/* $Id: ClpSimplexDual.hpp 1761 2011-07-06 16:06:24Z forrest $ */ // 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). /* Authors John Forrest */ #ifndef ClpSimplexDual_H #define ClpSimplexDual_H #include "ClpSimplex.hpp" /** This solves LPs using the dual simplex method It inherits from ClpSimplex. It has no data of its own and is never created - only cast from a ClpSimplex object at algorithm time. */ class ClpSimplexDual : public ClpSimplex { public: /**@name Description of algorithm */ //@{ /** Dual algorithm Method It tries to be a single phase approach with a weight of 1.0 being given to getting optimal and a weight of updatedDualBound_ being given to getting dual feasible. In this version I have used the idea that this weight can be thought of as a fake bound. If the distance between the lower and upper bounds on a variable is less than the feasibility weight then we are always better off flipping to other bound to make dual feasible. If the distance is greater then we make up a fake bound updatedDualBound_ away from one bound. If we end up optimal or primal infeasible, we check to see if bounds okay. If so we have finished, if not we increase updatedDualBound_ and continue (after checking if unbounded). I am undecided about free variables - there is coding but I am not sure about it. At present I put them in basis anyway. The code is designed to take advantage of sparsity so arrays are seldom zeroed out from scratch or gone over in their entirety. The only exception is a full scan to find outgoing variable for Dantzig row choice. For steepest edge we keep an updated list of infeasibilities (actually squares). On easy problems we don't need full scan - just pick first reasonable. One problem is how to tackle degeneracy and accuracy. At present I am using the modification of costs which I put in OSL and some of what I think is the dual analog of Gill et al. I am still not sure of the exact details. The flow of dual is three while loops as follows: while (not finished) { while (not clean solution) { Factorize and/or clean up solution by flipping variables so dual feasible. If looks finished check fake dual bounds. Repeat until status is iterating (-1) or finished (0,1,2) } while (status==-1) { Iterate until no pivot in or out or time to re-factorize. Flow is: choose pivot row (outgoing variable). if none then we are primal feasible so looks as if done but we need to break and check bounds etc. Get pivot row in tableau Choose incoming column. If we don't find one then we look primal infeasible so break and check bounds etc. (Also the pivot tolerance is larger after any iterations so that may be reason) If we do find incoming column, we may have to adjust costs to keep going forwards (anti-degeneracy). Check pivot will be stable and if unstable throw away iteration and break to re-factorize. If minor error re-factorize after iteration. Update everything (this may involve flipping variables to stay dual feasible. } } TODO's (or maybe not) At present we never check we are going forwards. I overdid that in OSL so will try and make a last resort. Needs partial scan pivot out option. May need other anti-degeneracy measures, especially if we try and use loose tolerances as a way to solve in fewer iterations. I like idea of dynamic scaling. This gives opportunity to decouple different implications of scaling for accuracy, iteration count and feasibility tolerance. for use of exotic parameter startFinishoptions see Clpsimplex.hpp */ int dual(int ifValuesPass, int startFinishOptions = 0); /** For strong branching. On input lower and upper are new bounds while on output they are change in objective function values (>1.0e50 infeasible). Return code is 0 if nothing interesting, -1 if infeasible both ways and +1 if infeasible one way (check values to see which one(s)) Solutions are filled in as well - even down, odd up - also status and number of iterations */ int strongBranching(int numberVariables, const int * variables, double * newLower, double * newUpper, double ** outputSolution, int * outputStatus, int * outputIterations, bool stopOnFirstInfeasible = true, bool alwaysFinish = false, int startFinishOptions = 0); /// This does first part of StrongBranching ClpFactorization * setupForStrongBranching(char * arrays, int numberRows, int numberColumns, bool solveLp = false); /// This cleans up after strong branching void cleanupAfterStrongBranching(ClpFactorization * factorization); //@} /**@name Functions used in dual */ //@{ /** This has the flow between re-factorizations Broken out for clarity and will be used by strong branching Reasons to come out: -1 iterations etc -2 inaccuracy -3 slight inaccuracy (and done iterations) +0 looks optimal (might be unbounded - but we will investigate) +1 looks infeasible +3 max iterations If givenPi not NULL then in values pass */ int whileIterating(double * & givenPi, int ifValuesPass); /** The duals are updated by the given arrays. Returns number of infeasibilities. After rowArray and columnArray will just have those which