/**
* This method is the access point to the planning procedure. Initially, it adds all variables
* from axioms to the set of found vars, then does the linear planning. If lp does not solve the
* problem and there are subtasks, goal-driven recursive planning with backtracking is invoked.
* Planning is performed until no new variables are introduced into the algorithm.
*/
public EvaluationAlgorithm invokePlaning(Problem problem, boolean _computeAll) {
long startTime = System.currentTimeMillis();
computeAll = _computeAll;
EvaluationAlgorithm algorithm = new EvaluationAlgorithm();
PlanningContext context = problem.getCurrentContext();
// add all axioms at the beginning of an algorithm
Collection<Var> flattened = new HashSet<Var>();
for (Iterator<Rel> axiomIter = problem.getAxioms().iterator(); axiomIter.hasNext(); ) {
Rel rel = axiomIter.next();
unfoldVarsToSet(rel.getOutputs(), flattened);
// do not overwrite values of variables that come via args of compute() or as inputs of
// independent subtasks
if (!problem.getAssumptions().containsAll(flattened)
// do not overwrite values of already known variables.
// typically this is the case when a value of a variable
// is given in a scheme via a properties window
// && !problem.getKnownVars().containsAll( flattened )
) {
algorithm.addRel(rel);
}
axiomIter.remove();
context.getKnownVars().addAll(flattened);
flattened.clear();
}
context.getFoundVars().addAll(context.getKnownVars());
// remove all known vars with no relations
for (Iterator<Var> varIter = context.getKnownVars().iterator(); varIter.hasNext(); ) {
if (varIter.next().getRels().isEmpty()) {
varIter.remove();
}
}
// start planning
if (problem.getRelsWithSubtasks().isEmpty()
&& linearForwardSearch(context, algorithm, computeAll)) {
if (isLinearLoggingOn()) logger.debug("Problem solved without subtasks");
} else if (!problem.getRelsWithSubtasks().isEmpty() && subtaskPlanning(problem, algorithm)) {
if (isLinearLoggingOn()) logger.debug("Problem solved with subtasks");
} else if (!computeAll) {
if (isLinearLoggingOn()) logger.debug("Problem not solved");
}
if (!nested) {
logger.info("Planning time: " + (System.currentTimeMillis() - startTime) + "ms.");
}
return algorithm;
}
public static void MLalgo() {
try {
Problem problem = new Problem();
problem.l = train_count; // number of training examples
problem.n = max_feature_count; // number of features
problem.x = train_matrix; // feature nodes
problem.y = ylable; // target values;
SolverType solver = SolverType.L2R_LR; // -s 0
double C = 1.0; // cost of constraints violation
double eps = 0.01; // stopping criteria
Parameter parameter = new Parameter(solver, C, eps);
model = Linear.train(problem, parameter);
File modelFile = new File("model");
model.save(modelFile);
// load model or use it directly
model = Model.load(modelFile);
} catch (Exception e) {
e.printStackTrace();
}
}
void getInfo() {
Scanner console = new Scanner(System.in);
int maxCost = 0, maxSols = 0;
Problem prob = null;
Solver solve = null;
ArrStack stk = new ArrStack();
System.out.print("Enter problem type, solution type, " + "max cost and max # of solutions: ");
try {
prob = (Problem) Class.forName(console.next()).newInstance();
solve = (Solver) Class.forName(console.next()).newInstance();
maxCost = console.nextInt();
maxSols = console.nextInt();
} catch (Exception e) {
System.out.println("" + e);
}
try {
prob.read(console);
} catch (Exception e) {
