limp-cbc-0.3.2.0: cbits/coin/CbcHeuristic.hpp
/* $Id: CbcHeuristic.hpp 1883 2013-04-06 13:33:15Z stefan $ */
// Copyright (C) 2002, International Business Machines
// Corporation and others. All Rights Reserved.
// This code is licensed under the terms of the Eclipse Public License (EPL).
#ifndef CbcHeuristic_H
#define CbcHeuristic_H
#include <string>
#include <vector>
#include "CoinPackedMatrix.hpp"
#include "OsiCuts.hpp"
#include "CoinHelperFunctions.hpp"
#include "OsiBranchingObject.hpp"
class OsiSolverInterface;
class CbcModel;
//#############################################################################
class CbcHeuristicNodeList;
class CbcBranchingObject;
/** A class describing the branching decisions that were made to get
to the node where a heuristic was invoked from */
class CbcHeuristicNode {
private:
void gutsOfConstructor(CbcModel& model);
CbcHeuristicNode();
CbcHeuristicNode& operator=(const CbcHeuristicNode&);
private:
/// The number of branching decisions made
int numObjects_;
/** The indices of the branching objects. Note: an index may be
listed multiple times. E.g., a general integer variable that has
been branched on multiple times. */
CbcBranchingObject** brObj_;
public:
CbcHeuristicNode(CbcModel& model);
CbcHeuristicNode(const CbcHeuristicNode& rhs);
~CbcHeuristicNode();
double distance(const CbcHeuristicNode* node) const;
double minDistance(const CbcHeuristicNodeList& nodeList) const;
bool minDistanceIsSmall(const CbcHeuristicNodeList& nodeList,
const double threshold) const;
double avgDistance(const CbcHeuristicNodeList& nodeList) const;
};
class CbcHeuristicNodeList {
private:
void gutsOfDelete();
void gutsOfCopy(const CbcHeuristicNodeList& rhs);
private:
std::vector<CbcHeuristicNode*> nodes_;
public:
CbcHeuristicNodeList() {}
CbcHeuristicNodeList(const CbcHeuristicNodeList& rhs);
CbcHeuristicNodeList& operator=(const CbcHeuristicNodeList& rhs);
~CbcHeuristicNodeList();
void append(CbcHeuristicNode*& node);
void append(const CbcHeuristicNodeList& nodes);
inline const CbcHeuristicNode* node(int i) const {
return nodes_[i];
}
inline int size() const {
return static_cast<int>(nodes_.size());
}
};
//#############################################################################
/** Heuristic base class */
class CbcHeuristic {
private:
void gutsOfDelete() {}
void gutsOfCopy(const CbcHeuristic & rhs);
public:
// Default Constructor
CbcHeuristic ();
// Constructor with model - assumed before cuts
CbcHeuristic (CbcModel & model);
// Copy constructor
CbcHeuristic ( const CbcHeuristic &);
virtual ~CbcHeuristic();
/// Clone
virtual CbcHeuristic * clone() const = 0;
/// Assignment operator
CbcHeuristic & operator=(const CbcHeuristic& rhs);
/// update model (This is needed if cliques update matrix etc)
virtual void setModel(CbcModel * model);
/// Resets stuff if model changes
virtual void resetModel(CbcModel * model) = 0;
/** returns 0 if no solution, 1 if valid solution
with better objective value than one passed in
Sets solution values if good, sets objective value
This is called after cuts have been added - so can not add cuts
*/
virtual int solution(double & objectiveValue,
double * newSolution) = 0;
/** returns 0 if no solution, 1 if valid solution, -1 if just
returning an estimate of best possible solution
with better objective value than one passed in
Sets solution values if good, sets objective value (only if nonzero code)
This is called at same time as cut generators - so can add cuts
Default is do nothing
*/
virtual int solution2(double & /*objectiveValue*/,
double * /*newSolution*/,
OsiCuts & /*cs*/) {
return 0;
}
/// Validate model i.e. sets when_ to 0 if necessary (may be NULL)
virtual void validate() {}
/** Sets "when" flag - 0 off, 1 at root, 2 other than root, 3 always.