have been flipped. Variables may be flipped between bounds to stay dual feasible. The output vector has movement of primal solution (row length array) */ int updateDualsInDual(CoinIndexedVector * rowArray, CoinIndexedVector * columnArray, CoinIndexedVector * outputArray, double theta, double & objectiveChange, bool fullRecompute); /** The duals are updated by the given arrays. This is in values pass - so no changes to primal is made */ void updateDualsInValuesPass(CoinIndexedVector * rowArray, CoinIndexedVector * columnArray, double theta); /** While updateDualsInDual sees what effect is of flip this does actual flipping. */ void flipBounds(CoinIndexedVector * rowArray, CoinIndexedVector * columnArray); /** Row array has row part of pivot row Column array has column part. This chooses pivot column. Spare arrays are used to save pivots which will go infeasible We will check for basic so spare array will never overflow. If necessary will modify costs For speed, we may need to go to a bucket approach when many variables are being flipped. Returns best possible pivot value */ double dualColumn(CoinIndexedVector * rowArray, CoinIndexedVector * columnArray, CoinIndexedVector * spareArray, CoinIndexedVector * spareArray2, double accpetablePivot, CoinBigIndex * dubiousWeights); /// Does first bit of dualColumn int dualColumn0(const CoinIndexedVector * rowArray, const CoinIndexedVector * columnArray, CoinIndexedVector * spareArray, double acceptablePivot, double & upperReturn, double &bestReturn, double & badFree); /** Row array has row part of pivot row Column array has column part. This sees what is best thing to do in dual values pass if sequenceIn==sequenceOut can change dual on chosen row and leave variable in basis */ void checkPossibleValuesMove(CoinIndexedVector * rowArray, CoinIndexedVector * columnArray, double acceptablePivot); /** Row array has row part of pivot row Column array has column part. This sees what is best thing to do in branch and bound cleanup If sequenceIn_ < 0 then can't do anything */ void checkPossibleCleanup(CoinIndexedVector * rowArray, CoinIndexedVector * columnArray, double acceptablePivot); /** This sees if we can move duals in dual values pass. This is done before any pivoting */ void doEasyOnesInValuesPass(double * givenReducedCosts); /** Chooses dual pivot row Would be faster with separate region to scan and will have this (with square of infeasibility) when steepest For easy problems we can just choose one of the first rows we look at If alreadyChosen >=0 then in values pass and that row has been selected */ void dualRow(int alreadyChosen); /** Checks if any fake bounds active - if so returns number and modifies updatedDualBound_ and everything. Free variables will be left as free Returns number of bounds changed if >=0 Returns -1 if not initialize and no effect Fills in changeVector which can be used to see if unbounded and cost of change vector If 2 sets to original (just changed) */ int changeBounds(int initialize, CoinIndexedVector * outputArray, double & changeCost); /** As changeBounds but just changes new bounds for a single variable. Returns true if change */ bool changeBound( int iSequence); /// Restores bound to original bound void originalBound(int iSequence); /** Checks if tentative optimal actually means unbounded in dual Returns -3 if not, 2 if is unbounded */ int checkUnbounded(CoinIndexedVector * ray, CoinIndexedVector * spare, double changeCost); /** Refactorizes if necessary Checks if finished. Updates status. lastCleaned refers to iteration at which some objective/feasibility cleaning too place. type - 0 initial so set up save arrays etc - 1 normal -if good update save - 2 restoring from saved */ void statusOfProblemInDual(int & lastCleaned, int type, double * givenDjs, ClpDataSave & saveData, int ifValuesPass); /** Perturbs problem (method depends on perturbation()) returns nonzero if should go to dual */ int perturb(); /** Fast iterations. Misses out a lot of initialization. Normally stops on maximum iterations, first re-factorization or tentative optimum. If looks interesting then continues as normal. Returns 0 if finished properly, 1 otherwise. */ int fastDual(bool alwaysFinish = false); /** Checks number of variables at fake bounds. This is used by fastDual so can exit gracefully before end */ int numberAtFakeBound(); /** Pivot in a variable and choose an outgoing one. Assumes dual feasible - will not go through a reduced cost. Returns step length in theta Return codes as before but -1 means no acceptable pivot */ int pivotResultPart1(); /** Get next free , -1 if none */ int nextSuperBasic(); /** Startup part of dual (may be extended to other algorithms) returns 0 if good, 1 if bad */ int startupSolve(int ifValuesPass, double * saveDuals, int startFinishOptions); void finishSolve(int startFinishOptions); void gutsOfDual(int ifValuesPass, double * & saveDuals, int initialStatus, ClpDataSave & saveData); //int dual2(int ifValuesPass,int startFinishOptions=0); void resetFakeBounds(int type); //@} }; #endif