Glossary   A - E

Absorption:  Node absorption is the process of removing a node from a Bayes net or decision net, and adjusting the remaining links and node tables so that subsequent inference done on the remaining nodes will yield the same results.  Usually Netica has to add some new links to maintain the global relationships between the nodes.  more info

Active Window:  Menu and key commands generally apply to the active window, which is the window with the non-dim title bar, and is the front most window (except possibly for some dialog boxes and help windows).  You can make a window active by clicking on an exposed part of it, or by choosing its title from the Window menu.

ASCII:  ASCII is a text character encoding based on the English alphabet, and was first released as a standard in 1967.  It represents each character with 7 bits, although there are many 8-bit extensions of it.  ASCII and its extensions have been by far the dominant way to represent text in a computer, but are now being overtaken by Unicode, which can represent many more characters.  In Netica, identifiers (i.e. IDnames) are always composed only of ASCII characters, while titles and descriptions may be in Unicode.

Assessment:  Probability assessment is the process of humans determining the probabilistic or deterministic relationships between nodes and their parents (usually in the form of conditional probability tables) after all the nodes and the link structure have been created.  Alternatively, they can be determined automatically by some learning procedure.

AutoNetica:  AutoNetica is a system to deploy your Bayes nets over the internet as a question-answer system. Such a system asks the user questions, or provides a dashboard to enter relevant information, and presents the user with conclusions.  more info

Auto-updating:  If a Bayes net is auto-updating, then whenever its beliefs become invalid, perhaps due to entering new findings, they are automatically recalculated (provided the net is compiled).  Since updating may be time consuming, you may prefer to only have the net updated when you manually request it.  The default for new nets is that they are auto-updating.  more info

Background:  The background of a Bayes net or decision net is the area (within its window) which is not covered by a node or a link.

Barren node:  A barren node is a node with no children, and that is not a findings node or a target node.  During belief updating and finding optimal decisions, nodes that are barren don’t influence the results, and may simply be removed.

Bayes net:  A Bayes net (also known as a belief net) is composed of a set of nodes representing variables of interest, connected by links to indicate dependencies, and containing information about the relationships between the nodes (often in the form of conditional probabilities).  Usages include prediction, diagnosis, probabilistic modeling, learning from data and forming a basis for building decision nets.  more info

Belief:  The belief of a node is the set of probabilities (one for each of its possible states), taking into account the currently entered findings by using the knowledge encoded in the Bayes net.  Technically speaking, it is the marginal posterior probability distribution of the node, given the findings and the Bayes net model.  Sometimes the plural form “beliefs” is used to mean each of the probabilities in the set.

Belief updating:  Belief updating is the process of finding new beliefs for the nodes of a Bayes net to account for the findings that are currently known.  It is a form of probabilistic inference.  During belief updating the Bayes net model (in particular, the conditional probability tables between the nodes) is not modified at all; for that probability revision is used.  more info

Bernoulli process:  A Bernoulli process consists of a series of independent trials, each with two possible outcomes (often labeled "success" and "failure"), with a constant probability, p, of success (such as a set of coin tosses).  If there are n trials, it is also called a binomial experiment, and the total number of successes, k, is given by the binomial distribution.  The number of trials needed before getting a fixed number of successes is given by a negative binomial distribution, and the number of trials needed to get one success is given by a geometric distribution.  If there are more than two possible outcomes, it is a multinomial experiment, and its results are given by a multinomial distribution.

Binary node:  A binary node is a node with exactly two states.  If those states correspond to ‘true’ and ‘false’, then it is also called a boolean node.

Binomial coefficient:  The binomial coefficient, denoted binomial(n,k), is the number of different k-sized groups that can be drawn from a set of n distinct elements.  Its value is given by: binomial(n,k) = n! / (k! * (n-k)!) Within a Netica equation, you can represent it with the binomial function.

Boolean node:  A boolean node is a node with two states, and whose state names are (true, false), (yes, no), (present, absent) or (on, off).  These state names can be in either order, and in lower or upper case (such as True or TRUE).  Some people refer to them as “propositional nodes”.  They are examples of binary nodes.

