/**
* Extracts the parameters for the subModels from the parameters vector returned by
* DTOM.makeParameters(...) Basically the same as BNet.makeSubParamList(...), but for input as
* follows - [ ( [intraslice parents], [temporal parents], (submodel,subparams) ) ]
*/
protected static Value[] makeSubParamListDBN(Value.Vector params) {
Value[] value = new Value[params.length()];
for (int i = 0; i < value.length; i++) {
Value.Structured temp = (Value.Structured) params.elt(i);
// value[i] = ((Value.Structured)(temp.cmpnt(2))).cmpnt(1);
value[i] =
((Value.Structured) (temp.cmpnt(3)))
.cmpnt(1); // One extra component in input vs TOM params
}
return value;
}
/**
* Extracts the subModels from the parameters vector returned by DTOM.makeParameters(...).
* Basically the same as BNet.makeSubModelList(...), but for input as follows - [ ( [intraslice
* parents], [temporal parents], (submodel,subparams) ) ]
*/
protected static Value.Model[] makeSubModelListDBN(Value.Vector params) {
Value.Model[] model = new Value.Model[params.length()];
for (int i = 0; i < model.length; i++) {
Value.Structured temp = (Value.Structured) params.elt(i);
// model[i] = (Value.Model)((Value.Structured)(temp.cmpnt(2))).cmpnt(0);
model[i] =
(Value.Model)
((Value.Structured) (temp.cmpnt(3)))
.cmpnt(0); // One extra component in input vs TOM params
}
return model;
}
/** Do a quick traversal of the dTree to find the number of leaves on the tree. */
public int findNumLeaves(Value.Structured params) {
int splitAttribute = params.intCmpnt(0);
int numLeaves = 0;
// we have hit a leaf.
if (splitAttribute == -1) {
numLeaves = 1;
} else {
Value.Vector paramVector = (Value.Vector) params.cmpnt(2);
for (int i = 0; i < paramVector.length(); i++) {
numLeaves += findNumLeaves((Value.Structured) paramVector.elt(i));
}
}
return numLeaves;
}
/** Do a quick traversal of the dTree to find the maximum depth reached. */
public int findTreeDepth(Value.Structured params) {
int splitAttribute = params.intCmpnt(0);
int maxDepth = 0;
// we have hit a leaf.
if (splitAttribute == -1) {
return 0;
} else {
Value.Vector paramVector = (Value.Vector) params.cmpnt(2);
for (int i = 0; i < paramVector.length(); i++) {
int newDepth = findTreeDepth((Value.Structured) paramVector.elt(i));
if (newDepth > maxDepth) {
maxDepth = newDepth;
}
}
}
return maxDepth + 1;
}
/** Set the current parameters to a given value. */
public void updateParams(Value newParams) {
if (this.params == newParams) return;
this.params = newParams;
Value.Structured params = (Value.Structured) this.params; // save us doing repeated typecasts.
int numLeaves = findNumLeaves(params);
int depth = findTreeDepth(params) + 1;
setLayout(new GridLayout(numLeaves, depth));
int currentDepth = 0; // Keep track of how deep into the tree we are
int currentLeaf = 0; // Keep track of how many leaf noded have been placed so far.
Component[][] component = new Component[depth][numLeaves];
// Use a stack to keep track of parameters not yet drawn (could also be done recursively.)
