📄 estimator.h

📁 Amis - A maximum entropy estimator 一个最大熵模型统计工具
💻 H
字号:
////////////////////////////////////////////////////////////////////////////  Copyright (c) 2000, Yusuke Miyao///  You may distribute under the terms of the Artistic License.//////  <id>$Id: Estimator.h,v 1.8 2003/05/16 12:10:44 yusuke Exp $</id>///  <collection>Maximum Entropy Estimator</collection>///  <name>Estimator.h</name>///  <overview>Base class for Maximum Entropy Estimators</overview>/////////////////////////////////////////////////////////////////////////#ifndef Amis_Estimator_h_#define Amis_Estimator_h_#include <amis/configure.h>#include <amis/Real.h>#include <amis/ModelBase.h>#include <amis/EventSpaceBase.h>#include <amis/Profile.h>#include <amis/LogStream.h>#include <amis/EmpiricalExpect.h>#include <iomanip>#include <iostream>#include <memory>AMIS_NAMESPACE_BEGIN///////////////////////////////////////////////////////////////////////// <classdef>/// <name>EstimatorInitError</name>/// <overview>Error for failing initialization of an estimator</overview>/// <desc>/// The error is thrown when an initialization of an estimator failed./// </desc>/// <body>class EstimatorInitError : public ErrorBase { public:  EstimatorInitError( const std::string& s ) : ErrorBase( s ) {}  /// Initialize with an error message  EstimatorInitError( char* s ) : ErrorBase( s ) {}  /// Initialize with an error message};/// </body>/// </classdef>///////////////////////////////////////////////////////////////////////// <classdef>/// <name>Estimator</name>/// <overview>Base class for Maximum Entropy Estimators</overview>/// <desc>/// This class is a base class for maximum entropy estimators/// such as GIS and IIS./// </desc>/// <body>class Estimator {protected:  int iteration_count;  /// Iteration counter  int report_interval;  /// Number of report interval  Real machine_epsilon;  /// Minimum significant number  int  estimation_precision;    Real log_likelihood;  /// Log-likelihood of the event space with the current model  Real log_likelihood_per_event;  /// Log-likelihood of the event space with the current model (per event average)    Real objective_function_value;  /// The current value of the objective function of the estimation algorithm (as maximization)  Real prev_objective_function_value;  /// The value of the objective function at the previous iteration    Real min_update;  /// Minimum value of model update  Real max_update;  /// Maximum value of model update  bool till_convergence;  /// Whether to iterate until convergence  bool suppress_message;  /// Whether to show messages  ProfTimer profile_timer;  /// Timer to measure the estimation time  public:  static const int DEFAULT_NUM_ITERATIONS = 100;  /// Default number of iterations  static const int DEFAULT_PRECISION = 6;  /// Default precision  static const int DEFAULT_REPORT_INTERVAL = 1;  /// Default number of report interval  static const int MAX_NUM_ITERATIONS = 5000;    static const int CONVERGED_ATOL = 0;  static const int CONVERGED_RTOL = 1;  static const int BAD_INCREASE   = 2;    Real ATOL;  Real RTOL;  public:  Estimator();  virtual ~Estimator() {}public:  virtual void setModel( ModelBase& model ) = 0;  virtual void setEventSpace( EventSpaceBase& event_space ) = 0;  virtual const ModelBase& getModel() const = 0;  virtual const EventSpaceBase& getEventSpace() const = 0;protected:  void initMinMaxUpdate() {    min_update = REAL_MAX;    max_update = -REAL_MAX;  }  void setMinMaxUpdate( Real m ) {    if ( min_update > m ) min_update = m;    if ( max_update < m ) max_update = m;  }public:  int iterationCount() const { return iteration_count; }  /// Get the iteration counter  int reportInterval() const { return report_interval; }  /// Get the report interval  bool isReportIteration() const { return ( iteration_count % report_interval == 0 ); }  /// Whether this iteration is a report iteration  Real machineEpsilon() const { return machine_epsilon; }  /// Get the machine-epsilon value  int estimationPrecision() const { return