cpolicygradient.h

强化学习算法（R-Learning）难得的珍贵资料

// Copyright (C) 2003
// Gerhard Neumann ([e-mail])

//                
// This file is part of RL Toolbox.
// http://www.igi.tugraz.at/ril_toolbox
//
// All rights reserved.
// 
// Redistribution and use in source and binary forms, with or without
// modification, are permitted provided that the following conditions
// are met:
// 1. Redistributions of source code must retain the above copyright
//    notice, this list of conditions and the following disclaimer.
// 2. Redistributions in binary form must reproduce the above copyright
//    notice, this list of conditions and the following disclaimer in the
//    documentation and/or other materials provided with the distribution.
// 3. The name of the author may not be used to endorse or promote products
//    derived from this software without specific prior written permission.
// 
// THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
// IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
// OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
// IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
// INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
// NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
// THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.

#ifndef C_POLICYGRADIENT__H
#define C_POLICYGRADIENT__H

#include "cparameters.h"
#include "cagentlistener.h"
#include "cfeaturefunction.h"
#include "caction.h"
#include "cpolicies.h"
#include "creinforce.h"
#include "cagent.h"
#include "cmontecarlo.h"
#include "ril_debug.h"

class CPolicyGradientCalculator : virtual public CParameterObject
{
protected:
	CAgentController *policy;
public:
	CPolicyGradientCalculator(CAgentController *policy);

	virtual void getGradient(CFeatureList *gradient) = 0;
};

class CGPOMDPGradientCalculator : public CPolicyGradientCalculator, public CSemiMDPRewardListener
{
protected:
	CFeatureList *localGradient;
	CFeatureList *localZTrace;

	CFeatureList *globalGradient;

	CAgent *agent; 
	CReinforcementBaseLineCalculator *baseLine;

	CStochasticPolicy *stochPolicy;

public:
	CGPOMDPGradientCalculator(CRewardFunction *reward, CStochasticPolicy *policy, CAgent *agent, CReinforcementBaseLineCalculator *baseLine, int TSteps, int nEpisodes, rlt_real beta);
	~CGPOMDPGradientCalculator();

	virtual void nextStep(CStateCollection *oldState, CAction *action, rlt_real reward, CStateCollection *newState);
	virtual void newEpisode();

	virtual void getGradient(CFeatureList *gradient);

	virtual CFeatureList* getGlobalGradient();
	virtual void setGlobalGradient(CFeatureList *globalGradient);
};

class CPolicyGradientUpdater : virtual public CParameterObject
{
protected:
	CGradientUpdateFunction *updateFunction;
public:
	CPolicyGradientUpdater(CGradientUpdateFunction *updateFunction);

	virtual void updateWeights(CFeatureList *gradient) = 0;
	void addRandomParams(rlt_real randSize);

	CGradientUpdateFunction *getUpdateFunction() {return updateFunction;};
};

class CConstantPolicyGradientUpdater : public CPolicyGradientUpdater
{
public:
	CConstantPolicyGradientUpdater(CGradientUpdateFunction *updateFunction, rlt_real learningRate);

	virtual void updateWeights(CFeatureList *gradient);
};

class CGSearchPolicyGradientUpdater : public CPolicyGradientUpdater
{
protected:
	CPolicyGradientCalculator *gradientCalculator;

	rlt_real *startParameters;
	rlt_real *workParameters;

	rlt_real lastStepSize;


	void setWorkingParamters(CFeatureList *gradient, rlt_real stepSize, rlt_real *startParameters, rlt_real *workParameters);
public:

	CGSearchPolicyGradientUpdater(CGradientUpdateFunction *updateFunction, CPolicyGradientCalculator *gradientCalculator, rlt_real s0, rlt_real epsilon);
	~CGSearchPolicyGradientUpdater();

	virtual void updateWeights(CFeatureList *gradient);
};

class CLineSearchPolicyGradientUpdater : public CPolicyGradientUpdater
{
protected:
	CPolicySameStateEvaluator *evaluator;

	rlt_real *startParameters;
	rlt_real *workParameters;

	rlt_real *startStepSizes;
	int numStepSizes;
	int maxSteps;

	void setWorkingParamters(CFeatureList *gradient, rlt_real stepSize, rlt_real *startParameters, rlt_real *workParameters);
public:

	CLineSearchPolicyGradientUpdater(CGradientUpdateFunction *updateFunction, CPolicySameStateEvaluator *policyEvaluator, rlt_real *startStepSizes, int numStepSizes, int maxSteps);
	virtual ~CLineSearchPolicyGradientUpdater();

	virtual void updateWeights(CFeatureList *gradient);
};

class CPolicyGradientLearner : virtual public CParameterObject
{
protected:
	CPolicyGradientCalculator *gradientCalculator;
	CPolicyGradientUpdater *gradientUpdater;

	CFeatureList *gradient;
	CFeatureList *hGradient;
	CFeatureList *gGradient;

	virtual void doUpdate(CFeatureList *gradient);

public:
	CPolicyGradientLearner(CPolicyGradientCalculator *gradientCalculator, CPolicyGradientUpdater *gradientUpdater, rlt_real epsilon);
	~CPolicyGradientLearner();

	rlt_real learnPolicy(int maxGradientUpdates, CPolicyEvaluator *evaluator, bool useOldGradient = false);
	
};

class CPolicyGradientWeightDecayListener : public CSemiMDPListener
{
protected:
	CGradientUpdateFunction *updateFunction;
	rlt_real *parameters;
public:
	CPolicyGradientWeightDecayListener(CGradientUpdateFunction *updateFunction, rlt_real weightdecay);
	~CPolicyGradientWeightDecayListener();

	virtual void newEpisode();
};



#endif