How to use the Bayes Net Toolbox

This documentation was last updated on 29 October 2007.
Click here for a French version of this documentation (last updated in 2005).

Creating your first Bayes net

To define a Bayes net, you must specify the graph structure and then the parameters. We look at each in turn, using a simple example (adapted from Russell and Norvig, "Artificial Intelligence: a Modern Approach", Prentice Hall, 1995, p454).

Graph structure

Consider the following network.

To specify this directed acyclic graph (dag), we create an adjacency matrix:

N = 4; 
dag = zeros(N,N);
C = 1; S = 2; R = 3; W = 4;
dag(C,[R S]) = 1;
dag(R,W) = 1;

We have numbered the nodes as follows: Cloudy = 1, Sprinkler = 2, Rain = 3, WetGrass = 4. The nodes must always be numbered in topological order, i.e., ancestors before descendants. For a more complicated graph, this is a little inconvenient: we will see how to get around this below.

In Matlab 6, you can use logical arrays instead of double arrays, which are 4 times smaller:

dag = false(N,N);
dag(C,[R S]) = true;
However, some graph functions (eg acyclic) do not work on logical arrays!

You can visualize the resulting graph structure using the methods discussed below. For details on GUIs, click here.

Creating the Bayes net shell

In addition to specifying the graph structure, we must specify the size and type of each node. If a node is discrete, its size is the number of possible values each node can take on; if a node is continuous, it can be a vector, and its size is the length of this vector. In this case, we will assume all nodes are discrete and binary.
discrete_nodes = 1:N;
node_sizes = 2*ones(1,N); 
If the nodes were not binary, you could type e.g.,
node_sizes = [4 2 3 5];
meaning that Cloudy has 4 possible values, Sprinkler has 2 possible values, etc. Note that these are cardinal values, not ordinal, i.e., they are not ordered in any way, like 'low', 'medium', 'high'.

We are now ready to make the Bayes net:

bnet = mk_bnet(dag, node_sizes, 'discrete', discrete_nodes);
By default, all nodes are assumed to be discrete, so we can also just write
bnet = mk_bnet(dag, node_sizes);
You may also specify which nodes will be observed. If you don't know, or if this not fixed in advance, just use the empty list (the default).
onodes = [];
bnet = mk_bnet(dag, node_sizes, 'discrete', discrete_nodes, 'observed', onodes);
Note that optional arguments are specified using a name/value syntax. This is common for many BNT functions. In general, to find out more about a function (e.g., which optional arguments it takes), please see its documentation string by typing
help mk_bnet
See also other useful Matlab tips.

It is possible to associate names with nodes, as follows:

bnet = mk_bnet(dag, node_sizes, 'names', {'cloudy','S','R','W'}, 'discrete', 1:4);
You can then refer to a node by its name:
C = bnet.names('cloudy'); % bnet.names is an associative array
bnet.CPD{C} = tabular_CPD(bnet, C, [0.5 0.5]);
This feature uses my own associative array class.


A model consists of the graph structure and the parameters. The parameters are represented by CPD objects (CPD = Conditional Probability Distribution), which define the probability distribution of a node given its parents. (We will use the terms "node" and "random variable" interchangeably.) The simplest kind of CPD is a table (multi-dimensional array), which is suitable when all the nodes are discrete-valued. Note that the discrete values are not assumed to be ordered in any way; that is, they represent categorical quantities, like male and female, rather than ordinal quantities, like low, medium and high. (We will discuss CPDs in more detail below.)

Tabular CPDs, also called CPTs (conditional probability tables), are stored as multidimensional arrays, where the dimensions are arranged in the same order as the nodes, e.g., the CPT for node 4 (WetGrass) is indexed by Sprinkler (2), Rain (3) and then WetGrass (4) itself. Hence the child is always the last dimension. If a node has no parents, its CPT is a column vector representing its prior. Note that in Matlab (unlike C), arrays are indexed from 1, and are layed out in memory such that the first index toggles fastest, e.g., the CPT for node 4 (WetGrass) is as follows

where we have used the convention that false==1, true==2. We can create this CPT in Matlab as follows

CPT = zeros(2,2,2);
CPT(1,1,1) = 1.0;
CPT(2,1,1) = 0.1;
Here is an easier way:
CPT = reshape([1 0.1 0.1 0.01 0 0.9 0.9 0.99], [2 2 2]);
In fact, we don't need to reshape the array, since the CPD constructor will do that for us. So we can just write
bnet.CPD{W} = tabular_CPD(bnet, W, 'CPT', [1 0.1 0.1 0.01 0 0.9 0.9 0.99]);
The other nodes are created similarly (using the old syntax for optional parameters)
bnet.CPD{C} = tabular_CPD(bnet, C, [0.5 0.5]);
bnet.CPD{R} = tabular_CPD(bnet, R, [0.8 0.2 0.2 0.8]);
bnet.CPD{S} = tabular_CPD(bnet, S, [0.5 0.9 0.5 0.1]);
bnet.CPD{W} = tabular_CPD(bnet, W, [1 0.1 0.1 0.01 0 0.9 0.9 0.99]);

Random Parameters

If we do not specify the CPT, random parameters will be created, i.e., each "row" of the CPT will be drawn from the uniform distribution. To ensure repeatable results, use
rand('state', seed);
randn('state', seed);
To control the degree of randomness (entropy), you can sample each row of the CPT from a Dirichlet(p,p,...) distribution. If p << 1, this encourages "deterministic" CPTs (one entry near 1, the rest near 0). If p = 1, each entry is drawn from U[0,1]. If p >> 1, the entries will all be near 1/k, where k is the arity of this node, i.e., each row will be nearly uniform. You can do this as follows, assuming this node is number i, and ns is the node_sizes.
k = ns(i);
ps = parents(dag, i);
psz = prod(ns(ps));
CPT = sample_dirichlet(p*ones(1,k), psz);
bnet.CPD{i} = tabular_CPD(bnet, i, 'CPT', CPT);

Loading a network from a file

If you already have a Bayes net represented in the XML-based Bayes Net Interchange Format (BNIF) (e.g., downloaded from the Bayes Net repository), you can convert it to BNT format using the BIF-BNT Java program written by Ken Shan. (This is not necessarily up-to-date.)

It is currently not possible to save/load a BNT matlab object to file, but this is easily fixed if you modify all the constructors for all the classes (see matlab documentation).

Creating a model using a GUI

Graph visualization

Click here for more information on graph visualization.


Having created the BN, we can now use it for inference. There are many different algorithms for doing inference in Bayes nets, that make different tradeoffs between speed, complexity, generality, and accuracy. BNT therefore offers a variety of different inference "engines". We will discuss these in more detail below. For now, we will use the junction tree engine, which is the mother of all exact inference algorithms. This can be created as follows.
engine = jtree_inf_engine(bnet);
The other engines have similar constructors, but might take additional, algorithm-specific parameters. All engines are used in the same way, once they have been created. We illustrate this in the following sections.

Computing marginal distributions

Suppose we want to compute the probability that the sprinker was on given that the grass is wet. The evidence consists of the fact that W=2. All the other nodes are hidden (unobserved). We can specify this as follows.
evidence = cell(1,N);
evidence{W} = 2;
We use a 1D cell array instead of a vector to cope with the fact that nodes can be vectors of different lengths. In addition, the value [] can be used to denote 'no evidence', instead of having to specify the observation pattern as a separate argument. (Click
here for a quick tutorial on cell arrays in matlab.)

We are now ready to add the evidence to the engine.

[engine, loglik] = enter_evidence(engine, evidence);
The behavior of this function is algorithm-specific, and is discussed in more detail below. In the case of the jtree engine, enter_evidence implements a two-pass message-passing scheme. The first return argument contains the modified engine, which incorporates the evidence. The second return argument contains the log-likelihood of the evidence. (Not all engines are capable of computing the log-likelihood.)

Finally, we can compute p=P(S=2|W=2) as follows.

marg = marginal_nodes(engine, S);
ans =
p = marg.T(2);
We see that p = 0.4298.

Now let us add the evidence that it was raining, and see what difference it makes.

evidence{R} = 2;
[engine, loglik] = enter_evidence(engine, evidence);
marg = marginal_nodes(engine, S);
p = marg.T(2);
We find that p = P(S=2|W=2,R=2) = 0.1945, which is lower than before, because the rain can ``explain away'' the fact that the grass is wet.

You can plot a marginal distribution over a discrete variable as a barchart using the built 'bar' function:

This is what it looks like

Observed nodes

What happens if we ask for the marginal on an observed node, e.g. P(W|W=2)? An observed discrete node effectively only has 1 value (the observed one) --- all other values would result in 0 probability. For efficiency, BNT treats observed (discrete) nodes as if they were set to 1, as we see below:
evidence = cell(1,N);
evidence{W} = 2;
engine = enter_evidence(engine, evidence);
m = marginal_nodes(engine, W);
ans =
This can get a little confusing, since we assigned W=2. So we can ask BNT to add the evidence back in by passing in an optional argument:
m = marginal_nodes(engine, W, 1);
ans =
This shows that P(W=1|W=2) = 0 and P(W=2|W=2) = 1.