If 10 added then don't worry if validate says there are funny objects
as user knows it will be fine
*/
inline void setWhen(int value) {
when_ = value;
}
/// Gets "when" flag - 0 off, 1 at root, 2 other than root, 3 always
inline int when() const {
return when_;
}
/// Sets number of nodes in subtree (default 200)
inline void setNumberNodes(int value) {
numberNodes_ = value;
}
/// Gets number of nodes in a subtree (default 200)
inline int numberNodes() const {
return numberNodes_;
}
/** Switches (does not apply equally to all heuristics)
1 bit - stop once allowable gap on objective reached
2 bit - always do given number of passes
4 bit - weaken cutoff by 5% every 50 passes?
8 bit - if has cutoff and suminf bobbling for 20 passes then
first try halving distance to best possible then
try keep halving distance to known cutoff
16 bit - needs new solution to run
1024 bit - stop all heuristics on max time
*/
inline void setSwitches(int value) {
switches_ = value;
}
/** Switches (does not apply equally to all heuristics)
1 bit - stop once allowable gap on objective reached
2 bit - always do given number of passes
4 bit - weaken cutoff by 5% every 50 passes?
8 bit - if has cutoff and suminf bobbling for 20 passes then
first try halving distance to best possible then
try keep halving distance to known cutoff
16 bit - needs new solution to run
1024 bit - stop all heuristics on max time
*/
inline int switches() const {
return switches_;
}
/// Whether to exit at once on gap
bool exitNow(double bestObjective) const;
/// Sets feasibility pump options (-1 is off)
inline void setFeasibilityPumpOptions(int value) {
feasibilityPumpOptions_ = value;
}
/// Gets feasibility pump options (-1 is off)
inline int feasibilityPumpOptions() const {
return feasibilityPumpOptions_;
}
/// Just set model - do not do anything else
inline void setModelOnly(CbcModel * model) {
model_ = model;
}
/// Sets fraction of new(rows+columns)/old(rows+columns) before doing small branch and bound (default 1.0)
inline void setFractionSmall(double value) {
fractionSmall_ = value;
}
/// Gets fraction of new(rows+columns)/old(rows+columns) before doing small branch and bound (default 1.0)
inline double fractionSmall() const {
return fractionSmall_;
}
/// Get how many solutions the heuristic thought it got
inline int numberSolutionsFound() const {
return numberSolutionsFound_;
}
/// Increment how many solutions the heuristic thought it got
inline void incrementNumberSolutionsFound() {
numberSolutionsFound_++;
}
/** Do mini branch and bound - return
0 not finished - no solution
1 not finished - solution
2 finished - no solution
3 finished - solution
(could add global cut if finished)
-1 returned on size
-2 time or user event
*/
int smallBranchAndBound(OsiSolverInterface * solver, int numberNodes,
double * newSolution, double & newSolutionValue,
double cutoff , std::string name) const;
/// Create C++ lines to get to current state
virtual void generateCpp( FILE * ) {}
/// Create C++ lines to get to current state - does work for base class
void generateCpp( FILE * fp, const char * heuristic) ;
/// Returns true if can deal with "odd" problems e.g. sos type 2
virtual bool canDealWithOdd() const {
return false;
}
/// return name of heuristic
inline const char *heuristicName() const {
return heuristicName_.c_str();
}
/// set name of heuristic
inline void setHeuristicName(const char *name) {
heuristicName_ = name;