System.out.println("Read error: " + e);
return;
}
Solver.Solution[] sols;
sols = solve.solveProblem(prob, maxCost, maxSols);
if (sols == null) {
System.out.println("No solutions ");
return;
}
System.out.println("Answers are: ");
for (int ans = 0; ans < sols.length && sols[ans] != null; ans++) {
System.out.println("Answer " + ans + " with cost " + (sols[ans].mSteps.length - 1));
for (int stepNdx = 1; stepNdx < sols[ans].mSteps.length; stepNdx++)
System.out.println(" " + sols[ans].mSteps[stepNdx]);
}
}
private void mainClassifierFunction(int option, String trainFile, String testFile, String ddgFile)
throws IOException {
// SentimentClassifierHindi this = new SentimentClassifierHindi();
// int finalSize = this.SentimentClassifierHindi();
int finalSize = this.generateFeature(option, trainFile, testFile, ddgFile);
System.out.println("Hello aspectCategorizationSemEval2016!");
// Create features
Problem problem = new Problem();
// Save X to problem
double a[] = new double[this.trainingFeature.size()];
File file = new File(rootDirectory + "\\dataset\\trainingLabels.txt");
BufferedReader reader = new BufferedReader(new FileReader(file));
String read;
int count = 0;
while ((read = reader.readLine()) != null) {
// System.out.println(read);
a[count++] = Double.parseDouble(read.toString());
}
// Feature[][] f = new Feature[][]{ {}, {}, {}, {}, {}, {} };
// trainingFeature = trainingObject.getList();
Feature[][] trainFeatureVector = new Feature[trainingFeature.size()][finalSize];
System.out.println("Training Instances: " + trainingFeature.size());
System.out.println("Feature Length: " + finalSize);
System.out.println("Test Instances: " + testFeature.size());
for (int i = 0; i < trainingFeature.size(); i++) {
// System.out.println();
// System.out.println(trainingFeature.get(i));
System.out.println(i + " trained.");
for (int j = 0; j < finalSize; j++) {
// System.out.print(trainingFeature.get(i).get(j + 1)+" ");
// trainingFeature.get(i).
if (trainingFeature.get(i).containsKey(j + 1)) {
// System.out.print(j + 1 + ", ");
trainFeatureVector[i][j] = new FeatureNode(j + 1, trainingFeature.get(i).get(j + 1));
} else {
trainFeatureVector[i][j] = new FeatureNode(j + 1, 0.0);
}
}
// System.out.println();
}
problem.l = trainingFeature.size(); // number of training examples
problem.n = finalSize; // number of features
problem.x = trainFeatureVector; // feature nodes
problem.y = a; // target values ----
BasicParser bp = new BasicParser();
SolverType solver = SolverType.L2R_LR; // -s 7
double C = 0.75; // cost of constraints violation
double eps = 0.0001; // stopping criteria
Parameter parameter = new Parameter(solver, C, eps);
Model model = Linear.train(problem, parameter);
File modelFile = new File("model");
model.save(modelFile);
// PrintWriter write = new PrintWriter(new BufferedWriter(new FileWriter(rootDirectory +
// "\\dataset\\predictedLabels.txt")));
PrintWriter write =
new PrintWriter(
new BufferedWriter(
new FileWriter(
rootDirectory
+ "\\dataset\\dataset_aspectCategorization\\predictedHotelsLabels.txt")));
if (option == 1) {
BufferedReader trainReader =
new BufferedReader(
new FileReader(
new File(
rootDirectory + "\\dataset\\dataset_aspectCategorization\\" + trainFile)));
HashMap<String, Integer> id = new HashMap<String, Integer>();
HashMap<String, String> review = new HashMap<String, String>();
double[] val = new double[trainingFeature.size()];
double[] tempVal = new double[trainingFeature.size()];