Computing joint distributions

We can compute the joint probability on a set of nodes as in the following example.
evidence = cell(1,N);
[engine, ll] = enter_evidence(engine, evidence);
m = marginal_nodes(engine, [S R W]);
m is a structure. The 'T' field is a multi-dimensional array (in this case, 3-dimensional) that contains the joint probability distribution on the specified nodes.
>> m.T
ans(:,:,1) =
    0.2900    0.0410
    0.0210    0.0009
ans(:,:,2) =
         0    0.3690
    0.1890    0.0891
We see that P(S=1,R=1,W=2) = 0, since it is impossible for the grass to be wet if both the rain and sprinkler are off.

Let us now add some evidence to R.

evidence{R} = 2;
[engine, ll] = enter_evidence(engine, evidence);
m = marginal_nodes(engine, [S R W])
m = 
    domain: [2 3 4]
         T: [2x1x2 double]
>> m.T
ans(:,:,1) =
ans(:,:,2) =
The joint T(i,j,k) = P(S=i,R=j,W=k|evidence) should have T(i,1,k) = 0 for all i,k, since R=1 is incompatible with the evidence that R=2. Instead of creating large tables with many 0s, BNT sets the effective size of observed (discrete) nodes to 1, as explained above. This is why m.T has size 2x1x2. To get a 2x2x2 table, type
m = marginal_nodes(engine, [S R W], 1)
m = 
    domain: [2 3 4]
         T: [2x2x2 double]
>> m.T
ans(:,:,1) =
            0        0.082
            0       0.0018
ans(:,:,2) =
            0        0.738
            0       0.1782

Note: It is not always possible to compute the joint on arbitrary sets of nodes: it depends on which inference engine you use, as discussed in more detail below.

Soft/virtual evidence

Sometimes a node is not observed, but we have some distribution over its possible values; this is often called "soft" or "virtual" evidence. One can use this as follows
[engine, loglik] = enter_evidence(engine, evidence, 'soft', soft_evidence);
where soft_evidence{i} is either [] (if node i has no soft evidence) or is a vector representing the probability distribution over i's possible values. For example, if we don't know i's exact value, but we know its likelihood ratio is 60/40, we can write evidence{i} = [] and soft_evidence{i} = [0.6 0.4].

Currently only jtree_inf_engine supports this option. It assumes that all hidden nodes, and all nodes for which we have soft evidence, are discrete. For a longer example, see BNT/examples/static/softev1.m.

Most probable explanation

To compute the most probable explanation (MPE) of the evidence (i.e., the most probable assignment, or a mode of the joint), use
[mpe, ll] = calc_mpe(engine, evidence);     
mpe{i} is the most likely value of node i. This calls enter_evidence with the 'maximize' flag set to 1, which causes the engine to do max-product instead of sum-product. The resulting max-marginals are then thresholded. If there is more than one maximum probability assignment, we must take care to break ties in a consistent manner (thresholding the max-marginals may give the wrong result). To force this behavior, type
[mpe, ll] = calc_mpe(engine, evidence, 1);     
Note that computing the MPE is someties called abductive reasoning.

You can also use calc_mpe_bucket written by Ron Zohar, that does a forwards max-product pass, and then a backwards traceback pass, which is how Viterbi is traditionally implemented.

Conditional Probability Distributions

A Conditional Probability Distributions (CPD) defines P(X(i) | X(Pa(i))), where X(i) is the i'th node, and X(Pa(i)) are the parents of node i. There are many ways to represent this distribution, which depend in part on whether X(i) and X(Pa(i)) are discrete, continuous, or a combination. We will discuss various representations below.

Tabular nodes

If the CPD is represented as a table (i.e., if it is a multinomial distribution), it has a number of parameters that is exponential in the number of parents. See the example above.

Noisy-or nodes

A noisy-OR node is like a regular logical OR gate except that sometimes the effects of parents that are on get inhibited. Let the prob. that parent i gets inhibited be q(i). Then a node, C, with 2 parents, A and B, has the following CPD, where we use F and T to represent off and on (1 and 2 in BNT).
A  B  P(C=off)      P(C=on)
F  F  1.0           0.0
T  F  q(A)          1-q(A)
F  T  q(B)          1-q(B)
T  T  q(A)q(B)      1-q(A)q(B)
Thus we see that the causes get inhibited independently. It is common to associate a "leak" node with a noisy-or CPD, which is like a parent that is always on. This can account for all other unmodelled causes which might turn the node on.

The noisy-or distribution is similar to the logistic distribution. To see this, let the nodes, S(i), have values in {0,1}, and let q(i,j) be the prob. that j inhibits i. Then

Pr(S(i)=1 | parents(S(i))) = 1 - prod_{j} q(i,j)^S(j)
Now define w(i,j) = -ln q(i,j) and rho(x) = 1-exp(-x). Then
Pr(S(i)=1 | parents(S(i))) = rho(sum_j w(i,j) S(j))
For a sigmoid node, we have
Pr(S(i)=1 | parents(S(i))) = sigma(-sum_j w(i,j) S(j))
where sigma(x) = 1/(1+exp(-x)). Hence they differ in the choice of the activation function (although both are monotonically increasing). In addition, in the case of a noisy-or, the weights are constrained to be positive, since they derive from probabilities q(i,j). In both cases, the number of parameters is linear in the number of parents, unlike the case of a multinomial distribution, where the number of parameters is exponential in the number of parents. We will see an example of noisy-OR nodes

Other (noisy) deterministic nodes

Deterministic CPDs for discrete random variables can be created using the deterministic_CPD class. It is also possible to 'flip' the output of the function with some probability, to simulate noise. The boolean_CPD class is just a special case of a deterministic CPD, where the parents and child are all binary.

Both of these classes are just "syntactic sugar" for the tabular_CPD class.

Softmax nodes

If we have a discrete node with a continuous parent, we can define its CPD using a softmax function (also known as the multinomial logit function). This acts like a soft thresholding operator, and is defined as follows:
                    exp(w(:,i)'*x + b(i)) 
Pr(Q=i | X=x)  =  -----------------------------
                  sum_j   exp(w(:,j)'*x + b(j))

The parameters of a softmax node, w(:,i) and b(i), i=1..|Q|, have the following interpretation: w(:,i)-w(:,j) is the normal vector to the decision boundary between classes i and j, and b(i)-b(j) is its offset (bias). For example, suppose X is a 2-vector, and Q is binary. Then
w = [1 -1;
     0 0];

b = [0 0];
means class 1 are points in the 2D plane with positive x coordinate, and class 2 are points in the 2D plane with negative x coordinate. If w has large magnitude, the decision boundary is sharp, otherwise it is soft. In the special case that Q is binary (0/1), the softmax function reduces to the logistic (sigmoid) function.

Fitting a softmax function can be done using the iteratively reweighted least squares (IRLS) algorithm. We use the implementation from Netlab. Note that since the softmax distribution is not in the exponential family, it does not have finite sufficient statistics, and hence we must store all the training data in uncompressed form. If this takes too much space, one should use online (stochastic) gradient descent (not implemented in BNT).

If a softmax node also has discrete parents, we use a different set of w/b parameters for each combination of parent values, as in the conditional linear Gaussian CPD. This feature was implemented by Pierpaolo Brutti. He is currently extending it so that discrete parents can be treated as if they were continuous, by adding indicator variables to the X vector.

We will see an example of softmax nodes below.

Neural network nodes

Pierpaolo Brutti has implemented the mlp_CPD class, which uses a multi layer perceptron to implement a mapping from continuous parents to discrete children, similar to the softmax function. (If there are also discrete parents, it creates a mixture of MLPs.) It uses code from Netlab. This is work in progress.

Root nodes

A root node has no parents and no parameters; it can be used to model an observed, exogeneous input variable, i.e., one which is "outside" the model. This is useful for conditional density models. We will see an example of root nodes below.

Gaussian nodes

We now consider a distribution suitable for the continuous-valued nodes. Suppose the node is called Y, its continuous parents (if any) are called X, and its discrete parents (if any) are called Q. The distribution on Y is defined as follows:
- no parents: Y ~ N(mu, Sigma)
- cts parents : Y|X=x ~ N(mu + W x, Sigma)
- discrete parents: Y|Q=i ~ N(mu(:,i), Sigma(:,:,i))
- cts and discrete parents: Y|X=x,Q=i ~ N(mu(:,i) + W(:,:,i) * x, Sigma(:,:,i))
where N(mu, Sigma) denotes a Normal distribution with mean mu and covariance Sigma. Let |X|, |Y| and |Q| denote the sizes of X, Y and Q respectively. If there are no discrete parents, |Q|=1; if there is more than one, then |Q| = a vector of the sizes of each discrete parent. If there are no continuous parents, |X|=0; if there is more than one, then |X| = the sum of their sizes. Then mu is a |Y|*|Q| vector, Sigma is a |Y|*|Y|*|Q| positive semi-definite matrix, and W is a |Y|*|X|*|Q| regression (weight) matrix.