Cartesian product:  The cartesian product of two sets is the set of all possible pairs of elements, where the first element of each pair is taken from the first set, and the second element is taken from the second set.  For example, the cartesian product of {low, medium, high} with {true, false} is {(low, true), (low, false), (medium, true), (medium, false), (high, true), (high, false)}  more info

Case:  A case is a set of findings that go together to provide information on one object, event, history, person, or other thing.  more info

Case symbol:  The case symbol is a tiny yellow page with writing to depict a list of attributes and values.  It looks like this: image\Case_Symbol.gif  It is used on toolbar buttons to indicate a case.  

For example, the image\DelCaseTOOL.gif  button removes the current case, the image\SaveCaseTOOL.gif  button saves the case to a file, and the image\OpenCaseTOOL.gif  button reads a case from file.

Central limit theorem:  The central limit theorem states that the distribution of the sum of a set of random variables approaches the normal distribution as the number of variables increases, provided they are independent and some other weak conditions (such as Liapounov's conditions) are met.  These conditions will always be met if they are identically distributed and their means and standard deviations exist.

Chance node:  A chance node is a nature node whose relationship with its parents is probabilistic (i.e. not deterministic).  If its parents values are all known, and there is no further information, then its value can only be inferred as a probability distribution over possible values.  Compare with deterministic node.

Changed Indicator:  A changed-indicator (sometimes known as a "dirty indicator") is a * or + after the title of the net in the net window's title bar, and means that the net currently displayed has been changed from the version saved to file.  * means a meaningful (i.e. "semantic") change, while + means just a change to the visual display.  Entering findings will not trigger the changed-indicator, unless they are for a constant node.  The changed-indicator properly responds to undo and redo.  more info

A table dialog can also have a * changed-indicator.  Or it can have a (*) indicator, which means that something else has changed the target, which makes it different from the version in the dialog box.  more info

Check:  Menu item, Table Check.  Not supported after version 3.08.

Child node:  If there is a link going from node A to node B, then B is said to be a child node of A.  Some people refer to it as a direct successor.

Clipboard:  Whenever you do a copy operation (for example by choosing Edit Copy, or pressing ctrl+c), the material you copy goes to the clipboard, from which you can paste it where you like.

Clique:  A clique is a set of nodes in which each node is connected to all the other nodes of the set, and there isn’t any other node in the net which is connected to all the nodes in the set.  Some people call this a “maximal clique”.  When Netica compiles a Bayes net, one step is to find the cliques of the triangulated Markov network.

Compile: When Netica compiles a Bayes net it takes a representation of the net similar to what you see on the screen, and from it builds a new representation called a junction tree, which it can use to do fast probabilistic inference.  more info

Conditional probability:  The conditional probability of an event is the probability of the event occurring under certain given conditions.  more info

Conditionally independent:  A variable X is said to be conditionally independent of another variable Y given knowledge Z, if obtaining knowledge about the value of X does not change your beliefs about the value of Y when you already know Z.

Conjugate distributions:  Conjugate distributions are probability distributions from a family of distributions, such that if the prior distribution belongs to the family, then for any sample size and any observations, the posterior distribution also belongs to the family.  Usually each distribution in the family can be specified by one or two parameters, so only these parameters need to be kept track of during probabilistic inference, which is particularly convenient.

Constant node:  Sometimes it is useful to have something that normally acts as a fixed constant, but which you can change from time to time.  That is the purpose of a constant node.  more info

Contingency table:  A contingency table provides a value (or set of values) for each possible configuration of values for some given variables.  In other words, it specifies a function of the cartesian product of the variables.  Netica’s table dialog box can be used to view or edit contingency tables.

Continuous variable:  A continuous variable is one which can take on a value between any other two values, such as: indoor temperature, time spent waiting, water consumed, color wavelength, and direction of travel.  A discrete variable corresponds to a digital quantity, while a continuous variable corresponds to an analog quantity.  more info

CPT:  CPT is an abbreviation for conditional probability table (also known as “link matrix”), which is the contingency table of conditional probabilities stored at each node, containing the probabilities of the node given each configuration of parent values.  Sometimes CPT is used to refer to the deterministic function table of a node, since the node's conditional probabilities can easily be found from that.  It is a form of node relation, so you use the table dialog box to change or view it.