LinearCopy.crossValidation(problem, parameter, 5, val, tempVal);
for (int i = 0; i < trainingFeature.size(); i++) {
int flag = 0;
String tokens[] = trainReader.readLine().split("\\|");
if (id.containsKey(tokens[1]) == true || tokens[2].compareToIgnoreCase("True") == 0) {
} else {
// System.out.println(tokens[1]);
/*int max = -1;
double probMax = -1.0;
for(int j=0; j<13; j++){
if(probMax<val[i][j]){
probMax = val[i][j];
max = j;
}
}*/
// System.out.println(tempVal[i]);
write.println((int) (val[i]));
write.println("next");
id.put(tokens[1], 1);
System.out.println(tokens[1] + "\t" + (int) (val[i]));
if (review.containsKey(tokens[1])) {
System.out.println(tokens[3]);
System.out.println(review.get(tokens[1]));
} else {
review.put(tokens[1], tokens[3]);
}
} /*else{
for (int j = 0; j < 13; j++) {
//System.out.print(val[i][j]+", ");
if (val[i] >= 0.185) {
flag = 1;
//System.out.println("i");
write.println(j + 1);
}
}
if (flag == 1) {
write.println("next");
} else {
write.println("-1");
write.println("next");
}
//write.println(prediction);
id.put(tokens[1], 1);
//System.out.println();
}*/
}
write.close();
return;
}
if (option == 3) {
System.out.println(rootDirectory);
BufferedReader testReader =
new BufferedReader(
new FileReader(
new File(
rootDirectory + "\\dataset\\dataset_aspectCategorization\\" + testFile)));
HashMap<String, Integer> id = new HashMap<String, Integer>();
model = Model.load(modelFile);
int countNext = 0;
for (int i = 0; i < testFeature.size(); i++) {
// System.out.println(i+", "+testFeature.size()+", "+testFeature.get(i).size());
Feature[] instance = new Feature[testFeature.get(i).size()];
int j = 0;
for (Map.Entry<Integer, Double> entry : testFeature.get(i).entrySet()) {
// System.out.print(entry.getKey() + ": " + entry.getValue() + "; ");
// listOfMaps.get(i).put(start + entry.getKey(), entry.getValue());
// do stuff
instance[j++] = new FeatureNode(entry.getKey(), entry.getValue());
}
// double d = LinearCopy.predict(model, instance);
double[] predict = new double[85];
double prediction = LinearCopy.predictProbability(model, instance, predict);
int labelMap[] = new int[13];
labelMap = model.getLabels();
for (int ar = 0; ar < labelMap.length; ar++) {
System.out.println("********************** " + ar + ": " + labelMap[ar]);
}
// System.out.println(prediction);
// Arrays.sort(predict, Collections.reverseOrder());
// System.out.println();
// double prediction = LinearCopy.predict(model, instance);
String tokens[] = testReader.readLine().split("\\|");
// System.out.println(tokens[1]);
int flag = -1;
if (id.containsKey(tokens[1]) == true || tokens[2].compareToIgnoreCase("True") == 0) {
flag = 4;
// System.out.println("OutofScope: "+tokens[1]);
} else if (tokens[3].compareToIgnoreCase("abc") == 0) {
flag = 2;
System.out.println(tokens[1]);
write.println("-1");
write.println("next");
countNext++;
id.put(tokens[1], 1);
} else {
flag = 0;
for (int p = 0; p < 85; p++) {
if (predict[p] >= 0.128) {
flag = 1;
write.println(labelMap[p]);
}
}
if (flag == 1) {
countNext++;
write.println("next");
} else {
countNext++;
write.println("-1");
write.println("next");
}
// write.println((int)d);
// write.println("next");
/*write.println(prediction);
write.println("next");*/
id.put(tokens[1], 1);
}
if (flag == -1) {
System.out.println("-1, " + tokens[1]);
}
}
write.close();
System.out.println("count " + countNext);
}
write.close();
}
/**