We can create a Gaussian node with random parameters as follows.

bnet.CPD{i} = gaussian_CPD(bnet, i);
We can specify the value of one or more of the parameters as in the following example, in which |Y|=2, and |Q|=1.
bnet.CPD{i} = gaussian_CPD(bnet, i, 'mean', [0; 0], 'weights', randn(Y,X), 'cov', eye(Y));

We will see an example of conditional linear Gaussian nodes below.

When learning Gaussians from data, it is helpful to ensure the data has a small magnitde (see e.g., KPMstats/standardize) to prevent numerical problems. Unless you have a lot of data, it is also a very good idea to use diagonal instead of full covariance matrices. (BNT does not currently support spherical covariances, although it would be easy to add, since KPMstats/clg_Mstep supports this option; you would just need to modify gaussian_CPD/update_ess to accumulate weighted inner products.)

Other continuous distributions

Currently BNT does not support any CPDs for continuous nodes other than the Gaussian. However, you can use a mixture of Gaussians to approximate other continuous distributions. We will see some an example of this with the IFA model below.

Generalized linear model nodes

In the future, we may incorporate some of the functionality of glmlab into BNT.

Classification/regression tree nodes

We plan to add classification and regression trees to define CPDs for discrete and continuous nodes, respectively. Trees have many advantages: they are easy to interpret, they can do feature selection, they can handle discrete and continuous inputs, they do not make strong assumptions about the form of the distribution, the number of parameters can grow in a data-dependent way (i.e., they are semi-parametric), they can handle missing data, etc. However, they are not yet implemented.

Summary of CPD types

We list all the different types of CPDs supported by BNT. For each CPD, we specify if the child and parents can be discrete (D) or continuous (C) (Binary (B) nodes are a special case). We also specify which methods each class supports. If a method is inherited, the name of the parent class is mentioned. If a parent class calls a child method, this is mentioned.

The CPD_to_CPT method converts a CPD to a table; this requires that the child and all parents are discrete. The CPT might be exponentially big... convert_to_table evaluates a CPD with evidence, and represents the the resulting potential as an array. This requires that the child is discrete, and any continuous parents are observed. convert_to_pot evaluates a CPD with evidence, and represents the resulting potential as a dpot, gpot, cgpot or upot, as requested. (d=discrete, g=Gaussian, cg = conditional Gaussian, u = utility).

When we sample a node, all the parents are observed. When we compute the (log) probability of a node, all the parents and the child are observed.

We also specify if the parameters are learnable. For learning with EM, we require the methods reset_ess, update_ess and maximize_params. For learning from fully observed data, we require the method learn_params. By default, all classes inherit this from generic_CPD, which simply calls update_ess N times, once for each data case, followed by maximize_params, i.e., it is like EM, without the E step. Some classes implement a batch formula, which is quicker.

Bayesian learning means computing a posterior over the parameters given fully observed data.

Pearl means we implement the methods compute_pi and compute_lambda_msg, used by pearl_inf_engine, which runs on directed graphs. belprop_inf_engine only needs convert_to_pot.H The pearl methods can exploit special properties of the CPDs for computing the messages efficiently, whereas belprop does not.

The only method implemented by generic_CPD is adjustable_CPD, which is not shown, since it is not very interesting.

Name Child Parents Comments CPD_to_CPT conv_to_table conv_to_pot sample prob learn Bayes Pearl
boolean B B Syntactic sugar for tabular - - - - - - - -
deterministic D D Syntactic sugar for tabular - - - - - - - -
Discrete D C/D Virtual class N Calls CPD_to_CPT Calls conv_to_table Calls conv_to_table Calls conv_to_table N N N
Gaussian C C/D - N N Y Y Y Y N N
gmux C C/D multiplexer N N Y N N N N Y
MLP D C/D multi layer perceptron N Y Inherits from discrete Inherits from discrete Inherits from discrete Y N N
noisy-or B B - Y Inherits from discrete Inherits from discrete Inherits from discrete Inherits from discrete N N Y
root C/D none no params N N Y Y Y N N N
softmax D C/D - N Y Inherits from discrete Inherits from discrete Inherits from discrete Y N N
generic C/D C/D Virtual class N N N N N N N N
Tabular D D - Y Inherits from discrete Inherits from discrete Inherits from discrete Inherits from discrete Y Y Y

Example models

Gaussian mixture models

Richard W. DeVaul has made a detailed tutorial on how to fit mixtures of Gaussians using BNT. Available

PCA, ICA, and all that

In Figure (a) below, we show how Factor Analysis can be thought of as a graphical model. Here, X has an N(0,I) prior, and Y|X=x ~ N(mu + Wx, Psi), where Psi is diagonal and W is called the "factor loading matrix". Since the noise on both X and Y is diagonal, the components of these vectors are uncorrelated, and hence can be represented as individual scalar nodes, as we show in (b). (This is useful if parts of the observations on the Y vector are occasionally missing.) We usually take k=|X| << |Y|=D, so the model tries to explain many observations using a low-dimensional subspace.
(a) (b) (c) (d)

We can create this model in BNT as follows.

ns = [k D];
dag = zeros(2,2);
dag(1,2) = 1;
bnet = mk_bnet(dag, ns, 'discrete', []);
bnet.CPD{1} = gaussian_CPD(bnet, 1, 'mean', zeros(k,1), 'cov', eye(k), ...
   'cov_type', 'diag', 'clamp_mean', 1, 'clamp_cov', 1);
bnet.CPD{2} = gaussian_CPD(bnet, 2, 'mean', zeros(D,1), 'cov', diag(Psi0), 'weights', W0, ...
   'cov_type', 'diag', 'clamp_mean', 1);
The root node is clamped to the N(0,I) distribution, so that we will not update these parameters during learning. The mean of the leaf node is clamped to 0, since we assume the data has been centered (had its mean subtracted off); this is just for simplicity. Finally, the covariance of the leaf node is constrained to be diagonal. W0 and Psi0 are the initial parameter guesses.

We can fit this model (i.e., estimate its parameters in a maximum likelihood (ML) sense) using EM, as we explain below. Not surprisingly, the ML estimates for mu and Psi turn out to be identical to the sample mean and variance, which can be computed directly as

mu_ML = mean(data);
Psi_ML = diag(cov(data));
Note that W can only be identified up to a rotation matrix, because of the spherical symmetry of the source.

If we restrict Psi to be spherical, i.e., Psi = sigma*I, there is a closed-form solution for W as well, i.e., we do not need to use EM. In particular, W contains the first |X| eigenvectors of the sample covariance matrix, with scalings determined by the eigenvalues and sigma. Classical PCA can be obtained by taking the sigma->0 limit. For details, see

By adding a hidden discrete variable, we can create mixtures of FA models, as shown in (c). Now we can explain the data using a set of subspaces. We can create this model in BNT as follows.

ns = [M k D];
dag = zeros(3);
dag(1,3) = 1;
dag(2,3) = 1;
bnet = mk_bnet(dag, ns, 'discrete', 1);
bnet.CPD{1} = tabular_CPD(bnet, 1, Pi0);
bnet.CPD{2} = gaussian_CPD(bnet, 2, 'mean', zeros(k, 1), 'cov', eye(k), 'cov_type', 'diag', ...
			   'clamp_mean', 1, 'clamp_cov', 1);
bnet.CPD{3} = gaussian_CPD(bnet, 3, 'mean', Mu0', 'cov', repmat(diag(Psi0), [1 1 M]), ...
			   'weights', W0, 'cov_type', 'diag', 'tied_cov', 1);
Notice how the covariance matrix for Y is the same for all values of Q; that is, the noise level in each sub-space is assumed the same. However, we allow the offset, mu, to vary. For details, see

I have included Zoubin's specialized MFA code (with his permission) with the toolbox, so you can check that BNT gives the same results: see 'BNT/examples/static/mfa1.m'.

Independent Factor Analysis (IFA) generalizes FA by allowing a non-Gaussian prior on each component of X. (Note that we can approximate a non-Gaussian prior using a mixture of Gaussians.) This means that the likelihood function is no longer rotationally invariant, so we can uniquely identify W and the hidden sources X. IFA also allows a non-diagonal Psi (i.e. correlations between the components of Y). We recover classical Independent Components Analysis (ICA) in the Psi -> 0 limit, and by assuming that |X|=|Y|, so that the weight matrix W is square and invertible. For details, see

Mixtures of experts

As an example of the use of the softmax function, we introduce the Mixture of Experts model. As before, circles denote continuous-valued nodes, squares denote discrete nodes, clear means hidden, and shaded means observed.

