Main Page   Namespace List   Class Hierarchy   Alphabetical List   Compound List   Namespace Members   Compound Members   Related Pages  

ObjCryst::MonteCarloObj Class Reference

Base object for Monte-Carlo Global Optimization methods. More...

#include <GlobalOptimObj.h>

Inheritance diagram for ObjCryst::MonteCarloObj::

ObjCryst::OptimizationObj List of all members.

Public Methods

 MonteCarloObj (const string name="")
 Constructor.

 MonteCarloObj (const bool internalUseOnly)
 Constructor. Using internalUseOnly=true will avoid registering the the object to any registry, and thus (for example) no display will be created, nor will this object be automatically be saved.

virtual ~MonteCarloObj ()
 Destructor.

void SetAlgorithmSimulAnnealing (const AnnealingSchedule scheduleTemp, const REAL tMax, const REAL tMin, const AnnealingSchedule scheduleMutation=ANNEALING_CONSTANT, const REAL mutMax=16., const REAL mutMin=.125, const long nbTrialRetry=0, const REAL minCostRetry=0., const long maxNbTrialSinceBest=0)
 Set the refinement method to simulated Annealing. Note that Parellel Tempering is more efficient to get out of local minima, so you sould rather use that method. More...

void SetAlgorithmParallTempering (const AnnealingSchedule scheduleTemp, const REAL tMax, const REAL tMin, const AnnealingSchedule scheduleMutation=ANNEALING_CONSTANT, const REAL mutMax=16., const REAL mutMin=.125)
 Set the refinement method to Parallel Tempering. More...

virtual void Optimize (long &nbSteps, const bool silent=false, const REAL finalcost=0)
 Launch optimization for N steps
Parameters:
nbSteps:  the number of steps to go. This number is modified (decreases!) as the refinement goes on.
silent  : if true, absolutely no message should be printed (except debugging).


void SaveOptimHistory () const
 Save history of the evolution of parameters to a file. Only non-fixed parameters are saved. This saves a very crude array in which can bve found the value of all non-fixed parameters for successive "best" configurations.

virtual void XMLOutput (ostream &os, int indent=0) const
 Output a description of the object in XML format to a stream. More...

virtual void XMLInput (istream &is, const XMLCrystTag &tag)
 Input in XML format from a stream, restoring the set of refined objects and the associated cost functions. Note that the corresponding objects must have been loaded in memory before, else shit happens.

virtual const string GetClassName () const
 Get the name for this class type.


Protected Methods

virtual void NewConfiguration ()
 Make a random change in the configuration. More...

virtual void InitOptions ()
 Initialization of options.


Protected Attributes

RefObjOpt mGlobalOptimType
 Method used for the global optimization. Should be removed when we switch to using several classes for different algorithms.

REAL mCurrentCost
 Current value of the cost function.

long mHistoryNb
 Total number of saved configurations.

CrystVector_long mHistoryTrialNumber
 Trials corresponding to each stored values.

CrystVector_REAL mHistoryCostFunction
 Evolution of cost function.

CrystVector_long mHistorySavedParamSetIndex
 Index of saved parameters set in mRefParList for each saved trial.

bool mHistorySaveAfterEachOptim
 Save the evolution of refined parameters after optimization ?

string mHistorySaveFileName
 Save the evolution of refined parameters after optimization ?

long mLastParSavedSetIndex
 Index of the 'last' parameter set.

REAL mTemperatureMax
 Beginning temperature for annealing.

REAL mTemperatureMin
 Lower temperature.

RefObjOpt mAnnealingScheduleTemp
 Schedule for the annealing.

REAL mMutationAmplitude
 Mutation amplitude. From .25 to 64. Random moves will have a maximum amplitude equal to this amplitude multiplied by the Global optimization step defined for each RefinablePar. Large amplitude should be used at the beginning of the refinement (high temeratures).

REAL mMutationAmplitudeMax
 Mutation amplitude at the beginning of the optimization.

REAL mMutationAmplitudeMin
 Mutation amplitude at the end of the optimization.

RefObjOpt mAnnealingScheduleMutation
 Schedule for the annealing.

long mNbTrialRetry
 Number of trials before testing if we are below the given minimum cost. If <=0, this will be ignored.

REAL mMinCostRetry
 Cost to reach unless an automatic randomization and retry is done.

long mMaxNbTrialSinceBest
 If more than mMaxNbTrialSinceBest trials have been made since the best configuration has been found, then revert to the best configuration. If <=0, then this is ignored. This must be large enough to have an ergodic algorithm (more strictly, should not be used ?).


