of a discrete node and the density of conditional Gaussian node.

Solution of sequential decision problems with multiple utility nodes and missing no-forgetting links.

Retrieve expected utility of a decision node.

Sampling using the Bayesian network or the junction tree of cliques and sampling of discrete, conditional Gaussian, and mixture of conditional Gaussian and discrete variables conditional. A configuration of the variables can be sampled according to the distribution determined by the evidence.

Sampling in uncompiled Bayesian networks when the set of nodes with evidence form an ancestral set of instantiated nodes.

Zero-compression.

Approximation.

Save to memory functionality to support efficient inference and initialization.

Parallel processing for probabilistic inference on multi-processor systems.

Junction tree navigation.

Save the junction tree and evidence entered into the junction tree in hkb-file.

Various triangulation options including optimal triangulation.

Revision

Sequential learning of a subset of the conditional probability distributions (adaptation). Also, available in influence diagrams. Beliefs in conditional probability tables can be specified using experience counts and the impact of prior knowledge can be faded away using fading factors.

Arc-reversal.

Rearrange content of conditional probability distributions and utility functions.

Switch node class for discrete nodes (chance <=> decision).

Analysis

Parameter sensitivity analysis.

Support for computing joint probability distributions including a Monte Carlo algorithm for finding the most probable configurations.

Computation of sensitivity data for multiple output probabilities simultaneously that allows for easy solution of constraints on the output probabilities.

Conflict analysis.

Value of information analysis on discrete random variables.

Possible suspicious findings can be located using fast retraction of evidence propagation.

Functionality to support implementation of sensitivity analysis.

d-separation/connection analysis.

Identification of requisite parents of a decision node in a LIMID.

Documentation

User defined attributes on nodes and domains.

User data on nodes and domains.

Readable network file format.

Possibility of changing name, label, position and size of nodes.

Many help facilities in the form of an extensive reference manual.

The Hugin Decision Engine is available with a C API, a C++ API, a Java API, a .NET API and as an ActiveX-server.

The Hugin Decision Engine is provided in two versions: a version using single precision floating point arithmetic and a version using double-precision floating point arithmetic. The double precision version may prove useful in computations with continuous random variables (at the cost of a larger space requirement). Domains saved in the hkb-format using a single precision version can be loaded by a double-precision version of the Hugin Decision Engine, and vice versa. Error handling: whenever a Hugin Decision Engine operation fails, an error indicator is returned - it is then up to the application to decide what action to take.

32-bit and 64-bit APIs (except the Visual Basic API) are available for all platforms except MAC OS X.