X is the observed input, Y is the output, and the Q nodes are hidden "gating" nodes, which select the appropriate set of parameters for Y. During training, Y is assumed observed, but for testing, the goal is to predict Y given X. Note that this is a conditional density model, so we don't associate any parameters with X. Hence X's CPD will be a root CPD, which is a way of modelling exogenous nodes. If the output is a continuous-valued quantity, we assume the "experts" are linear-regression units, and set Y's CPD to linear-Gaussian. If the output is discrete, we set Y's CPD to a softmax function. The Q CPDs will always be softmax functions.

As a concrete example, consider the mixture of experts model where X and Y are scalars, and Q is binary. This is just piecewise linear regression, where we have two line segments, i.e.,

We can create this model with random parameters as follows. (This code is bundled in BNT/examples/static/mixexp2.m.)

X = 1;
Q = 2;
Y = 3;
dag = zeros(3,3);
dag(X,[Q Y]) = 1
dag(Q,Y) = 1;
ns = [1 2 1]; % make X and Y scalars, and have 2 experts
onodes = [1 3];
bnet = mk_bnet(dag, ns, 'discrete', 2, 'observed', onodes);

rand('state', 0);
randn('state', 0);
bnet.CPD{1} = root_CPD(bnet, 1);
bnet.CPD{2} = softmax_CPD(bnet, 2);
bnet.CPD{3} = gaussian_CPD(bnet, 3);
Now let us fit this model using
EM. First we load the data (1000 training cases) and plot them.

data = load('/examples/static/Misc/mixexp_data.txt', '-ascii');        
plot(data(:,1), data(:,2), '.');

This is what the model looks like before training. (Thanks to Thomas Hofman for writing this plotting routine.)

Now let's train the model, and plot the final performance. (We will discuss how to train models in more detail below.)

ncases = size(data, 1); % each row of data is a training case
cases = cell(3, ncases);
cases([1 3], :) = num2cell(data'); % each column of cases is a training case
engine = jtree_inf_engine(bnet);
max_iter = 20;
[bnet2, LLtrace] = learn_params_em(engine, cases, max_iter);
(We specify which nodes will be observed when we create the engine. Hence BNT knows that the hidden nodes are all discrete. For complex models, this can lead to a significant speedup.) Below we show what the model looks like after 16 iterations of EM (with 100 IRLS iterations per M step), when it converged using the default convergence tolerance (that the fractional change in the log-likelihood be less than 1e-3). Before learning, the log-likelihood was -322.927442; afterwards, it was -13.728778.

(See BNT/examples/static/mixexp2.m for details of the code.)

Hierarchical mixtures of experts

A hierarchical mixture of experts (HME) extends the mixture of experts model by having more than one hidden node. A two-level example is shown below, along with its more traditional representation as a neural network. This is like a (balanced) probabilistic decision tree of height 2.

Pierpaolo Brutti has written an extensive set of routines for HMEs, which are bundled with BNT: see the examples/static/HME directory. These routines allow you to choose the number of hidden (gating) layers, and the form of the experts (softmax or MLP). See the file hmemenu, which provides a demo. For example, the figure below shows the decision boundaries learned for a ternary classification problem, using a 2 level HME with softmax gates and softmax experts; the training set is on the left, the testing set on the right.

For more details, see the following:


Bayes nets originally arose out of an attempt to add probabilities to expert systems, and this is still the most common use for BNs. A famous example is QMR-DT, a decision-theoretic reformulation of the Quick Medical Reference (QMR) model.

Here, the top layer represents hidden disease nodes, and the bottom layer represents observed symptom nodes. The goal is to infer the posterior probability of each disease given all the symptoms (which can be present, absent or unknown). Each node in the top layer has a Bernoulli prior (with a low prior probability that the disease is present). Since each node in the bottom layer has a high fan-in, we use a noisy-OR parameterization; each disease has an independent chance of causing each symptom. The real QMR-DT model is copyright, but we can create a random QMR-like model as follows.
function bnet = mk_qmr_bnet(G, inhibit, leak, prior)
% MK_QMR_BNET Make a QMR model
% bnet = mk_qmr_bnet(G, inhibit, leak, prior)
% G(i,j) = 1 iff there is an arc from disease i to finding j
% inhibit(i,j) = inhibition probability on i->j arc
% leak(j) = inhibition prob. on leak->j arc
% prior(i) = prob. disease i is on

[Ndiseases Nfindings] = size(inhibit);
N = Ndiseases + Nfindings;
finding_node = Ndiseases+1:N;
ns = 2*ones(1,N);
dag = zeros(N,N);
dag(1:Ndiseases, finding_node) = G;
bnet = mk_bnet(dag, ns, 'observed', finding_node);

for d=1:Ndiseases
  CPT = [1-prior(d) prior(d)];
  bnet.CPD{d} = tabular_CPD(bnet, d, CPT');

for i=1:Nfindings
  fnode = finding_node(i);
  ps = parents(G, i);
  bnet.CPD{fnode} = noisyor_CPD(bnet, fnode, leak(i), inhibit(ps, i));
In the file BNT/examples/static/qmr1, we create a random bipartite graph G, with 5 diseases and 10 findings, and random parameters. (In general, to create a random dag, use 'mk_random_dag'.) We can visualize the resulting graph structure using the methods discussed
below, with the following results:

Now let us put some random evidence on all the leaves except the very first and very last, and compute the disease posteriors.

pos = 2:floor(Nfindings/2);
neg = (pos(end)+1):(Nfindings-1);
onodes = myunion(pos, neg);
evidence = cell(1, N);
evidence(findings(pos)) = num2cell(repmat(2, 1, length(pos)));
evidence(findings(neg)) = num2cell(repmat(1, 1, length(neg)));

engine = jtree_inf_engine(bnet);
[engine, ll] = enter_evidence(engine, evidence);
post = zeros(1, Ndiseases);
for i=diseases(:)'
  m = marginal_nodes(engine, i);
  post(i) = m.T(2);
Junction tree can be quite slow on large QMR models. Fortunately, it is possible to exploit properties of the noisy-OR function to speed up exact inference using an algorithm called quickscore, discussed below.

Conditional Gaussian models

A conditional Gaussian model is one in which, conditioned on all the discrete nodes, the distribution over the remaining (continuous) nodes is multivariate Gaussian. This means we can have arcs from discrete (D) to continuous (C) nodes, but not vice versa. (We are allowed C->D arcs if the continuous nodes are observed, as in the mixture of experts model, since this distribution can be represented with a discrete potential.)

We now give an example of a CG model, from the paper "Propagation of Probabilities, Means amd Variances in Mixed Graphical Association Models", Steffen Lauritzen, JASA 87(420):1098--1108, 1992 (reprinted in the book "Probabilistic Networks and Expert Systems", R. G. Cowell, A. P. Dawid, S. L. Lauritzen and D. J. Spiegelhalter, Springer, 1999.)

Specifying the graph

Consider the model of waste emissions from an incinerator plant shown below. We follow the standard convention that shaded nodes are observed, clear nodes are hidden. We also use the non-standard convention that square nodes are discrete (tabular) and round nodes are Gaussian.

We can create this model as follows.

F = 1; W = 2; E = 3; B = 4; C = 5; D = 6; Min = 7; Mout = 8; L = 9;
n = 9;

dag = zeros(n);
dag(W,[E Min D]) = 1;
dag(B,[C D])=1;
dag(D,[L Mout])=1;

% node sizes - all cts nodes are scalar, all discrete nodes are binary
ns = ones(1, n);
dnodes = [F W B];
cnodes = mysetdiff(1:n, dnodes);
ns(dnodes) = 2;

bnet = mk_bnet(dag, ns, 'discrete', dnodes);
'dnodes' is a list of the discrete nodes; 'cnodes' is the continuous nodes. 'mysetdiff' is a faster version of the built-in 'setdiff'.

