Java Policy.evaluateBehavior方法代码示例

import burlap.behavior.policy.Policy; //导入方法依赖的package包/类
public static void IPSS(){

		InvertedPendulum ip = new InvertedPendulum();
		ip.physParams.actionNoise = 0.;
		Domain domain = ip.generateDomain();
		RewardFunction rf = new InvertedPendulum.InvertedPendulumRewardFunction(Math.PI/8.);
		TerminalFunction tf = new InvertedPendulum.InvertedPendulumTerminalFunction(Math.PI/8.);
		State initialState = InvertedPendulum.getInitialState(domain);

		SparseSampling ss = new SparseSampling(domain, rf, tf, 1, new SimpleHashableStateFactory(), 10 ,1);
		ss.setForgetPreviousPlanResults(true);
		ss.toggleDebugPrinting(false);
		Policy p = new GreedyQPolicy(ss);

		EpisodeAnalysis ea = p.evaluateBehavior(initialState, rf, tf, 500);
		System.out.println("Num steps: " + ea.maxTimeStep());
		Visualizer v = InvertedPendulumVisualizer.getInvertedPendulumVisualizer();
		new EpisodeSequenceVisualizer(v, domain, Arrays.asList(ea));

	}
 

import burlap.behavior.policy.Policy; //导入方法依赖的package包/类
public static void main(String [] args){

		GridWorldDomain gwd = new GridWorldDomain(11, 11);
		gwd.setMapToFourRooms();

		//only go in intended directon 80% of the time
		gwd.setProbSucceedTransitionDynamics(0.8);

		Domain domain = gwd.generateDomain();

		//get initial state with agent in 0,0
		State s = GridWorldDomain.getOneAgentNoLocationState(domain);
		GridWorldDomain.setAgent(s, 0, 0);

		//all transitions return -1
		RewardFunction rf = new UniformCostRF();

		//terminate in top right corner
		TerminalFunction tf = new GridWorldTerminalFunction(10, 10);

		//setup vi with 0.99 discount factor, a value
		//function initialization that initializes all states to value 0, and which will
		//run for 30 iterations over the state space
		VITutorial vi = new VITutorial(domain, rf, tf, 0.99, new SimpleHashableStateFactory(),
				new ValueFunctionInitialization.ConstantValueFunctionInitialization(0.0), 30);

		//run planning from our initial state
		Policy p = vi.planFromState(s);

		//evaluate the policy with one roll out visualize the trajectory
		EpisodeAnalysis ea = p.evaluateBehavior(s, rf, tf);

		Visualizer v = GridWorldVisualizer.getVisualizer(gwd.getMap());
		new EpisodeSequenceVisualizer(v, domain, Arrays.asList(ea));

	}
 

import burlap.behavior.policy.Policy; //导入方法依赖的package包/类
public static void MCLSPIFB(){

		MountainCar mcGen = new MountainCar();
		Domain domain = mcGen.generateDomain();
		TerminalFunction tf = new MountainCar.ClassicMCTF();
		RewardFunction rf = new GoalBasedRF(tf, 100);

		StateGenerator rStateGen = new MCRandomStateGenerator(domain);
		SARSCollector collector = new SARSCollector.UniformRandomSARSCollector(domain);
		SARSData dataset = collector.collectNInstances(rStateGen, rf, 5000, 20, tf, null);

		ConcatenatedObjectFeatureVectorGenerator featureVectorGenerator = new ConcatenatedObjectFeatureVectorGenerator(true, MountainCar.CLASSAGENT);
		FourierBasis fb = new FourierBasis(featureVectorGenerator, 4);

		LSPI lspi = new LSPI(domain, 0.99, fb, dataset);
		Policy p = lspi.runPolicyIteration(30, 1e-6);

		Visualizer v = MountainCarVisualizer.getVisualizer(mcGen);
		VisualActionObserver vob = new VisualActionObserver(domain, v);
		vob.initGUI();

		SimulatedEnvironment env = new SimulatedEnvironment(domain, rf, tf, MountainCar.getCleanState(domain, mcGen.physParams));
		EnvironmentServer envServ = new EnvironmentServer(env, vob);

		for(int i = 0; i < 5; i++){
			p.evaluateBehavior(envServ);
			envServ.resetEnvironment();
		}

		System.out.println("Finished");