}
/// Set random number generator seed
void setSeed(int value);
/// Get random number generator seed
int getSeed() const;
/// Sets decay factor (for howOften) on failure
inline void setDecayFactor(double value) {
decayFactor_ = value;
}
/// Set input solution
void setInputSolution(const double * solution, double objValue);
/* Runs if bit set
0 - before cuts at root node (or from doHeuristics)
1 - during cuts at root
2 - after root node cuts
3 - after cuts at other nodes
4 - during cuts at other nodes
8 added if previous heuristic in loop found solution
*/
inline void setWhereFrom(int value) {
whereFrom_ = value;
}
inline int whereFrom() const {
return whereFrom_;
}
/** Upto this depth we call the tree shallow and the heuristic can be called
multiple times. That is, the test whether the current node is far from
the others where the jeuristic was invoked will not be done, only the
frequency will be tested. After that depth the heuristic will can be
invoked only once per node, right before branching. That's when it'll be
tested whether the heur should run at all. */
inline void setShallowDepth(int value) {
shallowDepth_ = value;
}
/** How often to invoke the heuristics in the shallow part of the tree */
inline void setHowOftenShallow(int value) {
howOftenShallow_ = value;
}
/** How "far" should this node be from every other where the heuristic was
run in order to allow the heuristic to run in this node, too. Currently
this is tested, but we may switch to avgDistanceToRun_ in the future. */
inline void setMinDistanceToRun(int value) {
minDistanceToRun_ = value;
}
/** Check whether the heuristic should run at all
0 - before cuts at root node (or from doHeuristics)
1 - during cuts at root
2 - after root node cuts
3 - after cuts at other nodes
4 - during cuts at other nodes
8 added if previous heuristic in loop found solution
*/
virtual bool shouldHeurRun(int whereFrom);
/** Check whether the heuristic should run this time */
bool shouldHeurRun_randomChoice();
void debugNodes();
void printDistanceToNodes();
/// how many times the heuristic has actually run
inline int numRuns() const {
return numRuns_;
}
/// How many times the heuristic could run
inline int numCouldRun() const {
return numCouldRun_;
}
/*! \brief Clone, but ...
If type is
- 0 clone the solver for the model,
- 1 clone the continuous solver for the model
- Add 2 to say without integer variables which are at low priority
- Add 4 to say quite likely infeasible so give up easily (clp only).
*/
OsiSolverInterface * cloneBut(int type);
protected:
/// Model
CbcModel * model_;
/// When flag - 0 off, 1 at root, 2 other than root, 3 always
int when_;
/// Number of nodes in any sub tree
int numberNodes_;
/** Feasibility pump options , -1 is off
>=0 for feasibility pump itself
-2 quick proximity search
-3 longer proximity search
*/
int feasibilityPumpOptions_;
/// Fraction of new(rows+columns)/old(rows+columns) before doing small branch and bound
mutable double fractionSmall_;
/// Thread specific random number generator
CoinThreadRandom randomNumberGenerator_;
/// Name for printing
std::string heuristicName_;
/// How often to do (code can change)
int howOften_;
/// How much to increase how often
double decayFactor_;
/** Switches (does not apply equally to all heuristics)
1 bit - stop once allowable gap on objective reached
2 bit - always do given number of passes
4 bit - weaken cutoff by 5% every 50 passes?