Specifying the parameters

The parameters of the discrete nodes can be specified as follows.
bnet.CPD{B} = tabular_CPD(bnet, B, 'CPT', [0.85 0.15]); % 1=stable, 2=unstable
bnet.CPD{F} = tabular_CPD(bnet, F, 'CPT', [0.95 0.05]); % 1=intact, 2=defect
bnet.CPD{W} = tabular_CPD(bnet, W, 'CPT', [2/7 5/7]); % 1=industrial, 2=household

The parameters of the continuous nodes can be specified as follows.

bnet.CPD{E} = gaussian_CPD(bnet, E, 'mean', [-3.9 -0.4 -3.2 -0.5], ...
			   'cov', [0.00002 0.0001 0.00002 0.0001]);
bnet.CPD{D} = gaussian_CPD(bnet, D, 'mean', [6.5 6.0 7.5 7.0], ...
			   'cov', [0.03 0.04 0.1 0.1], 'weights', [1 1 1 1]);
bnet.CPD{C} = gaussian_CPD(bnet, C, 'mean', [-2 -1], 'cov', [0.1 0.3]);
bnet.CPD{L} = gaussian_CPD(bnet, L, 'mean', 3, 'cov', 0.25, 'weights', -0.5);
bnet.CPD{Min} = gaussian_CPD(bnet, Min, 'mean', [0.5 -0.5], 'cov', [0.01 0.005]);
bnet.CPD{Mout} = gaussian_CPD(bnet, Mout, 'mean', 0, 'cov', 0.002, 'weights', [1 1]);


First we compute the unconditional marginals.
engine = jtree_inf_engine(bnet);
evidence = cell(1,n);
[engine, ll] = enter_evidence(engine, evidence);
marg = marginal_nodes(engine, E);
'marg' is a structure that contains the fields 'mu' and 'Sigma', which contain the mean and (co)variance of the marginal on E. In this case, they are both scalars. Let us check they match the published figures (to 2 decimal places).
tol = 1e-2;
assert(approxeq(, -3.25, tol));
assert(approxeq(sqrt(marg.Sigma), 0.709, tol));
We can compute the other posteriors similarly. Now let us add some evidence.
evidence = cell(1,n);
evidence{W} = 1; % industrial
evidence{L} = 1.1;
evidence{C} = -0.9;
[engine, ll] = enter_evidence(engine, evidence);
Now we find
marg = marginal_nodes(engine, E);
assert(approxeq(, -3.8983, tol));
assert(approxeq(sqrt(marg.Sigma), 0.0763, tol));
We can also compute the joint probability on a set of nodes. For example, P(D, Mout | evidence) is a 2D Gaussian:
marg = marginal_nodes(engine, [D Mout])
marg = 
    domain: [6 8]
        mu: [2x1 double]
     Sigma: [2x2 double]
         T: 1.0000
The mean is
ans =
and the covariance matrix is
ans =
    0.1062    0.1062
    0.1062    0.1182
It is easy to visualize this posterior using standard Matlab plotting functions, e.g.,
gaussplot2d(, marg.Sigma);
produces the following picture.

The T field indicates that the mixing weight of this Gaussian component is 1.0. If the joint contains discrete and continuous variables, the result will be a mixture of Gaussians, e.g.,

marg = marginal_nodes(engine, [F E])
    domain: [1 3]
        mu: [-3.9000 -0.4003]
     Sigma: [1x1x2 double]
         T: [0.9995 4.7373e-04]
The interpretation is Sigma(i,j,k) = Cov[ E(i) E(j) | F=k ]. In this case, E is a scalar, so i=j=1; k specifies the mixture component.

We saw in the sprinkler network that BNT sets the effective size of observed discrete nodes to 1, since they only have one legal value. For continuous nodes, BNT sets their length to 0, since they have been reduced to a point. For example,

marg = marginal_nodes(engine, [B C])
    domain: [4 5]
        mu: []
     Sigma: []
         T: [0.0123 0.9877]
It is simple to post-process the output of marginal_nodes. For example, the file BNT/examples/static/cg1 sets the mu term of observed nodes to their observed value, and the Sigma term to 0 (since observed nodes have no variance).

Note that the implemented version of the junction tree is numerically unstable when using CG potentials (which is why, in the example above, we only required our answers to agree with the published ones to 2dp.) This is why you might want to use stab_cond_gauss_inf_engine, implemented by Shan Huang. This is described in

However, even the numerically stable version can be computationally intractable if there are many hidden discrete nodes, because the number of mixture components grows exponentially e.g., in a
switching linear dynamical system. In general, one must resort to approximate inference techniques: see the discussion on inference engines below.

Other hybrid models

When we have C->D arcs, where C is hidden, we need to use approximate inference. One approach (not implemented in BNT) is described in Of course, one can always use sampling methods for approximate inference in such models.

Parameter Learning

The parameter estimation routines in BNT can be classified into 4 types, depending on whether the goal is to compute a full (Bayesian) posterior over the parameters or just a point estimate (e.g., Maximum Likelihood or Maximum A Posteriori), and whether all the variables are fully observed or there is missing data/ hidden variables (partial observability).

Full obs Partial obs
Point learn_params learn_params_em
Bayes bayes_update_params not yet supported

Loading data from a file

To load numeric data from an ASCII text file called 'dat.txt', where each row is a case and columns are separated by white-space, such as
011979 1626.5 0.0
021979 1367.0 0.0
you can use
data = load('dat.txt');
load dat.txt -ascii
In the latter case, the data is stored in a variable called 'dat' (the filename minus the extension). Alternatively, suppose the data is stored in a .csv file (has commas separating the columns, and contains a header line), such as
header info goes here
You can load this using
[a,b,c,d] = textread('dat.txt', '%s %d %f %f', 'delimiter', ',', 'headerlines', 1);
If your file is not in either of these formats, you can either use Perl to convert it to this format, or use the Matlab scanf command. Type help iofun for more information on Matlab's file functions.

BNT learning routines require data to be stored in a cell array. data{i,m} is the value of node i in case (example) m, i.e., each column is a case. If node i is not observed in case m (missing value), set data{i,m} = []. (Not all the learning routines can cope with such missing values, however.) In the special case that all the nodes are observed and are scalar-valued (as opposed to vector-valued), the data can be stored in a matrix (as opposed to a cell-array).

Suppose, as in the mixture of experts example, that we have 3 nodes in the graph: X(1) is the observed input, X(3) is the observed output, and X(2) is a hidden (gating) node. We can create the dataset as follows.

data = load('dat.txt');
ncases = size(data, 1);
cases = cell(3, ncases);
cases([1 3], :) = num2cell(data');
Notice how we transposed the data, to convert rows into columns. Also, cases{2,m} = [] for all m, since X(2) is always hidden.

Maximum likelihood parameter estimation from complete data

As an example, let's generate some data from the sprinkler network, randomize the parameters, and then try to recover the original model. First we create some training data using forwards sampling.
samples = cell(N, nsamples);
for i=1:nsamples
  samples(:,i) = sample_bnet(bnet);
samples{j,i} contains the value of the j'th node in case i. sample_bnet returns a cell array because, in general, each node might be a vector of different length. In this case, all nodes are discrete (and hence scalars), so we could have used a regular array instead (which can be quicker):
data = cell2num(samples);
Now we create a network with random parameters. (The initial values of bnet2 don't matter in this case, since we can find the globally optimal MLE independent of where we start.)
% Make a tabula rasa
bnet2 = mk_bnet(dag, node_sizes);
seed = 0;
rand('state', seed);
bnet2.CPD{C} = tabular_CPD(bnet2, C);
bnet2.CPD{R} = tabular_CPD(bnet2, R);
bnet2.CPD{S} = tabular_CPD(bnet2, S);
bnet2.CPD{W} = tabular_CPD(bnet2, W);
Finally, we find the maximum likelihood estimates of the parameters.
bnet3 = learn_params(bnet2, samples);
To view the learned parameters, we use a little Matlab hackery.
CPT3 = cell(1,N);
for i=1:N
  s=struct(bnet3.CPD{i});  % violate object privacy
Here are the parameters learned for node 4.
1 1 : 1.0000 0.0000 
2 1 : 0.2000 0.8000 
1 2 : 0.2273 0.7727 
2 2 : 0.0000 1.0000 
So we see that the learned parameters are fairly close to the "true" ones, which we display below.
1 1 : 1.0000 0.0000 
2 1 : 0.1000 0.9000 
1 2 : 0.1000 0.9000 
2 2 : 0.0100 0.9900 
We can get better results by using a larger training set, or using informative priors (see

Parameter priors

Currently, only tabular CPDs can have priors on their parameters. The conjugate prior for a multinomial is the Dirichlet. (For binary random variables, the multinomial is the same as the Bernoulli, and the Dirichlet is the same as the Beta.)