	}
 

import burlap.behavior.policy.Policy; //导入方法依赖的package包/类
public static void main(String[] args) {

		MountainCar mcGen = new MountainCar();
		Domain domain = mcGen.generateDomain();
		TerminalFunction tf = new MountainCar.ClassicMCTF();
		RewardFunction rf = new GoalBasedRF(tf, 100);

		StateGenerator rStateGen = new MCRandomStateGenerator(domain);
		SARSCollector collector = new SARSCollector.UniformRandomSARSCollector(domain);
		SARSData dataset = collector.collectNInstances(rStateGen, rf, 5000, 20, tf, null);

		ConcatenatedObjectFeatureVectorGenerator fvGen = new ConcatenatedObjectFeatureVectorGenerator(true,
				MountainCar.CLASSAGENT);
		FourierBasis fb = new FourierBasis(fvGen, 4);

		LSPI lspi = new LSPI(domain, 0.99, fb, dataset);
		Policy p = lspi.runPolicyIteration(30, 1e-6);

		Visualizer v = MountainCarVisualizer.getVisualizer(mcGen);
		VisualActionObserver vob = new VisualActionObserver(domain, v);
		vob.initGUI();

		SimulatedEnvironment env = new SimulatedEnvironment(domain, rf, tf,
				MountainCar.getCleanState(domain, mcGen.physParams));
		EnvironmentServer envServ = new EnvironmentServer(env, vob);

		for(int i = 0; i < 100; i++){
			p.evaluateBehavior(envServ);
			envServ.resetEnvironment();
		}

		System.out.println("Finished");

	}
 

import burlap.behavior.policy.Policy; //导入方法依赖的package包/类
public static void MCLSPIRBF(){

		MountainCar mcGen = new MountainCar();
		Domain domain = mcGen.generateDomain();
		TerminalFunction tf = new MountainCar.ClassicMCTF();
		RewardFunction rf = new GoalBasedRF(tf, 100);
		State s = MountainCar.getCleanState(domain, mcGen.physParams);

		StateGenerator rStateGen = new MCRandomStateGenerator(domain);
		SARSCollector collector = new SARSCollector.UniformRandomSARSCollector(domain);
		SARSData dataset = collector.collectNInstances(rStateGen, rf, 5000, 20, tf, null);

		RBFFeatureDatabase rbf = new RBFFeatureDatabase(true);
		StateGridder gridder = new StateGridder();
		gridder.gridEntireDomainSpace(domain, 5);
		List<State> griddedStates = gridder.gridInputState(s);
		DistanceMetric metric = new EuclideanDistance(
				new ConcatenatedObjectFeatureVectorGenerator(true, MountainCar.CLASSAGENT));
		for(State g : griddedStates){
			rbf.addRBF(new GaussianRBF(g, metric, .2));
		}

		LSPI lspi = new LSPI(domain, 0.99, rbf, dataset);
		Policy p = lspi.runPolicyIteration(30, 1e-6);

		Visualizer v = MountainCarVisualizer.getVisualizer(mcGen);
		VisualActionObserver vob = new VisualActionObserver(domain, v);
		vob.initGUI();


		SimulatedEnvironment env = new SimulatedEnvironment(domain, rf, tf, s);
		EnvironmentServer envServ = new EnvironmentServer(env, vob);

		for(int i = 0; i < 5; i++){
			p.evaluateBehavior(envServ);
			envServ.resetEnvironment();
		}

		System.out.println("Finished");


	}
 

import burlap.behavior.policy.Policy; //导入方法依赖的package包/类
public static void main(String[] args) {
	GridWorldDomain gwd = new GridWorldDomain(11, 11);
	Domain domain = gwd.generateDomain();
	State s = GridWorldDomain.getOneAgentNoLocationState(domain, 1, 3);

	Policy p = new RandomPolicy(domain);
	EpisodeAnalysis ea = p.evaluateBehavior(s, new NullRewardFunction(), new NullTermination(), 30);

	String yamlOut = ea.serialize();

	System.out.println(yamlOut);

	System.out.println("\n\n");

	EpisodeAnalysis read = EpisodeAnalysis.parseEpisode(domain, yamlOut);

	System.out.println(read.getActionSequenceString());
	System.out.println(read.getState(0).toString());
	System.out.println(read.actionSequence.size());
	System.out.println(read.stateSequence.size());

}