CSV file:  CSV file is a commonly used term for a form of case file in which the names of the variables appear on the first line, and then below are all the cases (i.e. records), with each case on a single line and having a value for each of the variables, and with all the values and variables in text form and separated by commas (i.e. "Comma Separated Values").  See also tab delimited text.

Cutset:  A cutset for two sets of nodes A and B is a third set C, such that if all the nodes in C are removed from the net (along with all links involving them), then there is no path from a node in A to a node in B.  See also loop cutset.

d-separation:  The d-separation rule enables you to quickly determine whether a finding at one node can possibly change the beliefs at another node by only considering the link structure of a Bayes net.  more info

Dag:  A dag is a net with no directed cycles (although it may have undirected loops). The word is a shortened form of “directed acyclic graph”.

Decision net:  If decision nodes (representing variables that can be controlled) and utility nodes (representing variables to be optimized) are added to a Bayes net, then a decision net (also known as an “influence diagram”) is formed.  more info

Decision node:  A decision node is a node in a decision net which represents a variable (or choice) under the control of the decision maker.  When the net is solved, a decision rule is found for the node which optimizes the expected utility.  Decision nodes are normally drawn as rectangles (without rounded corners).

Decision rule:  A decision rule indicates which option to choose in making a certain decision, for each possible condition that may be known when the decision is to be made.  In other words, for a decision node, it is a function which provides a value for each member of the cartesian product of the parents of the decision node.

Decision theory:  Decision theory is a normative theory which indicates how a single agent should best make decisions to maximize his expected utility.  It considers sequences of decisions, what information the agent will have when he makes the decisions, uncertainties in the beliefs of the agent, and complex probabilistic interactions in the environment in which the agent is operating.

Default node style:  There is one default node style for the whole net, which is the style to display any node which doesn’t have an overriding style.  It is set by choosing it from the Style menu when no nodes are selected.  more info

Deselect:  To deselect all the nodes of a net, click on its background (i.e. within the window, but not on a node or link).  more info

Deterministic node:  A deterministic node is a nature node whose relationship with its parents is given as a function of the parent values (i.e. deterministic rather than probabilistic).  If the parent values are all known, its value can be determined with certainty.  Compare with chance node.

Deterministic updating:  Before doing belief updating, Netica does deterministic updating when it can, for greater speed and accuracy.  If all the parents of a node have findings, and the node has a function table or a deterministic equation, then its value can be found exactly (the equation is used without discretization) very quickly.

Directed cycle:  A directed cycle (sometimes just called a “cycle”) is a path through a net, following the direction of the arrows, which returns to its beginning (i.e. the first node of the path is the same as the last).  Compare with undirected loop.

Netica can find and display directed cycles.

Directed network:  A directed network is one where the links have direction (i.e. arrows).  Bayes nets and decision nets are examples.  Compare with an undirected network.

Directed path:  A directed path is a sequence of nodes from a net, such that you can get from one node of the sequence to the next node by traversing a link between them in the direction of its arrow.  Compare with path.

Dirty indicator:  see changed indicator.

Disconnected link:  A disconnected link is one that has been disconnected from the parent node it originally came from, and belief updating cannot proceed until the link is reconnected to its parent, or to another similar node. more info

Discrete node:  A discrete node is a node representing a discrete variable.  The states of a node constitute the domain of the categorical variable.  