The Dirichlet has a simple interpretation in terms of pseudo counts. If we let N_ijk = the num. times X_i=k and Pa_i=j occurs in the training set, where Pa_i are the parents of X_i, then the maximum likelihood (ML) estimate is T_ijk = N_ijk / N_ij (where N_ij = sum_k' N_ijk'), which will be 0 if N_ijk=0. To prevent us from declaring that (X_i=k, Pa_i=j) is impossible just because this event was not seen in the training set, we can pretend we saw value k of X_i, for each value j of Pa_i some number (alpha_ijk) of times in the past. The MAP (maximum a posterior) estimate is then

T_ijk = (N_ijk + alpha_ijk) / (N_ij + alpha_ij)
and is never 0 if all alpha_ijk > 0. For example, consider the network A->B, where A is binary and B has 3 values. A uniform prior for B has the form
    B=1 B=2 B=3
A=1 1   1   1
A=2 1   1   1
which can be created using
tabular_CPD(bnet, i, 'prior_type', 'dirichlet', 'dirichlet_type', 'unif');
This prior does not satisfy the likelihood equivalence principle, which says that
Markov equivalent models should have the same marginal likelihood. A prior that does satisfy this principle is shown below. Heckerman (1995) calls this the BDeu prior (likelihood equivalent uniform Bayesian Dirichlet).
    B=1 B=2 B=3
A=1 1/6 1/6 1/6
A=2 1/6 1/6 1/6
where we put N/(q*r) in each bin; N is the equivalent sample size, r=|A|, q = |B|. This can be created as follows
tabular_CPD(bnet, i, 'prior_type', 'dirichlet', 'dirichlet_type', 'BDeu');
Here, 1 is the equivalent sample size, and is the strength of the prior. You can change this using
tabular_CPD(bnet, i, 'prior_type', 'dirichlet', 'dirichlet_type', ...
   'BDeu', 'dirichlet_weight', 10);

(Sequential) Bayesian parameter updating from complete data

If we use conjugate priors and have fully observed data, we can compute the posterior over the parameters in batch form as follows.
cases = sample_bnet(bnet, nsamples);
bnet = bayes_update_params(bnet, cases);  
LL = log_marg_lik_complete(bnet, cases);   
bnet.CPD{i}.prior contains the new Dirichlet pseudocounts, and bnet.CPD{i}.CPT is set to the mean of the posterior (the normalized counts). (Hence if the initial pseudo counts are 0, bayes_update_params and learn_params will give the same result.)

We can compute the same result sequentially (on-line) as follows.

LL = 0;
for m=1:nsamples
  LL = LL + log_marg_lik_complete(bnet, cases(:,m));
  bnet = bayes_update_params(bnet, cases(:,m));
The file BNT/examples/static/StructLearn/model_select1 has an example of sequential model selection which uses the same idea. We generate data from the model A->B and compute the posterior prob of all 3 dags on 2 nodes: (1) A B, (2) A <- B , (3) A -> B Models 2 and 3 are
Markov equivalent, and therefore indistinguishable from observational data alone, so we expect their posteriors to be the same (assuming a prior which satisfies likelihood equivalence). If we use random parameters, the "true" model only gets a higher posterior after 2000 trials! However, if we make B a noisy NOT gate, the true model "wins" after 12 trials, as shown below (red = model 1, blue/green (superimposed) represents models 2/3).

The use of marginal likelihood for model selection is discussed in greater detail in the section on structure learning.

Maximum likelihood parameter estimation with missing values

Now we consider learning when some values are not observed. Let us randomly hide half the values generated from the water sprinkler example.
samples2 = samples;
hide = rand(N, nsamples) > 0.5;
for k=1:length(I)
  samples2{I(k), J(k)} = [];
samples2{i,l} is the value of node i in training case l, or [] if unobserved.

Now we will compute the MLEs using the EM algorithm. We need to use an inference algorithm to compute the expected sufficient statistics in the E step; the M (maximization) step is as above.

engine2 = jtree_inf_engine(bnet2);
max_iter = 10;
[bnet4, LLtrace] = learn_params_em(engine2, samples2, max_iter);
LLtrace(i) is the log-likelihood at iteration i. We can plot this as follows:
plot(LLtrace, 'x-')
Let's display the results after 10 iterations of EM.
CPT4{1} =
CPT4{2} =
    0.6510    0.3490
    0.8751    0.1249
CPT4{3} =
    0.8366    0.1634
    0.0197    0.9803
CPT4{4} =
(:,:,1) =
    0.8276    0.0546
    0.5452    0.1658
(:,:,2) =
    0.1724    0.9454
    0.4548    0.8342
We can get improved performance by using one or more of the following methods: Click
here for a discussion of learning Gaussians, which can cause numerical problems.

For a more complete example of learning with EM, see the script BNT/examples/static/learn1.m.

Parameter tying

In networks with repeated structure (e.g., chains and grids), it is common to assume that the parameters are the same at every node. This is called parameter tying, and reduces the amount of data needed for learning.

When we have tied parameters, there is no longer a one-to-one correspondence between nodes and CPDs. Rather, each CPD species the parameters for a whole equivalence class of nodes. It is easiest to see this by example. Consider the following hidden Markov model (HMM)

When HMMs are used for semi-infinite processes like speech recognition, we assume the transition matrix P(H(t+1)|H(t)) is the same for all t; this is called a time-invariant or homogenous Markov chain. Hence hidden nodes 2, 3, ..., T are all in the same equivalence class, say class Hclass. Similarly, the observation matrix P(O(t)|H(t)) is assumed to be the same for all t, so the observed nodes are all in the same equivalence class, say class Oclass. Finally, the prior term P(H(1)) is in a class all by itself, say class H1class. This is illustrated below, where we explicitly represent the parameters as random variables (dotted nodes).

In BNT, we cannot represent parameters as random variables (nodes). Instead, we "hide" the parameters inside one CPD for each equivalence class, and then specify that the other CPDs should share these parameters, as follows.

hnodes = 1:2:2*T;
onodes = 2:2:2*T;
H1class = 1; Hclass = 2; Oclass = 3;
eclass = ones(1,N);
eclass(hnodes(2:end)) = Hclass;
eclass(hnodes(1)) = H1class;
eclass(onodes) = Oclass;
% create dag and ns in the usual way
bnet = mk_bnet(dag, ns, 'discrete', dnodes, 'equiv_class', eclass);
Finally, we define the parameters for each equivalence class:
bnet.CPD{H1class} = tabular_CPD(bnet, hnodes(1)); % prior
bnet.CPD{Hclass} = tabular_CPD(bnet, hnodes(2)); % transition matrix
if cts_obs
  bnet.CPD{Oclass} = gaussian_CPD(bnet, onodes(1));
  bnet.CPD{Oclass} = tabular_CPD(bnet, onodes(1));
In general, if bnet.CPD{e} = xxx_CPD(bnet, j), then j should be a member of e's equivalence class; that is, it is not always the case that e == j. You can use bnet.rep_of_eclass(e) to return the representative of equivalence class e. BNT will look up the parents of j to determine the size of the CPT to use. It assumes that this is the same for all members of the equivalence class. Click here for a more complex example of parameter tying.

Note: Normally one would define an HMM as a Dynamic Bayes Net (see the function BNT/examples/dynamic/mk_chmm.m). However, one can define an HMM as a static BN using the function BNT/examples/static/Models/mk_hmm_bnet.m.

Structure learning

Update (9/29/03): Phillipe LeRay is developing some additional structure learning code on top of BNT. Click here for details.

There are two very different approaches to structure learning: constraint-based and search-and-score. In the constraint-based approach, we start with a fully connected graph, and remove edges if certain conditional independencies are measured in the data. This has the disadvantage that repeated independence tests lose statistical power.

In the more popular search-and-score approach, we perform a search through the space of possible DAGs, and either return the best one found (a point estimate), or return a sample of the models found (an approximation to the Bayesian posterior).

The number of DAGs as a function of the number of nodes, G(n), is super-exponential in n, and is given by the following recurrence

The first few values are shown below.
n G(n)
1 1
2 3
3 25
4 543
5 29,281
6 3,781,503
7 1.1 x 10^9
8 7.8 x 10^11
9 1.2 x 10^15
10 4.2 x 10^18
Since the number of DAGs is super-exponential in the number of nodes, we cannot exhaustively search the space, so we either use a local search algorithm (e.g., greedy hill climbining, perhaps with multiple restarts) or a global search algorithm (e.g., Markov Chain Monte Carlo).

If we know a total ordering on the nodes, finding the best structure amounts to picking the best set of parents for each node independently. This is what the K2 algorithm does. If the ordering is unknown, we can search over orderings, which is more efficient than searching over DAGs (Koller and Friedman, 2000).

In addition to the search procedure, we must specify the scoring function. There are two popular choices. The Bayesian score integrates out the parameters, i.e., it is the marginal likelihood of the model. The BIC (Bayesian Information Criterion) is defined as log P(D|theta_hat) - 0.5*d*log(N), where D is the data, theta_hat is the ML estimate of the parameters, d is the number of parameters, and N is the number of data cases. The BIC method has the advantage of not requiring a prior.

BIC can be derived as a large sample approximation to the marginal likelihood. (It is also equal to the Minimum Description Length of a model.) However, in practice, the sample size does not need to be very large for the approximation to be good. For example, in the figure below, we plot the ratio between the log marginal likelihood and the BIC score against data-set size; we see that the ratio rapidly approaches 1, especially for non-informative priors. (This plot was generated by the file BNT/examples/static/bic1.m. It uses the water sprinkler BN with BDeu Dirichlet priors with different equivalent sample sizes.)

