public void runSupervised() {
MyTimer timer = new MyTimer();
timer.start();
PuddleMapFV agentfv = new PuddleMapFV(this.puddleMap, 5, 20, 20);
PuddleMapFVComponent agentCompFV = new PuddleMapFVComponent(this.puddleMap, 5, 20, 20);
StateToFeatureVectorGenerator svar =
new ConcatenatedObjectFeatureVectorGenerator(false, GridWorldDomain.CLASSAGENT);
java.util.List<EpisodeAnalysis> eas =
EpisodeAnalysis.parseFilesIntoEAList(this.expertDir, domain, this.sp);
LinearStateDifferentiableRF objectiveRF =
new LinearStateDifferentiableRF(agentfv, agentfv.getDim());
objectiveRF.setParameters(new double[] {1., -10, -10, 0, -10, 0, 0, 0, 0, 0});
WekaPolicy p = new WekaPolicy(agentfv, new J48(), this.domain.getActions(), eas);
// WekaPolicy p = new WekaPolicy(svar, new Logistic(), this.domain.getActions(), eas);
timer.stop();
System.out.println("Training Time: " + timer.getTime());
String baseName = "Svar";
GridWorldTerminalFunction tf = new GridWorldTerminalFunction(20, 20);
State simple = this.initialState.copy();
GridWorldDomain.setAgent(simple, 18, 0);
EpisodeAnalysis trainedEp1 = p.evaluateBehavior(simple, objectiveRF, tf, 200);
trainedEp1.writeToFile(trainedDir + "/j48" + baseName + "EpSimple", this.sp);
State hardAgent = this.initialState.copy();
GridWorldDomain.setAgent(hardAgent, 0, 9);
EpisodeAnalysis trainedEp2 = p.evaluateBehavior(hardAgent, objectiveRF, tf, 200);
trainedEp2.writeToFile(trainedDir + "/j48" + baseName + "EpHardAgent", this.sp);
int ngx = 12;
int ngy = 14;
tf = new GridWorldTerminalFunction(ngx, ngy);
this.puddleMap[ngx][ngy] = 1;
this.puddleMap[20][20] = 0;
agentfv.setGoal(ngx, ngy);
agentCompFV.setGoal(ngx, ngy);
State hardGoal = this.initialState.copy();
GridWorldDomain.setAgent(hardGoal, 0, 0);
EpisodeAnalysis trainedEp3 = p.evaluateBehavior(hardGoal, objectiveRF, tf, 200);
trainedEp3.writeToFile(trainedDir + "/j48" + baseName + "EpHardGoal", this.sp);
new EpisodeSequenceVisualizer(this.v, this.domain, this.sp, this.trainedDir);
}
public void runIRL() {
MyTimer timer = new MyTimer();
timer.start();
PuddleMapFV ofvgen = new PuddleMapFV(this.puddleMap, 5, 20, 20);
// PuddleMapFV fvgen = new PuddleMapFV(this.puddleMap, 5, 20, 20);
PuddleMapFVComponent fvgen = new PuddleMapFVComponent(this.puddleMap, 5, 20, 20);
// StateToFeatureVectorGenerator fvgen = new ConcatenatedObjectFeatureVectorGenerator(false,
// GridWorldDomain.CLASSAGENT);
// PuddleMapExactFV fvgen = new PuddleMapExactFV(this.puddleMap, 5);
// LinearStateDifferentiableRF rf = new LinearStateDifferentiableRF(fvgen, fvgen.getDim());
LinearStateActionDifferentiableRF rf =
new LinearStateActionDifferentiableRF(
fvgen,
fvgen.getDim(),
new GroundedAction(this.domain.getAction(GridWorldDomain.ACTIONNORTH), ""),
new GroundedAction(this.domain.getAction(GridWorldDomain.ACTIONSOUTH), ""),
new GroundedAction(this.domain.getAction(GridWorldDomain.ACTIONEAST), ""),
new GroundedAction(this.domain.getAction(GridWorldDomain.ACTIONWEST), ""));
LinearStateDifferentiableRF objectiveRF =
new LinearStateDifferentiableRF(ofvgen, ofvgen.getDim());
objectiveRF.setParameters(new double[] {1., -10, -10, 0, -10, 0, 0, 0, 0, 0});
// objectiveRF.setParameters(new double[]{1., -10, -10, 0, -10});
java.util.List<EpisodeAnalysis> eas =
EpisodeAnalysis.parseFilesIntoEAList(this.expertDir, domain, this.sp);
int depth = 6;
double beta = 10;
// DifferentiableSparseSampling dss = new DifferentiableSparseSampling(domain, rf, new
// NullTermination(), 0.99, new NameDependentStateHashFactory(), depth, -1, beta);
DifferentiableZeroStepPlanner dss = new DifferentiableZeroStepPlanner(domain, rf);
dss.toggleDebugPrinting(false);
MLIRLRequest request = new MLIRLRequest(domain, dss, eas, rf);
request.setBoltzmannBeta(beta);
// MLIRL irl = new MLIRL(request, 0.001, 0.01, 10); //use this for only the given features
// MLIRL irl = new MLIRL(request, 0.00001, 0.01, 10);
MLIRL irl = new MLIRL(request, 0.0001, 0.01, 10);
irl.performIRL();
// System.out.println(this.getFVAndShapeString(rf.getParameters()));
timer.stop();
System.out.println("Training time: " + timer.getTime());
/*//uncomment to run test examples
String baseName = "RH45";
Policy p = new GreedyQPolicy(dss);
GridWorldTerminalFunction tf = new GridWorldTerminalFunction(20, 20);
State simple = this.initialState.copy();
GridWorldDomain.setAgent(simple, 18, 0);
EpisodeAnalysis trainedEp1 = p.evaluateBehavior(simple, objectiveRF, tf, 200);
trainedEp1.writeToFile(trainedDir+"/IRL" + baseName + "EpSimple", this.sp);
State hardAgent = this.initialState.copy();
GridWorldDomain.setAgent(hardAgent, 0, 9);
EpisodeAnalysis trainedEp2 = p.evaluateBehavior(hardAgent, objectiveRF, tf, 200);
trainedEp2.writeToFile(trainedDir+"/IRL" + baseName + "EpHardAgent", this.sp);
dss.resetPlannerResults();
int ngx = 12;
int ngy = 14;
tf = new GridWorldTerminalFunction(ngx, ngy);
this.puddleMap[ngx][ngy] = 1;
this.puddleMap[20][20] = 0;
//fvgen.setGoal(ngx, ngy);
State hardGoal = this.initialState.copy();
GridWorldDomain.setAgent(hardGoal, 0, 0);
EpisodeAnalysis trainedEp3 = p.evaluateBehavior(hardGoal, objectiveRF, tf, 200);
trainedEp3.writeToFile(trainedDir+"/IRL" + baseName + "EpHardGoal", this.sp);
new EpisodeSequenceVisualizer(this.v, this.domain, this.sp, this.trainedDir);
*/
}