Detailed Description

Base object for Monte-Carlo Global Optimization methods.

The algorithm is quite simple, whith two type of optimizations, either simulated Annealing or Parallel Tempering, the latter being recommanded for most real-world optimizations.


Member Function Documentation

virtual void ObjCryst::MonteCarloObj::NewConfiguration   [protected, virtual]
 

Make a random change in the configuration.

For internal use only.

This just generates a new configuration with random changes (according to current parameters). The old config is stored in mRefParList as the last config (index mLastParSavedSetIndex). The new one is not tested in this function vs temperature: this should be done in the OptimizationObj::Optimize() function, which also chooses whether to revert to the previous configuration.

Random moves are made by the objects and not by this function, because the new configuration can be specific (like, for example, permutations between some of the parameters (atoms)).

void ObjCryst::MonteCarloObj::SetAlgorithmParallTempering const AnnealingSchedule    scheduleTemp,
const REAL    tMax,
const REAL    tMin,
const AnnealingSchedule    scheduleMutation = ANNEALING_CONSTANT,
const REAL    mutMax = 16.,
const REAL    mutMin = .125
 

Set the refinement method to Parallel Tempering.

The refinement begins at max and finishes at min temperature.

Parameters:
scheduleTemp:  temperature schedule
tMax,tMin:  Max and Min temperatures. See AnnealingSchedule.
scheduleMutation:  the mutation schedule. For each new configuration, the variation of each refinable parameter is less than its RefinablePar::GlobalOptimStep(), multiplied by the current mutation amplitude. By default this mutation is equal to 1., but making bigger steps can be a good idea at the beginning of the refinement. Thus you can choose a schedule for the amplitude, exactly like for the temperature.
mutMax,mutMin:  Max and Min mutation amplitudes.
Warning:
do not use the 'smart' option for the temperature schedule, it is not yet implemented. Later it will be used to set the temperatures as a function of the amplitude schedule, so that we keep accepeted move between 30% and 70%.
Note:
this will be removed when we separate the different algorithms in different classes.

void ObjCryst::MonteCarloObj::SetAlgorithmSimulAnnealing const AnnealingSchedule    scheduleTemp,
const REAL    tMax,
const REAL    tMin,
const AnnealingSchedule    scheduleMutation = ANNEALING_CONSTANT,
const REAL    mutMax = 16.,
const REAL    mutMin = .125,
const long    nbTrialRetry = 0,
const REAL    minCostRetry = 0.,
const long    maxNbTrialSinceBest = 0
 

Set the refinement method to simulated Annealing. Note that Parellel Tempering is more efficient to get out of local minima, so you sould rather use that method.

The refinement begins at max and finishes at min temperature.

Parameters:
scheduleTemp:  temperature schedule. See AnnealingSchedule.
tMax,tMin:  Max and Min temperatures.
scheduleMutation:  the mutation schedule. For each new configuration, the variation of each refinable parameter is less than its RefinablePar::GlobalOptimStep(), multiplied by the current mutation amplitude. By default this mutation is equal to 1., but making bigger steps is a good idea at the beginning of the refinement (for higher temperatures). See AnnealingSchedule. See AnnealingSchedule.
mutMax,mutMin:  Max and Min mutation amplitudes.
minCostRetry, nbTrialRetry:  if after nbTrialRetry, the cost function is still above minCostRetry, then start again from a random configuration. No randomization is made if nbTrialRetry <= 0.
maxNbTrialSinceBest:  if more than maxNbTrialSinceBest trials have been made since the best configuration was recorded, then revert to that configuration. This should be large enough to have an ergodic search (the default is never to revert..)
Warning:
do not use the 'smart' option for the temperature schedule, it is not yet implemented. Later it will be used to set the temperatures as a function of the amplitude schedule, so that we accept between 30% and 70% moves.
Note:
this will be removed when we separate the different algorithms in different classes.

virtual void ObjCryst::MonteCarloObj::XMLOutput ostream &    os,
int    indent = 0
const [virtual]
 

Output a description of the object in XML format to a stream.

This saves the list of refined object and the cost functions, as well as options for the refinement. The refined objects are not saved, so this must be done somewhere else (they must be reloaded before this object).

Reimplemented from ObjCryst::OptimizationObj.


The documentation for this class was generated from the following file:
Generated on Tue Feb 5 17:37:21 2002 for ObjCryst++ by doxygen1.2.13.1 written by Dimitri van Heesch, © 1997-2001