As with parameter learning, handling missing data/ hidden variables is much harder than the fully observed case. The structure learning routines in BNT can therefore be classified into 4 types, analogously to the parameter learning case.

Full obs Partial obs
Point learn_struct_K2
not yet supported
Bayes learn_struct_mcmc not yet supported

Markov equivalence

If two DAGs encode the same conditional independencies, they are called Markov equivalent. The set of all DAGs can be paritioned into Markov equivalence classes. Graphs within the same class can have the direction of some of their arcs reversed without changing any of the CI relationships. Each class can be represented by a PDAG (partially directed acyclic graph) called an essential graph or pattern. This specifies which edges must be oriented in a certain direction, and which may be reversed.

When learning graph structure from observational data, the best one can hope to do is to identify the model up to Markov equivalence. To distinguish amongst graphs within the same equivalence class, one needs interventional data: see the discussion on active learning below.

Exhaustive search

The brute-force approach to structure learning is to enumerate all possible DAGs, and score each one. This provides a "gold standard" with which to compare other algorithms. We can do this as follows.
dags = mk_all_dags(N);
score = score_dags(data, ns, dags);
where data(i,m) is the value of node i in case m, and ns(i) is the size of node i. If the DAGs have a lot of families in common, we can cache the sufficient statistics, making this potentially more efficient than scoring the DAGs one at a time. (Caching is not currently implemented, however.)

By default, we use the Bayesian scoring metric, and assume CPDs are represented by tables with BDeu(1) priors. We can override these defaults as follows. If we want to use uniform priors, we can say

params = cell(1,N);
for i=1:N
  params{i} = {'prior', 'unif'};
score = score_dags(data, ns, dags, 'params', params);
params{i} is a cell-array, containing optional arguments that are passed to the constructor for CPD i.

Now suppose we want to use different node types, e.g., Suppose nodes 1 and 2 are Gaussian, and nodes 3 and 4 softmax (both these CPDs can support discrete and continuous parents, which is necessary since all other nodes will be considered as parents). The Bayesian scoring metric currently only works for tabular CPDs, so we will use BIC:

score = score_dags(data, ns, dags, 'discrete', [3 4], 'params', [], 
    'type', {'gaussian', 'gaussian', 'softmax', softmax'}, 'scoring_fn', 'bic')
In practice, one can't enumerate all possible DAGs for N > 5, but one can evaluate any reasonably-sized set of hypotheses in this way (e.g., nearest neighbors of your current best guess). Think of this as "computer assisted model refinement" as opposed to de novo learning.


The K2 algorithm (Cooper and Herskovits, 1992) is a greedy search algorithm that works as follows. Initially each node has no parents. It then adds incrementally that parent whose addition most increases the score of the resulting structure. When the addition of no single parent can increase the score, it stops adding parents to the node. Since we are using a fixed ordering, we do not need to check for cycles, and can choose the parents for each node independently.

The original paper used the Bayesian scoring metric with tabular CPDs and Dirichlet priors. BNT generalizes this to allow any kind of CPD, and either the Bayesian scoring metric or BIC, as in the example above. In addition, you can specify an optional upper bound on the number of parents for each node. The file BNT/examples/static/k2demo1.m gives an example of how to use K2. We use the water sprinkler network and sample 100 cases from it as before. Then we see how much data it takes to recover the generating structure:

order = [C S R W];
max_fan_in = 2;
sz = 5:5:100;
for i=1:length(sz)
  dag2 = learn_struct_K2(data(:,1:sz(i)), node_sizes, order, 'max_fan_in', max_fan_in);
  correct(i) = isequal(dag, dag2);
Here are the results.
correct =
  Columns 1 through 12 
     0     0     0     0     0     0     0     1     0     1     1     1
  Columns 13 through 20 
     1     1     1     1     1     1     1     1
So we see it takes about sz(10)=50 cases. (BIC behaves similarly, showing that the prior doesn't matter too much.) In general, we cannot hope to recover the "true" generating structure, only one that is in its Markov equivalence class.


Hill-climbing starts at a specific point in space, considers all nearest neighbors, and moves to the neighbor that has the highest score; if no neighbors have higher score than the current point (i.e., we have reached a local maximum), the algorithm stops. One can then restart in another part of the space.

A common definition of "neighbor" is all graphs that can be generated from the current graph by adding, deleting or reversing a single arc, subject to the acyclicity constraint. Other neighborhoods are possible: see Optimal Structure Identification with Greedy Search, Max Chickering, JMLR 2002.


We can use a Markov Chain Monte Carlo (MCMC) algorithm called Metropolis-Hastings (MH) to search the space of all DAGs. The standard proposal distribution is to consider moving to all nearest neighbors in the sense defined above.

The function can be called as in the following example.

[sampled_graphs, accept_ratio] = learn_struct_mcmc(data, ns, 'nsamples', 100, 'burnin', 10);
We can convert our set of sampled graphs to a histogram (empirical posterior over all the DAGs) thus
all_dags = mk_all_dags(N);
mcmc_post = mcmc_sample_to_hist(sampled_graphs, all_dags);
To see how well this performs, let us compute the exact posterior exhaustively.

score = score_dags(data, ns, all_dags);
post = normalise(exp(score)); % assuming uniform structural prior
We plot the results below. (The data set was 100 samples drawn from a random 4 node bnet; see the file BNT/examples/static/mcmc1.)

We can also plot the acceptance ratio versus number of MCMC steps, as a crude convergence diagnostic.


Even though the number of samples needed by MCMC is theoretically polynomial (not exponential) in the dimensionality of the search space, in practice it has been found that MCMC does not converge in reasonable time for graphs with more than about 10 nodes.

Active structure learning

As was mentioned above, one can only learn a DAG up to Markov equivalence, even given infinite data. If one is interested in learning the structure of a causal network, one needs interventional data. (By "intervention" we mean forcing a node to take on a specific value, thereby effectively severing its incoming arcs.)

Most of the scoring functions accept an optional argument that specifies whether a node was observed to have a certain value, or was forced to have that value: we set clamped(i,m)=1 if node i was forced in training case m. e.g., see the file BNT/examples/static/cooper_yoo.

An interesting question is to decide which interventions to perform (c.f., design of experiments). For details, see the following tech report

Structural EM

Computing the Bayesian score when there is partial observability is computationally challenging, because the parameter posterior becomes multimodal (the hidden nodes induce a mixture distribution). One therefore needs to use approximations such as BIC. Unfortunately, search algorithms are still expensive, because we need to run EM at each step to compute the MLE, which is needed to compute the score of each model. An alternative approach is to do the local search steps inside of the M step of EM, which is more efficient since the data has been "filled in" - this is called the structural EM algorithm (Friedman 1997), and provably converges to a local maximum of the BIC score.

Wei Hu has implemented SEM for discrete nodes. You can download his package from here. Please address all questions about this code to See also Phl's implementation of SEM.

Visualizing the graph

Click here for more information on graph visualization.

Constraint-based methods

The IC algorithm (Pearl and Verma, 1991), and the faster, but otherwise equivalent, PC algorithm (Spirtes, Glymour, and Scheines 1993), computes many conditional independence tests, and combines these constraints into a PDAG to represent the whole Markov equivalence class.

IC*/FCI extend IC/PC to handle latent variables: see below. (IC stands for inductive causation; PC stands for Peter and Clark, the first names of Spirtes and Glymour; FCI stands for fast causal inference. What we, following Pearl (2000), call IC* was called IC in the original Pearl and Verma paper.) For details, see

The PC algorithm takes as arguments a function f, the number of nodes N, the maximum fan in K, and additional arguments A which are passed to f. The function f(X,Y,S,A) returns 1 if X is conditionally independent of Y given S, and 0 otherwise. For example, suppose we cheat by passing in a CI "oracle" which has access to the true DAG; the oracle tests for d-separation in this DAG, i.e., f(X,Y,S) calls dsep(X,Y,S,dag). We can to this as follows.

pdag = learn_struct_pdag_pc('dsep', N, max_fan_in, dag);
pdag(i,j) = -1 if there is definitely an i->j arc, and pdag(i,j) = 1 if there is either an i->j or and i<-j arc.

Applied to the sprinkler network, this returns

pdag =
     0     1     1     0
     1     0     0    -1
     1     0     0    -1
     0     0     0     0
So as expected, we see that the V-structure at the W node is uniquely identified, but the other arcs have ambiguous orientation.

