Graphical Model

The FactorGraph supports the composite type ContinuousModel based on the synchronous message passing schedule, with three fields:

• ContinuousGraph;
• ContinuousInference;
• ContinuousSystem.

The subtype ContinuousGraph describes the factor graph obtained based on the input data. The GBP inference and marginal values are kept in the subtype ContinuousInference. The system of the linear equations being solved is preserved in the subtype ContinuousSystem. Note that the function continuousModel() returns the main FactorGraph composite type ContinuousModel with all subtypes.

In addition, we also provide several functions for factor graph manipulation.

Build graphical model

Input arguments of the function continuousModel() describe the graphical model, while the function returns ContinuousModel type.

Loads the system data passing arguments:

gbp = continuousModel(coefficient, observation, variances)

Virtual factor nodes

The function continuousModel() receives arguments by keyword to set the mean and variance of the virtual factor nodes. We advise the reader to read the Section initialisation procedure which provides a detailed description of the virtual factor nodes.

gbp = continuousModel(DATA; mean = value, variance = value)

Default setting of the mean value is mean = 0.0, while the default variance is equal to variance = 1e10.

Randomized damping parametars

The function continuousModel() receives arguments by keyword to set damping parametars. We advise the reader to read the section the GBP with randomized damping which provides a detailed description of the input parameters.

gbp = continuousModel(DATA; prob = value, alpha = value)

The keyword prob represents the probability of the Bernoulli random variable, independently sampled for each mean value message from a factor node to a variable node, applied for randomised damping iteration scheme with value between 0 and 1. Default setting is set to prob = 0.6. The damped message is evaluated as a linear combination of the message from the previous and the current iteration, with weights alpha = value and 1 - alpha, applied for randomised damping iteration scheme where alpha is between 0 and 1. Default setting is set to alpha = 0.4.

Using the function continuousModel(), the set of damp messages are fixed through GBP iterations. However, we provide the function that changes damp parameters prob and alpha on the fly:

damping!(gbp; prob = value, alpha = value)

Freeze factor node, variable node or edge

The functions freeze target factor or variable node, whereby all messages sent by the factor or variable node retain latest obtained values at the time of freezing.

freezeFactor!(gbp; factor = index)
freezeVariable!(gbp; variable = index)

We provide functions that freeze the target edge. More precisely, the function freezes the message from variable node to factor node, or the message from factor node to variable node. Hence, the frozen message keeps the last value obtained at the time of freezing.

freezeVariableFactor!(gbp; variable = index, factor = index)
freezeFactorVariable!(gbp; factor = index, variable = index)

The functions accept following parameters: composite type ContinuousModel; the factor node index corresponding to the row index of the coefficient matrix; and the variable node index corresponding to the column index of the coefficient matrix. Note that the singly connected factor nodes can not be frozen because they always send the same message.

Defreeze factor node, variable node or edge

The functions unfreeze the target frozen factor node or frozen variable node, allowing the factor or variable node to calculate outgoing messages.

defreezeFactor!(gbp; factor = index)
defreezeVariable!(gbp; variable = index)

Also, we provide functions that unfreeze the target edge, starting the calculation of messages either from variable node to factor node or from factor node to variable node calculates.

defreezeVariableFactor!(gbp; variable = index, factor = index)
defreezeFactorVariable!(gbp; factor = index, variable = index)

The functions accept following parameters: composite type ContinuousModel; the factor node index corresponding to the row index of the coefficient matrix; and the variable node index corresponding to the column index of the coefficient matrix. Since singly connected factors cannot be frozen, they cannot be unfreezed.

Hide factor node

Utilising a hiding mechanism, the function softly deletes factor node. Hence, the function obliterates the target factor node from the graph during the calculation. Soft delete actually removes the node, while preserving node numbering and keeping the same dimensions of the internal variables.

hideFactor!(gbp; factor = index)

If the function targets the singly connected factor node, the function obliterates the target factor only if there are two or more singly connected factor nodes at the same variable node. If there is only one singly connected factor node at the variable node, the function transforms the target factor node to the virtual factor node. Note that to maintain consistency, the function also affects ContinuousSystem.observation, ContinuousSystem.coefficient and ContinuousSystem.coefficientTranspose fields by setting non-zero elements to zero.

addFactors!(gbp; coefficient = matrix, observation = vector, variance = vector)
The function supports addition of the multiple factor nodes to initial (existing) formation of the factor graph using the same input data format. The function accepts the following parameters: composite type ContinuousModel; the observation and variance vectors representing new observation values and variances, respectively. The keyword coefficient with corresponding coefficients defines the set of equations describing new factor nodes. Also, function initializes messages from variable nodes to a new factor node using results from the last GBP iteration. Note that the function also affects ContinuousSystem.observation, ContinuousSystem.variance, ContinuousSystem.coefficient and ContinuousSystem.coefficientTranspose fields.