8 bit - if has cutoff and suminf bobbling for 20 passes then
first try halving distance to best possible then
try keep halving distance to known cutoff
16 bit - needs new solution to run
1024 bit - stop all heuristics on max time
*/
mutable int switches_;
/* Runs if bit set
0 - before cuts at root node (or from doHeuristics)
1 - during cuts at root
2 - after root node cuts
3 - after cuts at other nodes
4 - during cuts at other nodes
8 added if previous heuristic in loop found solution
*/
int whereFrom_;
/** Upto this depth we call the tree shallow and the heuristic can be called
multiple times. That is, the test whether the current node is far from
the others where the jeuristic was invoked will not be done, only the
frequency will be tested. After that depth the heuristic will can be
invoked only once per node, right before branching. That's when it'll be
tested whether the heur should run at all. */
int shallowDepth_;
/** How often to invoke the heuristics in the shallow part of the tree */
int howOftenShallow_;
/** How many invocations happened within the same node when in a shallow
part of the tree. */
int numInvocationsInShallow_;
/** How many invocations happened when in the deep part of the tree. For
every node we count only one invocation. */
int numInvocationsInDeep_;
/** After how many deep invocations was the heuristic run last time */
int lastRunDeep_;
/// how many times the heuristic has actually run
int numRuns_;
/** How "far" should this node be from every other where the heuristic was
run in order to allow the heuristic to run in this node, too. Currently
this is tested, but we may switch to avgDistanceToRun_ in the future. */
int minDistanceToRun_;
/// The description of the nodes where this heuristic has been applied
CbcHeuristicNodeList runNodes_;
/// How many times the heuristic could run
int numCouldRun_;
/// How many solutions the heuristic thought it got
int numberSolutionsFound_;
/// How many nodes the heuristic did this go
mutable int numberNodesDone_;
// Input solution - so can be used as seed
double * inputSolution_;
#ifdef JJF_ZERO
/// Lower bounds of last node where the heuristic found a solution
double * lowerBoundLastNode_;
/// Upper bounds of last node where the heuristic found a solution
double * upperBoundLastNode_;
#endif
};
/** Rounding class
*/
class CbcRounding : public CbcHeuristic {
public:
// Default Constructor
CbcRounding ();
// Constructor with model - assumed before cuts
CbcRounding (CbcModel & model);
// Copy constructor
CbcRounding ( const CbcRounding &);
// Destructor
~CbcRounding ();
/// Assignment operator
CbcRounding & operator=(const CbcRounding& rhs);
/// Clone
virtual CbcHeuristic * clone() const;
/// Create C++ lines to get to current state
virtual void generateCpp( FILE * fp) ;
/// Resets stuff if model changes
virtual void resetModel(CbcModel * model);
/// update model (This is needed if cliques update matrix etc)
virtual void setModel(CbcModel * model);
using CbcHeuristic::solution ;
/** returns 0 if no solution, 1 if valid solution
with better objective value than one passed in
Sets solution values if good, sets objective value (only if good)
This is called after cuts have been added - so can not add cuts
*/
virtual int solution(double & objectiveValue,
double * newSolution);
/** returns 0 if no solution, 1 if valid solution
with better objective value than one passed in
Sets solution values if good, sets objective value (only if good)
This is called after cuts have been added - so can not add cuts
Use solutionValue rather than solvers one
*/
virtual int solution(double & objectiveValue,
double * newSolution,
double solutionValue);
/// Validate model i.e. sets when_ to 0 if necessary (may be NULL)
virtual void validate();
/// Set seed
void setSeed(int value) {
seed_ = value;
}
protected:
// Data
// Original matrix by column
CoinPackedMatrix matrix_;
// Original matrix by
CoinPackedMatrix matrixByRow_;
// Down locks
unsigned short * down_;
// Up locks
unsigned short * up_;
// Equality locks
unsigned short * equal_;
// Seed for random stuff
int seed_;
};
/** Partial solution class
If user knows a partial solution this tries to get an integer solution
it uses hotstart information
*/
class CbcHeuristicPartial : public CbcHeuristic {
public:
// Default Constructor
CbcHeuristicPartial ();
/** Constructor with model - assumed before cuts
Fixes all variables with priority <= given
and does given number of nodes
*/
CbcHeuristicPartial (CbcModel & model, int fixPriority = 10000, int numberNodes = 200);
// Copy constructor