We now give an example from p141 (1st edn) / p103 (2nd end) of the SGS book. This example concerns the female orgasm. We are given a correlation matrix C between 7 measured factors (such as subjective experiences of coital and masturbatory experiences), derived from 281 samples, and want to learn a causal model of the data. We will not discuss the merits of this type of work here, but merely show how to reproduce the results in the SGS book. Their program, Tetrad, makes use of the Fisher Z-test for conditional independence, so we do the same:

max_fan_in = 4;
nsamples = 281;
alpha = 0.05;
pdag = learn_struct_pdag_pc('cond_indep_fisher_z', n, max_fan_in, C, nsamples, alpha);
In this case, the CI test is
f(X,Y,S) = cond_indep_fisher_z(X,Y,S,  C,nsamples,alpha)
The results match those of Fig 12a of SGS apart from two edge differences; presumably this is due to rounding error (although it could be a bug, either in BNT or in Tetrad). This example can be found in the file BNT/examples/static/pc2.m.

The IC* algorithm (Pearl and Verma, 1991), and the faster FCI algorithm (Spirtes, Glymour, and Scheines 1993), are like the IC/PC algorithm, except that they can detect the presence of latent variables. See the file learn_struct_pdag_ic_star written by Tamar Kushnir. The output is a matrix P, defined as follows (see Pearl (2000), p52 for details):

% P(i,j) = -1 if there is either a latent variable L such that i <-L->j OR there is a directed edge from i->j.
% P(i,j) = -2 if there is a marked directed i-*>j edge.
% P(i,j) = P(j,i) = 1 if there is and undirected edge i--j
% P(i,j) = P(j,i) = 2 if there is a latent variable L such that i<-L->j.

Philippe Leray's structure learning package

Philippe Leray has written a structure learning package that uses BNT. It currently (Juen 2003) has the following features:

Inference engines

Up until now, we have used the junction tree algorithm for inference. However, sometimes this is too slow, or not even applicable. In general, there are many inference algorithms each of which make different tradeoffs between speed, accuracy, complexity and generality. Furthermore, there might be many implementations of the same algorithm; for instance, a general purpose, readable version, and a highly-optimized, specialized one. To cope with this variety, we treat each inference algorithm as an object, which we call an inference engine.

An inference engine is an object that contains a bnet and supports the 'enter_evidence' and 'marginal_nodes' methods. The engine constructor takes the bnet as argument and may do some model-specific processing. When 'enter_evidence' is called, the engine may do some evidence-specific processing. Finally, when 'marginal_nodes' is called, the engine may do some query-specific processing.

The amount of work done when each stage is specified -- structure, parameters, evidence, and query -- depends on the engine. The cost of work done early in this sequence can be amortized. On the other hand, one can make better optimizations if one waits until later in the sequence. For example, the parameters might imply conditional indpendencies that are not evident in the graph structure, but can nevertheless be exploited; the evidence indicates which nodes are observed and hence can effectively be disconnected from the graph; and the query might indicate that large parts of the network are d-separated from the query nodes. (Since it is not the actual values of the evidence that matters, just which nodes are observed, many engines allow you to specify which nodes will be observed when they are constructed, i.e., before calling 'enter_evidence'. Some engines can still cope if the actual pattern of evidence is different, e.g., if there is missing data.)

Although being maximally lazy (i.e., only doing work when a query is issued) may seem desirable, this is not always the most efficient. For example, when learning using EM, we need to call marginal_nodes N times, where N is the number of nodes. Variable elimination would end up repeating a lot of work each time marginal_nodes is called, making it inefficient for learning. The junction tree algorithm, by contrast, uses dynamic programming to avoid this redundant computation --- it calculates all marginals in two passes during 'enter_evidence', so calling 'marginal_nodes' takes constant time.

We will discuss some of the inference algorithms implemented in BNT below, and finish with a summary of all of them.

Variable elimination

The variable elimination algorithm, also known as bucket elimination or peeling, is one of the simplest inference algorithms. The basic idea is to "push sums inside of products"; this is explained in more detail here.

The principle of distributing sums over products can be generalized greatly to apply to any commutative semiring. This forms the basis of many common algorithms, such as Viterbi decoding and the Fast Fourier Transform. For details, see

Choosing an order in which to sum out the variables so as to minimize computational cost is known to be NP-hard. The implementation of this algorithm in var_elim_inf_engine makes no attempt to optimize this ordering (in contrast, say, to jtree_inf_engine, which uses a greedy search procedure to find a good ordering).

Note: unlike most algorithms, var_elim does all its computational work inside of marginal_nodes, not inside of enter_evidence.

Global inference methods

The simplest inference algorithm of all is to explicitely construct the joint distribution over all the nodes, and then to marginalize it. This is implemented in global_joint_inf_engine. Since the size of the joint is exponential in the number of discrete (hidden) nodes, this is not a very practical algorithm. It is included merely for pedagogical and debugging purposes.

Three specialized versions of this algorithm have also been implemented, corresponding to the cases where all the nodes are discrete (D), all are Gaussian (G), and some are discrete and some Gaussian (CG). They are called enumerative_inf_engine, gaussian_inf_engine, and cond_gauss_inf_engine respectively.

Note: unlike most algorithms, these global inference algorithms do all their computational work inside of marginal_nodes, not inside of enter_evidence.


The junction tree algorithm is quite slow on the QMR network, since the cliques are so big. One simple trick we can use is to notice that hidden leaves do not affect the posteriors on the roots, and hence do not need to be included in the network. A second trick is to notice that the negative findings can be "absorbed" into the prior: see the file BNT/examples/static/mk_minimal_qmr_bnet for details.

A much more significant speedup is obtained by exploiting special properties of the noisy-or node, as done by the quickscore algorithm. For details, see

This has been implemented in BNT as a special-purpose inference engine, which can be created and used as follows:
engine = quickscore_inf_engine(inhibit, leak, prior);
engine = enter_evidence(engine, pos, neg);
m = marginal_nodes(engine, i);

Belief propagation

Even using quickscore, exact inference takes time that is exponential in the number of positive findings. Hence for large networks we need to resort to approximate inference techniques. See for example The latter approximation entails applying Pearl's belief propagation algorithm to a model even if it has loops (hence the name loopy belief propagation). Pearl's algorithm, implemented as pearl_inf_engine, gives exact results when applied to singly-connected graphs (a.k.a. polytrees, since the underlying undirected topology is a tree, but a node may have multiple parents). To apply this algorithm to a graph with loops, use pearl_inf_engine. This can use a centralized or distributed message passing protocol. You can use it as in the following example.
engine = pearl_inf_engine(bnet, 'max_iter', 30);
engine = enter_evidence(engine, evidence);
m = marginal_nodes(engine, i);
We found that this algorithm often converges, and when it does, often is very accurate, but it depends on the precise setting of the parameter values of the network. (See the file BNT/examples/static/qmr1 to repeat the experiment for yourself.) Understanding when and why belief propagation converges/ works is a topic of ongoing research.

pearl_inf_engine can exploit special structure in noisy-or and gmux nodes to compute messages efficiently.

belprop_inf_engine is like pearl, but uses potentials to represent messages. Hence this is slower.

belprop_fg_inf_engine is like belprop, but is designed for factor graphs.


BNT now (Mar '02) has two sampling (Monte Carlo) inference algorithms: Note: To generate samples from a network (which is not the same as inference!), use sample_bnet.

Summary of inference engines

The inference engines differ in many ways. Here are some of the major "axes":

In terms of topology, most engines handle any kind of DAG. belprop_fg does approximate inference on factor graphs (FG), which can be used to represent directed, undirected, and mixed (chain) graphs. (In the future, we plan to support exact inference on chain graphs.) quickscore only works on QMR-like models.

In terms of node types: algorithms that use potentials can handle discrete (D), Gaussian (G) or conditional Gaussian (CG) models. Sampling algorithms can essentially handle any kind of node (distribution). Other algorithms make more restrictive assumptions in exchange for speed.

Finally, most algorithms are designed to give the exact answer. The belief propagation algorithms are exact if applied to trees, and in some other cases. Sampling is considered approximate, even though, in the limit of an infinite number of samples, it gives the exact answer.

Here is a summary of the properties of all the engines in BNT which work on static networks.

Name Exact? Node type? topology
belprop approx D DAG
belprop_fg approx D factor graph
cond_gauss exact CG DAG
enumerative exact D DAG
gaussian exact G DAG
gibbs approx D DAG
global_joint exact D,G,CG DAG
jtree exact D,G,CG DAG b
likelihood_weighting approx any DAG
pearl approx D,G DAG
pearl exact D,G polytree
quickscore exact noisy-or QMR
stab_cond_gauss exact CG DAG
var_elim exact D,G,CG DAG

Influence diagrams/ decision making

BNT implements an exact algorithm for solving LIMIDs (limited memory influence diagrams), described in LIMIDs explicitely show all information arcs, rather than implicitely assuming no forgetting. This allows them to model forgetful controllers.

See the examples in BNT/examples/limids for details.

DBNs, HMMs, Kalman filters and all that

Click here for documentation about how to use BNT for dynamical systems and sequence data.