/* $Id: CbcFathomDynamicProgramming.hpp 1573 2011-01-05 01:12:36Z lou $ */
// Copyright (C) 2004, International Business Machines
// Corporation and others.  All Rights Reserved.
// This code is licensed under the terms of the Eclipse Public License (EPL).

#ifndef CbcFathomDynamicProgramming_H
#define CbcFathomDynamicProgramming_H

#include "CbcFathom.hpp"

//#############################################################################
/** FathomDynamicProgramming class.

    The idea is that after some branching the problem will be effectively smaller than
    the original problem and maybe there will be a more specialized technique which can completely
    fathom this branch quickly.

    This is a dynamic programming implementation which is very fast for some
    specialized problems.  It expects small integral rhs, an all integer problem
    and positive integral coefficients. At present it can not do general set covering
    problems just set partitioning.  It can find multiple optima for various rhs
    combinations.

    The main limiting factor is size of state space.  Each 1 rhs doubles the size of the problem.
    2 or 3 rhs quadruples, 4,5,6,7 by 8 etc.
 */

class CbcFathomDynamicProgramming : public CbcFathom {
public:
    // Default Constructor
    CbcFathomDynamicProgramming ();

    // Constructor with model - assumed before cuts
    CbcFathomDynamicProgramming (CbcModel & model);
    // Copy constructor
    CbcFathomDynamicProgramming(const CbcFathomDynamicProgramming & rhs);

    virtual ~CbcFathomDynamicProgramming();

    /// update model (This is needed if cliques update matrix etc)
    virtual void setModel(CbcModel * model);

    /// Clone
    virtual CbcFathom * clone() const;

    /// Resets stuff if model changes
    virtual void resetModel(CbcModel * model);

    /** returns 0 if no fathoming attempted, 1 fully fathomed ,
        2 incomplete search, 3 incomplete search but treat as complete.
        If solution then newSolution will not be NULL and
        will be freed by CbcModel.  It is expected that the solution is better
        than best so far but CbcModel will double check.

        If returns 3 then of course there is no guarantee of global optimum
    */
    virtual int fathom(double *& newSolution);

    /// Maximum size allowed
    inline int maximumSize() const {
        return maximumSizeAllowed_;
    }
    inline void setMaximumSize(int value) {
        maximumSizeAllowed_ = value;
    }
    /// Returns type of algorithm and sets up arrays
    int checkPossible(int allowableSize = 0);
    // set algorithm
    inline void setAlgorithm(int value) {
        algorithm_ = value;
    }
    /** Tries a column
        returns true if was used in making any changes.
    */
    bool tryColumn(int numberElements, const int * rows,
                   const double * coefficients, double cost,
                   int upper = COIN_INT_MAX);
    /// Returns cost array
    inline const double * cost() const {
        return cost_;
    }
    /// Returns back array
    inline const int * back() const {
        return back_;
    }
    /// Gets bit pattern for target result
    inline int target() const {
        return target_;
    }
    /// Sets bit pattern for target result
    inline void setTarget(int value) {
        target_ = value;
    }
private:
    /// Does deleteions
    void gutsOfDelete();

    /** Adds one attempt of one column of type 0,
        returns true if was used in making any changes
    */
    bool addOneColumn0(int numberElements, const int * rows,
                       double cost);
    /** Adds one attempt of one column of type 1,
        returns true if was used in making any changes.
        At present the user has to call it once for each possible value
    */
    bool addOneColumn1(int numberElements, const int * rows,
                       const int * coefficients, double cost);
    /** Adds one attempt of one column of type 1,
        returns true if was used in making any changes.
        At present the user has to call it once for each possible value.
        This version is when there are enough 1 rhs to do faster
    */
    bool addOneColumn1A(int numberElements, const int * rows,
                        const int * coefficients, double cost);
    /// Gets bit pattern from original column
    int bitPattern(int numberElements, const int * rows,
                   const int * coefficients);
    /// Gets bit pattern from original column
    int bitPattern(int numberElements, const int * rows,
                   const double * coefficients);
    /// Fills in original column (dense) from bit pattern - returning number nonzero
    int decodeBitPattern(int bitPattern, int * values, int numberRows);

protected:

    /// Size of states (power of 2 unless just one constraint)
    int size_;
    /** Type - 0 coefficients and rhs all 1,
        1 - coefficients > 1 or rhs > 1
    */
    int type_;
    /// Space for states
    double * cost_;
    /// Which state produced this cheapest one
    int * back_;
    /// Some rows may be satisified so we need a lookup
    int * lookup_;
    /// Space for sorted indices
    int * indices_;
    /// Number of active rows
    int numberActive_;
    /// Maximum size allowed
    int maximumSizeAllowed_;
    /// Start bit for each active row
    int * startBit_;
    /// Number bits for each active row
    int * numberBits_;
    /// Effective rhs
    int * rhs_;
    /// Space for sorted coefficients
    int * coefficients_;
    /// Target pattern
    int target_;
    /// Number of Non 1 rhs
    int numberNonOne_;
    /// Current bit pattern
    int bitPattern_;
    /// Current algorithm
    int algorithm_;
private:

    /// Illegal Assignment operator
    CbcFathomDynamicProgramming & operator=(const CbcFathomDynamicProgramming& rhs);

};

#endif