java.util.Stack paramStack = new java.util.Stack();
paramStack.add(params);
while (true) {
Value.Structured currentParams = (Value.Structured) paramStack.pop();
int splitAttribute = currentParams.intCmpnt(0);
if (splitAttribute == -1) {
Value.Structured subModelParams = (Value.Structured) currentParams.cmpnt(2);
Value.Model subModel = (Value.Model) subModelParams.cmpnt(0);
Value subParams = subModelParams.cmpnt(1);
Component leaf = DualVisualiser.makeComponent(varName, parentName, subModel, subParams);
component[currentDepth][currentLeaf] = leaf;
currentLeaf++;
if (currentLeaf < numLeaves) {
while (component[currentDepth - 1][currentLeaf] == null) {
currentDepth--;
}
}
} else {
Value.Vector paramVector = (Value.Vector) currentParams.cmpnt(2);
for (int i = 0; i < paramVector.length(); i++) {
Value.Structured elt = (Value.Structured) paramVector.elt(paramVector.length() - i - 1);
paramStack.push(elt);
}
int x = currentLeaf;
for (int value = 0; value < paramVector.length(); value++) {
Value.Structured elt = (Value.Structured) paramVector.elt(value);
int subLeaves = findNumLeaves(elt);
Color colour = getColour(value);
for (int j = 0; j < subLeaves; j++) {
if (component[currentDepth][x] != null) {
throw new RuntimeException(
"SHouldn't be overwriting! [" + currentDepth + "," + x + "]");
}
if (j == 0)
component[currentDepth][x] =
new JLabel(currentParams.cmpnt(0).toString() + " = " + value);
else component[currentDepth][x] = new JLabel("");
component[currentDepth][x].setBackground(colour);
((JComponent) component[currentDepth][x]).setOpaque(true);
x++;
}
}
currentDepth++;
}
if (currentLeaf == numLeaves) break;
}
for (int i = 0; i < numLeaves; i++) {
for (int j = 0; j < depth; j++) {
// if ( component[j][i] == null ) component[j][i] = new JLabel("("+j+","+i+")");
if (component[j][i] == null) component[j][i] = new JLabel("");
this.add(component[j][i]);
}
}
}
/**
* Returns a time series (a vector) sampled from the DBN (passed as a DTOM) Based on
* BNetStochastic.generate(...). Used (perhaps amongst other things) during the calculation of
* (approximate) KL divergence between models
*
* @param rand RNG to use
* @param dtom DTOM, parameters of which to generate data from (must have caseInfo set in DTOM)
* @param n length of time series to generate from parameters
* @return Time series sampled/generated from the DTOM
*/
public static Value.Vector generateTimeSeriesDTOM(Random rand, DTOM dtom, int n) {
// Get the variable names and number of nodes; also get the ordering of the nodes (used for
// forward sampling)
int numNodes = dtom.getNumNodes();
String[] name =
((cdms.core.Type.Structured) ((cdms.core.Type.Vector) (dtom.caseInfo.data).t).elt).labels;
int[] order = dtom.getTotalOrderCopy();
// Initialize the vectors and structures etc for storing the time series data: (Value.Vector)
Value.Vector origData = dtom.caseInfo.data;
Type.Structured inputTypes =
(Type.Structured)
((Type.Vector) origData.t).elt; // Type info for generated data is same as original data
// Store the assignments in an integer array:
int[][] newData = new int[numNodes][n];
// Get the parameters for each node (parameters for first time slice, and second time slice)
Value.Vector[] paramsT0 = new Value.Vector[numNodes];
Value.Vector[] paramsT1 = new Value.Vector[numNodes];
for (int i = 0; i < numNodes; i++) {
DNode node = (DNode) dtom.getNode(i);
try {
// Learn parameters - T1 (inc. temporal arcs)
Value.Structured model = node.learnModel(dtom.caseInfo.mmlModelLearner, dtom.caseInfo.data);
paramsT1[i] = (Value.Vector) model.cmpnt(2);
// Learn parameters - T0 (no temporal arcs)
Value.Structured modelT0 =
node.learnModelT0(dtom.caseInfo.mmlModelLearner, dtom.caseInfo.data);
paramsT0[i] = (Value.Vector) modelT0.cmpnt(2);
} catch (LearnerException e) {
throw new RuntimeException("Error learning models. " + e);
}
}
// Determine the arity of each node: (Using a very inelegant method...)
final Type.Structured datatype = (Type.Structured) ((Type.Vector) (dtom.caseInfo.data).t).elt;
int[] arity = new int[numNodes];
for (int i = 0; i < numNodes; i++) {
Type.Symbolic sType = (Type.Symbolic) datatype.cmpnts[i];
arity[i] = NeticaFn.makeValidNeticaNames(sType.ids, true).length;
}
// Generate a set of assignments for the FIRST time slice
// (This needs to be done in order, to avoid sampling children before parents...)
int[] assignmentT0 = new int[numNodes];
for (int i = 0; i < numNodes; i++) {
DNode currNode = (DNode) dtom.getNode(order[i]); // ith node in total order
Value.Vector currParams =
paramsT0[order[i]].cmpnt(1); // parameters for ith node in total order
// Get the appropriate distribution to sample from (given values of parents)
Structured vals;
if (currNode.getNumParents() == 0) { // No parents
vals =
(Value.Structured)
currParams.elt(
0); // Contains the actual probability values; only one element in structure if
// no parents...