* Goal-driven recursive (depth-first, exhaustive) search with backtracking
*
* @param problem
* @param algorithm
* @param subtaskRelsInPath
* @param depth
*/
private boolean subtaskPlanningImpl(
PlanningContext context,
Set<Rel> relsWithSubtasks,
EvaluationAlgorithm algorithm,
LinkedList<Rel> subtaskRelsInPath,
int depth) {
Set<Rel> relsWithSubtasksCopy = new LinkedHashSet<Rel>(relsWithSubtasks);
Set<Rel> relsWithSubtasksToRemove = new LinkedHashSet<Rel>();
boolean firstMLB = true;
// start building Maximal Linear Branch (MLB)
MLB:
while (!relsWithSubtasksCopy.isEmpty()) {
if (isSubtaskLoggingOn()) {
String print = p(depth) + "Starting new MLB with: ";
for (Rel rel : relsWithSubtasksCopy) {
print +=
"\n" + p(depth) + " " + rel.getParent().getFullName() + " : " + rel.getDeclaration();
}
/*
print += "\n" + p( depth ) + " All remaining rels in problem:";
for ( Rel rel : problem.getAllRels() ) {
print += "\n" + p( depth ) + " " + rel.getParentObjectName() + " : " + rel.getDeclaration();
}
print += "\n" + p( depth ) + "All found variables: ";
for ( Var var : problem.getFoundVars() ) {
print += "\n" + p( depth ) + " " + var.toString();
}
*/
logger.debug(print);
}
// if this is a first attempt to construct an MLB to solve a subtask(i.e. depth>0),
// do not invoke linear planning because it has already been done
if ((depth == 0) || !firstMLB) {
boolean solvedIntermediately = linearForwardSearch(context, algorithm, true);
// Having constructed some MLBs the (sub)problem may be solved
// and there is no need in wasting precious time planning unnecessary branches
if (solvedIntermediately
&& ( // on the top level optimize only if computing goals
(depth == 0 && !computeAll)
// otherwise (inside subtasks) always optimize
|| (depth != 0))) {
// If the problem is solved, optimize and return
if (!isOptDisabled) Optimizer.optimize(context, algorithm);
return true;
}
} else {
firstMLB = false;
}
// or children
OR:
for (Iterator<Rel> subtaskRelIterator = relsWithSubtasksCopy.iterator();
subtaskRelIterator.hasNext(); ) {
Rel subtaskRel = subtaskRelIterator.next();
if (isSubtaskLoggingOn())
logger.debug(
p(depth)
+ "OR: depth: "
+ (depth + 1)
+ " rel - "
+ subtaskRel.getParent().getFullName()
+ " : "
+ subtaskRel.getDeclaration());
if (subtaskRel.equals(subtaskRelsInPath.peekLast())
|| (!context.isRelReadyToUse(subtaskRel))
|| context.getFoundVars().containsAll(subtaskRel.getOutputs())
|| (!isSubtaskRepetitionAllowed && subtaskRelsInPath.contains(subtaskRel))) {
if (isSubtaskLoggingOn()) {
logger.debug(p(depth) + "skipped");
if (!context.isRelReadyToUse(subtaskRel)) {
logger.debug(p(depth) + "because it has unknown inputs"); // TODO print unknown
} else if (context.getFoundVars().containsAll(subtaskRel.getOutputs())) {
logger.debug(p(depth) + "because all outputs in FoundVars");
} else if (subtaskRel.equals(subtaskRelsInPath.peekLast())) {
logger.debug(p(depth) + "because it is nested in itself");
} else if (!isSubtaskRepetitionAllowed && subtaskRelsInPath.contains(subtaskRel)) {
logger.debug(
p(depth)
+ "This rel with subtasks is already in use, path: "
+ subtaskRelsInPath);
}
}
continue OR;
}
LinkedList<Rel> newPath = new LinkedList<Rel>(subtaskRelsInPath);
newPath.add(subtaskRel);
PlanningResult result = new PlanningResult(subtaskRel, true);
// this is true if all subtasks are solvable
boolean allSolved = true;
// and children
AND:
for (SubtaskRel subtask : subtaskRel.getSubtasks()) {
