ICMOptC -

User Documentation

Centre for Vision, Speech & Signal Processing

Comments:
------------------------------------------------------------------------- ********** Iterated Conditional Modes *********************************** ------------------------------------------------------------------------- ICM - otherwise known as steepest descent or zero temp simulated annealing This is a minimizer. Major choices: Apart from the usual parameters you must choose - assignment or not. The assignment flag selects that the solution must be a permutation matrix. - some variables may be clamped (fixed) and the optimisation proceeds only over other variables HOW DOES IT WORK? ----------------- The non-assignment method in one pass - looks at each object and computes the energy of all possible labels - the label is reassigned to have the lowest energy The assignment method in one pass - starts with a supplied solution - computes the energy of a swap - swaps to the lower energy state The execution times are NOT comparable as the permutation algorithm has a much smaller search space. Parameters ---------- nobjs nlabs idum ------------------------------------------------------------------------- -------------------------------------------------------------------------

Parent Classes:

BaseOptC

Variables:

Methods:

ICMOptC(int _nobjs,int _nlabs,int _idum = -1,TraceC tr = TraceC(NoTrace,cout),BooleanT _assignment = FALSE)
~ICMOptC(): Destructor
void ResetSeed(): set the seed to its first given value if negative this restarts the random seqeunce
HardAssignC Optimize(ICMCostC & func)
HardAssignC Optimize(ICMCostC & func,SArray1dC<BooleanT> _clamped,HardAssignC & start): Optimize the specified function according to previously specified parameters. An initial guess may be supplied and is important for clamped problems! For assignment problems start must be a valid permutation solution!
HardAssignC MultiOptimize(ICMCostC & func,int n)
HardAssignC MultiOptimize(ICMCostC & func,int n,SArray1dC<BooleanT> _clamped,HardAssignC & start): multiple starts of the previous function.
HardAssignC DoOptimize(ICMCostC & cost,HardAssignC & work): does the real work! Starts at given labelling takes account of clamp array

#include "amma/BaseOpt.hh"

HardAssignC Optimize(BaseCostC & func)
HardAssignC Optimize(BaseCostC & func,SArray1dC<BooleanT> _clamped,HardAssignC & start)
void ResetSeed()
int NObjs() const
int NLabs() const
HardAssignC RanHardLabel(int & idum,BooleanT ass = FALSE): Random initial assignment of labels
SoftAssignC RanSoftLabel(int & idum): Random initial assignment of labels
void AddNoise(SoftAssignC & work,double x,int & idum,double minp = 1e-10): add uniform noise scaled by x to work set negatives to positive and set minimum to minp does NOT normalize
void NormaliseSoft(SoftAssignC & w): normalise the label set
SoftAssignC NormaliseSoftLabels(SoftAssignC & wold): normalise the label set
SoftAssignC NormaliseSoftObjects(SoftAssignC & wold): normalise the object set
double Difference(SoftAssignC & w1,SoftAssignC & w2): sum of absolute differences
HardAssignC SoftToHard(const SoftAssignC & ass): Random initial assignment of labels
void PrintSoftLabel(SoftAssignC & r): print out
void PrintHardLabel(HardAssignC & r): print out
BooleanT CheckPermute(HardAssignC & r): checks to see that the labelling is a permutation
void ForcePermute(HardAssignC & r): forces the solution to be a permutation by a crude and possibly slow algorithm

#include "amma/Trace.hh"

Inline TraceC(int tl = NoTrace,ostream & os = cout): Constructor
Inline ~TraceC(): Destructor
Inline TraceC(const TraceC & t): Copy Constructor

Programmer:Andrew Stoddart, Documentation by CxxDoc: Tue Mar 20 10:48:08 2001

ICMOptC

include	"amma/SAOpt.hh"
User Level:	Default
Library:	CombOpt
Example:	doNumClassComb.cc
Section:	Optimisation.Implementation Numerical Methods.Optimisation.Implementation
In Scope:	std