CbcHeuristicPartial ( const CbcHeuristicPartial &);
// Destructor
~CbcHeuristicPartial ();
/// Assignment operator
CbcHeuristicPartial & operator=(const CbcHeuristicPartial& rhs);
/// Clone
virtual CbcHeuristic * clone() const;
/// Create C++ lines to get to current state
virtual void generateCpp( FILE * fp) ;
/// Resets stuff if model changes
virtual void resetModel(CbcModel * model);
/// update model (This is needed if cliques update matrix etc)
virtual void setModel(CbcModel * model);
using CbcHeuristic::solution ;
/** returns 0 if no solution, 1 if valid solution
with better objective value than one passed in
Sets solution values if good, sets objective value (only if good)
This is called after cuts have been added - so can not add cuts
*/
virtual int solution(double & objectiveValue,
double * newSolution);
/// Validate model i.e. sets when_ to 0 if necessary (may be NULL)
virtual void validate();
/// Set priority level
void setFixPriority(int value) {
fixPriority_ = value;
}
/** Check whether the heuristic should run at all */
virtual bool shouldHeurRun(int whereFrom);
protected:
// Data
// All variables with abs priority <= this will be fixed
int fixPriority_;
};
/** heuristic - just picks up any good solution
found by solver - see OsiBabSolver
*/
class CbcSerendipity : public CbcHeuristic {
public:
// Default Constructor
CbcSerendipity ();
/* Constructor with model
*/
CbcSerendipity (CbcModel & model);
// Copy constructor
CbcSerendipity ( const CbcSerendipity &);
// Destructor
~CbcSerendipity ();
/// Assignment operator
CbcSerendipity & operator=(const CbcSerendipity& rhs);
/// Clone
virtual CbcHeuristic * clone() const;
/// Create C++ lines to get to current state
virtual void generateCpp( FILE * fp) ;
/// update model
virtual void setModel(CbcModel * model);
using CbcHeuristic::solution ;
/** returns 0 if no solution, 1 if valid solution.
Sets solution values if good, sets objective value (only if good)
We leave all variables which are at one at this node of the
tree to that value and will
initially set all others to zero. We then sort all variables in order of their cost
divided by the number of entries in rows which are not yet covered. We randomize that
value a bit so that ties will be broken in different ways on different runs of the heuristic.
We then choose the best one and set it to one and repeat the exercise.
*/
virtual int solution(double & objectiveValue,
double * newSolution);
/// Resets stuff if model changes
virtual void resetModel(CbcModel * model);
protected:
};
/** Just One class - this chooses one at random
*/
class CbcHeuristicJustOne : public CbcHeuristic {
public:
// Default Constructor
CbcHeuristicJustOne ();
// Constructor with model - assumed before cuts
CbcHeuristicJustOne (CbcModel & model);
// Copy constructor
CbcHeuristicJustOne ( const CbcHeuristicJustOne &);
// Destructor
~CbcHeuristicJustOne ();
/// Clone
virtual CbcHeuristicJustOne * clone() const;
/// Assignment operator
CbcHeuristicJustOne & operator=(const CbcHeuristicJustOne& rhs);
/// Create C++ lines to get to current state
virtual void generateCpp( FILE * fp) ;
/** returns 0 if no solution, 1 if valid solution
with better objective value than one passed in
Sets solution values if good, sets objective value (only if good)
This is called after cuts have been added - so can not add cuts
This does Fractional Diving
*/
virtual int solution(double & objectiveValue,
double * newSolution);
/// Resets stuff if model changes
virtual void resetModel(CbcModel * model);
/// update model (This is needed if cliques update matrix etc)
virtual void setModel(CbcModel * model);
/// Selects the next variable to branch on
/** Returns true if all the fractional variables can be trivially
rounded. Returns false, if there is at least one fractional variable
that is not trivially roundable. In this case, the bestColumn
returned will not be trivially roundable.
This is dummy as never called
*/
virtual bool selectVariableToBranch(OsiSolverInterface* /*solver*/,
const double* /*newSolution*/,
int& /*bestColumn*/,
int& /*bestRound*/) {
return true;
}
/// Validate model i.e. sets when_ to 0 if necessary (may be NULL)
virtual void validate();
/// Adds an heuristic with probability
void addHeuristic(const CbcHeuristic * heuristic, double probability);
/// Normalize probabilities
void normalizeProbabilities();
protected:
// Data
// Probability of running a heuristic
double * probabilities_;
// Heuristics
CbcHeuristic ** heuristic_;
// Number of heuristics
int numberHeuristics_;
};
#endif