if (isSubtaskLoggingOn()) logger.debug(p(depth) + "AND: subtask - " + subtask);
EvaluationAlgorithm sbtAlgorithm = null;
////////////////////// INDEPENDENT SUBTASK////////////////////////////////////////
if (subtask.isIndependent()) {
if (isSubtaskLoggingOn()) logger.debug("Independent!!!");
if (subtask.isSolvable() == null) {
if (isSubtaskLoggingOn())
logger.debug("Start solving independent subtask " + subtask.getDeclaration());
// independent subtask is solved only once
Problem problemContext = subtask.getContext();
DepthFirstPlanner planner = new DepthFirstPlanner();
planner.indSubtasks = indSubtasks;
planner.nested = true;
sbtAlgorithm = planner.invokePlaning(problemContext, isOptDisabled);
PlanningContext indCntx = problemContext.getCurrentContext();
boolean solved = indCntx.getFoundVars().containsAll(indCntx.getAllGoals());
if (solved) {
subtask.setSolvable(Boolean.TRUE);
indSubtasks.put(subtask, sbtAlgorithm);
if (isSubtaskLoggingOn()) logger.debug("Solved " + subtask.getDeclaration());
} else {
subtask.setSolvable(Boolean.FALSE);
if (RuntimeProperties.isLogInfoEnabled()) {
logger.debug("Unable to solve " + subtask.getDeclaration());
}
}
allSolved &= solved;
} else if (subtask.isSolvable() == Boolean.TRUE) {
if (isSubtaskLoggingOn()) logger.debug("Already solved");
allSolved &= true;
sbtAlgorithm = indSubtasks.get(subtask);
} else {
if (isSubtaskLoggingOn()) logger.debug("Not solvable");
allSolved &= false;
}
if (isSubtaskLoggingOn()) logger.debug("End of independent subtask " + subtask);
if (!allSolved) {
continue OR;
}
assert sbtAlgorithm != null;
result.addSubtaskAlgorithm(subtask, sbtAlgorithm);
}
////////////////////// DEPENDENT SUBTASK//////////////////////////////////////
else {
// lets clone the environment
PlanningContext newContext = prepareNewContext(context, subtask);
sbtAlgorithm = new EvaluationAlgorithm();
// during linear planning, if some goals are found, they are removed from the set
// "goals"
boolean solved =
linearForwardSearch(
newContext,
sbtAlgorithm,
// do not optimize here, because the solution may require additional rels with
// subtasks
true);
if (solved) {
if (isSubtaskLoggingOn()) logger.debug(p(depth) + "SOLVED subtask: " + subtask);
if (!isOptDisabled) {
// if a subtask has been solved, optimize its algorithm
Optimizer.optimize(newContext, sbtAlgorithm);
}
result.addSubtaskAlgorithm(subtask, sbtAlgorithm);
allSolved &= solved;
continue AND;
} else if (!solved && (depth == maxDepth)) {
if (isSubtaskLoggingOn())
logger.debug(p(depth) + "NOT SOLVED and cannot go any deeper, subtask: " + subtask);
continue OR;
}
if (isSubtaskLoggingOn()) logger.debug(p(depth) + "Recursing deeper");
solved =
subtaskPlanningImpl(newContext, relsWithSubtasks, sbtAlgorithm, newPath, depth + 1);
if (isSubtaskLoggingOn()) logger.debug(p(depth) + "Back to depth " + (depth + 1));
// the linear planning has been performed at the end of MLB on the depth+1,
// if the problem was solved, there is no need to run linear planning again
if ((solved || (solved = linearForwardSearch(newContext, sbtAlgorithm, true)))
&& !isOptDisabled) {
// if solved, optimize here with full list of goals in order to get rid of
// unnecessary subtask instances and other relations
Optimizer.optimize(newContext, sbtAlgorithm);
}
if (isSubtaskLoggingOn())
logger.debug(p(depth) + (solved ? "" : "NOT") + " SOLVED subtask: " + subtask);
allSolved &= solved;
// if at least one subtask is not solvable, try another