estimation_precision; }    Real logLikelihood() const { return log_likelihood; }  /// Get the log-likelihood of the event space with the current model  Real logLikelihoodPerEvent() const { return log_likelihood_per_event; }  /// Get the log-likelihood of the event space with the current model (per event average)  Real objectiveFunctionValue() const { return objective_function_value; }  /// Get the value of the objective function  Real previousObjectiveFunctionValue() const { return prev_objective_function_value; }  /// Get the previous value of the objective function  Real minUpdate() const { return min_update; }  /// Get the minimum value of the model  Real maxUpdate() const { return max_update; }  /// Get the maximum value of the model  bool tillConvergence() { return till_convergence; }  void setTillConvergence( bool t = true ) { till_convergence = t; }  bool suppressMessage() { return suppress_message; }  void setSuppressMessage( bool t = true ) { suppress_message = t; }  const ProfTimer& getEstimationTimer() { return profile_timer; }protected:  void setLogLikelihood( Real ll_pe, Real sum_event_count ) {    log_likelihood_per_event = ll_pe;    log_likelihood = ll_pe * sum_event_count;  }  void setObjectiveFunctionValue( Real fv ) {	  objective_function_value = fv;  }  void setPreviousObjectiveFunctionValue( Real fv ) {	  prev_objective_function_value = fv;  }protected:  virtual void initialize() = 0;  /// Initialization of the internal data before estimation  virtual void finalize() = 0;  /// Finalization of the internal data after estimation  virtual void iteration() = 0;  /// Each iteration for estimation  virtual void postIteration() {}  /// After one iteration  virtual void startupMessage();  /// Print start-up message  virtual void reportMessage();  /// Print a message for each report iteration  virtual void endingMessage();  /// Print ending messagepublic:  virtual bool isConverged() = 0;  /// Check whether the model is converged  virtual const std::string estimatorName() const = 0;  /// Name of the estimator  virtual const EmpiricalExpectBase& getEmpiricalExpectation() = 0;  /// Get the empirical expectation  //virtual Real setLogLikelihood() = 0;  virtual void estimate( int num_iterations = DEFAULT_NUM_ITERATIONS,                         int precision = DEFAULT_PRECISION,                         int r = DEFAULT_REPORT_INTERVAL );  /// Estimation of parameters of a maximum entropy model  /*    virtual bool checkConvergence( double pf, double f, int* reason ) {	  	  /// Use Amis style (I don't know which is better. 	  return checkConvergenceAmisStyle( pf, f, reason );  }  virtual bool checkConvergenceTaoStyle( double pf, double f, int* reason ) {	  AMIS_DEBUG_MESSAGE( 3, "checkConvergenceTaoStyle\n" );	  AMIS_DEBUG_MESSAGE( 3, "pf = " << pf << " f=" << f << " atol=" << atol << " rtol=" << rtol << '\n' );	  if( pf - f <= ATOL ) {		  AMIS_DEBUG_MESSAGE( 3, "ABS_CONV=true\n" );		  *reason = CONVERGED_ATOL;		  return true;	  }	  	  if( fabs( pf - f ) / ( fabs( pf ) + 1 ) <= RTOL ) {		  AMIS_DEBUG_MESSAGE( 3, "REL_CONV=true\n" );		  *reason = CONVERGED_RTOL;		  return true;	  }	  	  return false;  }    virtual bool checkConvergenceAmisStyle( double pf, double f, int* reason ) {	  /// From the implementation of BFGSSolver by Y. Miyao	  AMIS_DEBUG_MESSAGE( 3, "checkConvergenceAmisStyle\n" );	  AMIS_DEBUG_MESSAGE( 3, "pf = " << pf << " f=" << f << " atol=" << ATOL << " rtol=" << RTOL << '\n' );	  	  if( pf - f <= ATOL ) {		  		  if( pf - f <= 0.0 ) {			  *reason = BAD_INCREASE;		  }		  else{			  *reason = CONVERGED_ATOL;		  }		  return true;	  }	  	  	  if( fabs( 1.0 - fabs( f/ pf )  ) <= RTOL ) {		  *reason = CONVERGED_RTOL;		  return true;	  }	  //std::cerr << "check failed" << std::endl;	  return false;  }  */};typedef std::auto_ptr< Estimator > EstimatorPtr;AMIS_NAMESPACE_END/// </body>/// </classdef>#endif // Estimator_h_// end of Estimator.h
⌨️ 快捷键说明

复制代码 Ctrl + C
搜索代码 Ctrl + F
全屏模式 F11
切换主题 Ctrl + Shift + D
显示快捷键 ?
增大字号 Ctrl + =
减小字号 Ctrl + -