Discrete variable:  A discrete variable is one with a well defined finite set of possible values, called states.  Examples are: the number of dimes in a purse, the ‘true’ or ‘false’ value of a statement, which party will win the election, the country of origin, and the place a roulette wheel stops.  A discrete variable corresponds to a digital quantity, while a continuous variable corresponds to an analog quantity.  more info

Discretize:  Often it is useful to have a continuous variable behave like a discrete one.  To do this, select the node(s) and choose Modify Discretize Node.  You can break up the total range of the continuous variable into a number of intervals by supplying numbers showing where one interval ends and the next begins (called thresholds).  This is known as discretizing the variable.  Each interval results in one state of the discrete version of the variable.   more info

Elimination order:  The elimination order is simply an ordered list of all the nodes in the net, which specifies how to triangulate the net during compiling (see Spiegelhalter&DLC93).  Triangulation is the most critical step in producing an efficient compilation, so the order is included when the net is next saved to file if an “Optimized Compile” command (which finds a very good elimination order) has been done.  The elimination order is indicated by the numbers following the node names when you view the net in “triangulated style”.  more info

Ellipsis:  An ellipsis is three closely spaced dots, used to indicate that there is more to follow, but that it has been left out to save space.  For example: LongNameFor…

E-mail:  The Norsys team very much welcomes questions and comments about Netica or this onscreen help document.  All inquires should be sent to: support@norsys.com.  To do this directly from Netica, choose Help Email Norsys.

Entering findings:  When a Bayes net is applied to a particular situation, or case, then the known information about that case is entered into the Bayes net by assigning values (called "findings", or "evidence") to the known variables (i.e. nodes), and that process is known as entering findings into the nodes.  Entering a finding into a particular node does not retract existing findings at that node or other nodes (but for convenience, in Netica Application, if the new finding for a node flat-out contradicts a previously entered finding for that node, the previous finding will be retracted first).  more info

Equation:  Within Netica, the probabilistic or deterministic relationship between nodes may be expressed using an equation, as an alternative to CPT or function tables.  The equation follows a syntax common in mathematics and computer programming, and may use any of a large set of special functions built-in to Netica.   more info

Ergo:  A program available from Noetic Systems Incorporated which works with Bayes nets.  If you have Bayes net files in the old Ergo format (i.e. not .ent files) which you wish to use with Netica, then give them a file extension of ".ergo" and Netica will be able to read them.

Error rate:  If a classifier is tested on a number of cases, each of known class, one can determine how many times it misclassified a case (i.e. said the case belonged to some class when in fact it belonged to a different one).  That, divided by the number of classifications made, is the error rate (usually expressed as a percentage).  The error rate is only with respect to the probability distribution of the test cases.

Examples:  After you install Netica on your computer, within the "Netica" folder will be a folder called "Examples".  It contains several Bayes nets suitable for an introduction to the field of Bayes nets, and for use in learning about Netica.  Within it is a folder called "Tutorial - View these first" containing on ordered set of Bayes nets that contain some tutorial information and exercises in their net description window; if you are working through them, you might be interested in reading the Quick Tour of this onscreen help (if it mentions some file you don't see in the Example folder, don't forget to look in its Tutorial sub-folder).  For many more examples, be sure to see our great online Bayes net library, available by choosing Help Net Library Website.

Expected value:  The expected value (also known as mean value) is not the value you “expect” to see, and usually it isn’t even the value most likely to occur.  This term, from probability theory, means the average value that will occur, where the average is weighted by the probability of occurrence.  For example if the value will be 3 with probability 0.2 and 9 with probability 0.8, then the expected value is: (0.2 x 3) + (0.8 x 9) = 7.8.

Experience:  For learning and communicating the knowledge contained within a Bayes net, it is useful to indicate a “confidence” in the conditional probabilities it contains, so that one knows how much to change them when new information about them becomes known.  That confidence is called the experience, and it is calibrated to be equivalent to seeing a certain number of relevant cases.  Sometimes this is called the “equivalent sample size”.  more info

Explaining-Away:  If A and B are both possible causes of E, then when we find that E occurred, providing there are no complicating factors, our beliefs that A occurred and B occurred will both increase.  But if we further discover that A occurred, then our belief that B occurred will go back down a little.  This is called explaining-away, because A provides the explanation for E, and there is no need for B to have occurred to explain E.  Bayes nets automatically handle explaining away in the correct manner, while evidence-based and rule-based systems are generally poor at it.