/**
* 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;
}
private PlanningContext prepareNewContext(PlanningContext context, Rel subtask) {
PlanningContext newContext = context.getCopy();
// [x->y] assume x is known
Set<Var> allSubtaskInputs = new LinkedHashSet<Var>();
unfoldVarsToSet(subtask.getInputs(), allSubtaskInputs);
newContext.getKnownVars().addAll(allSubtaskInputs);
newContext.getFoundVars().addAll(allSubtaskInputs);
Set<Var> goals = new LinkedHashSet<Var>();
unfoldVarsToSet(subtask.getOutputs(), goals);
newContext.addGoals(goals);
return newContext;
}
/**
* 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;
}
/**
* Linear forward search algorithm
*
* @param p
* @param algorithm
* @param targetVars
* @param _computeAll
* @return
*/
private boolean linearForwardSearch(
PlanningContext context, EvaluationAlgorithm algorithm, boolean _computeAll) {
/*
* while iterating through hashset, items cant be removed from/added to
* that set. Theyre collected into these sets and added/removedall
* together after iteration is finished
*/
Set<Var> newVars = new LinkedHashSet<Var>();
Set<Var> relOutputs = new LinkedHashSet<Var>();
Set<Var> removableVars = new LinkedHashSet<Var>();
boolean changed = true;
if (isLinearLoggingOn())
logger.debug(
"------Starting linear planning with (sub)goals: "
+ context.getRemainingGoals()
+ "--------");
if (isLinearLoggingOn()) logger.debug("Algorithm " + algorithm);
int counter = 1;
while ((!_computeAll && changed && !context.getRemainingGoals().isEmpty())
|| (changed && _computeAll)) {
if (isLinearLoggingOn()) logger.debug("----Iteration " + counter + " ----");
counter++;
changed = false;
// iterate through all knownvars
if (isLinearLoggingOn()) logger.debug("Known:" + context.getKnownVars());
for (Var var : context.getKnownVars()) {
if (isLinearLoggingOn()) logger.debug("Current Known: " + var);
// Check the relations of all components
for (Rel rel : var.getRels()) {
if (isLinearLoggingOn()) logger.debug("And its rel: " + rel);
if (context.isAvailableRel(rel)) {
context.removeUnknownInput(rel, var);
if (isLinearLoggingOn()) logger.debug("problem contains it " + rel);
removableVars.add(var);
if (context.isRelReadyToUse(rel) && rel.getType() != RelType.TYPE_METHOD_WITH_SUBTASK) {
if (isLinearLoggingOn()) logger.debug("rel is ready to be used " + rel);
boolean relIsNeeded = false;
if (isLinearLoggingOn()) logger.debug("its outputs " + rel.getOutputs());
for (Var relVar : rel.getOutputs()) {
if (!context.getFoundVars().contains(relVar)) {
relIsNeeded = true;
}
}
if (rel.getOutputs().isEmpty()) {
relIsNeeded = true;
}
if (isLinearLoggingOn()) logger.debug("relIsNeeded " + relIsNeeded);
if (relIsNeeded) {
if (isLinearLoggingOn()) logger.debug("needed rel: " + rel);
if (!rel.getOutputs().isEmpty()) {
relOutputs.clear();
unfoldVarsToSet(rel.getOutputs(), relOutputs);
newVars.addAll(relOutputs);
context.getFoundVars().addAll(relOutputs);
}
algorithm.addRel(rel);
if (isLinearLoggingOn()) logger.debug("algorithm " + algorithm);
}
context.removeRel(rel);
changed = true;
}
}
}
}
// remove targets if they have already been found
for (Iterator<Var> targetIter = context.getRemainingGoals().iterator();
targetIter.hasNext(); ) {
Var targetVar = targetIter.next();
if (context.getFoundVars().contains(targetVar)) {
targetIter.remove();
}
}
if (isLinearLoggingOn()) logger.debug("foundvars " + context.getFoundVars());
context.getKnownVars().addAll(newVars);
context.getKnownVars().removeAll(removableVars);
newVars.clear();
}
if (isLinearLoggingOn()) logger.debug("algorithm " + algorithm);
if (!_computeAll) {
Optimizer.optimize(context, algorithm);
if (isLinearLoggingOn()) logger.debug("optimized algorithm " + algorithm);
}
if (isLinearLoggingOn()) logger.debug("\n---!!!Finished linear planning!!!---\n");
return context.getRemainingGoals().isEmpty()
|| context.getFoundVars().containsAll(context.getAllGoals());
}