// branch
if (!allSolved) {
continue OR;
}
result.addSubtaskAlgorithm(subtask, sbtAlgorithm);
}
} // AND
if (allSolved) {
algorithm.add(result);
Set<Var> newVars = new LinkedHashSet<Var>();
unfoldVarsToSet(subtaskRel.getOutputs(), newVars);
context.getKnownVars().addAll(newVars);
context.getFoundVars().addAll(newVars);
subtaskRelIterator.remove();
if (isSubtaskLoggingOn()) {
logger.debug(
p(depth)
+ "SOLVED ALL SUBTASKS for "
+ subtaskRel.getParent().getFullName()
+ " : "
+ subtaskRel.getDeclaration());
logger.debug(p(depth) + "Updating the problem graph and continuing building new MLB");
}
// this is used for incremental dfs
if (depth == 0) {
relsWithSubtasksToRemove.add(subtaskRel);
}
continue MLB;
}
if (isSubtaskLoggingOn())
logger.debug(
p(depth)
+ "NOT SOLVED ALL subtasks, removing from path "
+ subtaskRel.getParent().getFullName()
+ " : "
+ subtaskRel.getDeclaration());
newPath.remove(subtaskRel);
} // end OR
// exit loop because there are no more rels with subtasks to be
// applied
// (i.e. no more rels can introduce new variables into the
// algorithm)
if (isSubtaskLoggingOn()) logger.debug(p(depth) + "No more MLB can be constructed");
break MLB;
}
// incremental dfs, remove solved subtasks
if (depth == 0) {
relsWithSubtasks.removeAll(relsWithSubtasksToRemove);
}
return false;
}
private boolean subtaskPlanning(Problem problem, EvaluationAlgorithm algorithm) {
if (isSubtaskLoggingOn())
logger.debug("!!!--------- Starting Planning With Subtasks ---------!!!");
final int maxDepthBackup = maxDepth;
if (isSubtaskLoggingOn())
logger.debug(
"maxDepthBackup:" + maxDepthBackup + " sbt: " + problem.getRelsWithSubtasks().size());
PlanningContext context = problem.getCurrentContext();
try {
Set<Rel> relsWithSubtasks = new LinkedHashSet<Rel>(problem.getRelsWithSubtasks());
if (isIncremental) {
int incrementalDepth = 0;
while (incrementalDepth
<= (isSubtaskRepetitionAllowed
? maxDepthBackup
: problem.getRelsWithSubtasks().size() - 1)) {
if (isSubtaskLoggingOn())
logger.debug(
"Incremental dfs, with max depth "
+ (incrementalDepth + 1)
+ " and "
+ problem.getRelsWithSubtasks().size()
+ " subtasks to solve");
maxDepth = incrementalDepth++;
// if we need to compute some specific goals, after reaching a certain depth, but not the
// maximal depth,
// the problem may be solved and there is no need to go any deeper.
if (subtaskPlanningImpl(context, relsWithSubtasks, algorithm, new LinkedList<Rel>(), 0)) {
if (isSubtaskLoggingOn())
logger.debug("The problem was solved during idfs after some intermediate MLB");
return true;
}
if (isSubtaskLoggingOn())
logger.debug("Unsolved subtask left: " + problem.getRelsWithSubtasks().size());
}
if (isSubtaskLoggingOn()) logger.debug("Fininshed incremental dfs");
} else {
if (!isSubtaskRepetitionAllowed) {
maxDepth = problem.getRelsWithSubtasks().size() - 1;
}
if (isSubtaskLoggingOn())
logger.debug("Starting subtask dfs with maxDepth: " + (maxDepth + 1));
if (subtaskPlanningImpl(context, relsWithSubtasks, algorithm, new LinkedList<Rel>(), 0)) {
if (isSubtaskLoggingOn())
logger.debug("The problem was solved during dfs after some intermediate MLB");
return true;
}
}
} finally {
if (isSubtaskLoggingOn()) logger.debug("Fininshed dfs");
maxDepth = maxDepthBackup;
indSubtasks.clear();
}
if (isSubtaskLoggingOn()) logger.debug("Invoking final linear planning");
return linearForwardSearch(context, algorithm, computeAll);
}