} else { // This node has parents (which already have assigned values)
// Need to work out the index of the relevent parameters given the assignments of parents
// Parameters are in order of [0,0,0], [0,0,1], [0,0,2], ..., [A,B,C]
// Index given by: sum_x( val[pa[x]]*prod( arity[pa[x+1...end]] )
int[] currParents = currNode.getParentCopy(); // Intraslice parents
// Collect assignments and arity for the current parents
int[] assignment = new int[currParents.length];
int[] ar = new int[currParents.length];
for (int z = 0; z < currParents.length; z++) {
assignment[z] = assignmentT0[currParents[z]];
ar[z] = arity[currParents[z]];
}
int index = assignmentToIndexReverse(assignment, ar);
// Find the set of parameters for the current parent assignment:
vals =
(Value.Structured)
currParams.elt(
index); // Contains the actual probability values for the current assignment of
// parents
}
// Now, sample a value according to the probability distribution:
double rv = rand.nextDouble(); // Random value b/w 0 and 1
double cumProb = 0.0;
for (int idx = 0; idx < arity[order[i]]; idx++) { // i.e. loop through each value
cumProb += vals.doubleCmpnt(idx);
if (rv < cumProb) { // Assignment to node[ order[i] ] is idx
assignmentT0[order[i]] = idx;
break;
}
}
}
// Generate data from SECOND time slice CPDs - repeatedly...
int[] assignmentT1 = new int[numNodes];
for (int lineNum = 0; lineNum < n; lineNum++) {
// First: record the first time slice assignemnts.
// Then: copy the second time slice assignments to the first time slice assignments
if (lineNum > 0) {
// System.out.println("Assignments line " + (lineNum-1) + " - " +
// Arrays.toString(assignmentT0) );
for (int j = 0; j < numNodes; j++) { // j is variable number
newData[j][lineNum - 1] = assignmentT0[j];
}
assignmentT0 = assignmentT1;
assignmentT1 = new int[numNodes];
}
// Now, generate data for second time slice given values of first time slice:
for (int i = 0; i < numNodes; i++) {
DNode currNode = (DNode) dtom.getNode(order[i]); // ith node in total order
Value.Vector currParams =
paramsT1[order[i]].cmpnt(1); // parameters for ith node in total order
// Get the appropriate distribution to sample from (given values of parents)
Structured vals;
if (currNode.getNumParents() == 0) { // No parents
vals =
(Value.Structured)
currParams.elt(
0); // Contains the actual probability values; only one element in structure
// if no parents...
} else { // This node has parents (which already have assigned values)
// Need to work out the index of the relevent parameters given the assignments of parents
// Parameters are in order of [0,0,0], [0,0,1], [0,0,2], ..., [A,B,C]
// Index given by: sum_x( val[pa[x]]*prod( arity[pa[x+1...end]] )
// Complicated by the fact that we have temporal parents and intraslice parents...
// [intraslice_parents, temporal_parents]
int[] currParents = currNode.getParentCopy(); // Intraslice parents
int[] currParentsTemporal =
currNode.getTemporalParentCopy(); // Temporal (interslice) parents
// Collect the parent assignments and arities
int numParents = currParents.length + currParentsTemporal.length;
int[] assignment = new int[numParents];
int[] ar = new int[numParents];
for (int z = 0; z < numParents; z++) {
if (z < currParents.length) { // Dealing with intraslice parents
assignment[z] = assignmentT1[currParents[z]];
ar[z] = arity[currParents[z]];
} else { // Dealing with interslice (t0) parents
assignment[z] = assignmentT0[currParentsTemporal[z - currParents.length]];
ar[z] = arity[currParentsTemporal[z - currParents.length]];
}
}
int index = assignmentToIndexReverse(assignment, ar);
// Find the set of parameters for the current parent assignment:
vals =
(Value.Structured)
currParams.elt(
index); // Contains the actual probability values for the current assignment
// of parents
}
// Now, sample a value according to the probability distribution:
double rv = rand.nextDouble(); // Random value b/w 0 and 1
double cumProb = 0.0;
for (int idx = 0; idx < arity[order[i]]; idx++) { // i.e. loop through each value
cumProb += vals.doubleCmpnt(idx);
if (rv < cumProb) { // Assignment to node[ order[i] ] is idx
assignmentT1[order[i]] = idx;
break;
}
}
}
}
// Copy the very last line of data:
for (int j = 0; j < numNodes; j++) {
newData[j][n - 1] = assignmentT1[j];
}
// Now, combine type and value (i.e. assignments) together for each variable:
Value.Vector[] vecArray = new Value.Vector[numNodes];
for (int i = 0; i < numNodes; i++) {
vecArray[i] =
new VectorFN.FastDiscreteVector(newData[i], (Type.Symbolic) inputTypes.cmpnts[i]);
}
// And create the overall data structure:
Value.Structured vecStruct = new Value.DefStructured(vecArray, name);
Value.Vector newDataVector = new VectorFN.MultiCol(vecStruct);
// Return the new time series vector...
return newDataVector;
}