libncbm.cpp

浏览该文件的文档.

/*
 * This program is free software; you can redistribute it and/or modify
 * it under the terms of the GNU General Public License as published by
 * the Free Software Foundation; either version 3 of the License, or
 * (at your option) any later version.
 *
 * Copyright (C) 2012 Viktor Gal
 *
 */

#include <shogun/structure/libncbm.h>
#include <shogun/lib/Time.h>
#include <shogun/mathematics/Math.h>

#include <shogun/lib/external/libqp.h>
#include <shogun/multiclass/GMNPLib.h>

#include <vector>

namespace shogun
{

static float64_t* HMatrix;
static uint32_t maxCPs;
static const float64_t epsilon=0.0;

static const float64_t *get_col(uint32_t i)
{
    return (&HMatrix[maxCPs*i]);
}

IGNORE_IN_CLASSLIST struct line_search_res
{
    /* */
    float64_t a;
    float64_t fval;
    float64_t reg;
    SGVector<float64_t> solution;
    SGVector<float64_t> gradient;
};

inline static line_search_res zoom
(
 CDualLibQPBMSOSVM *machine,
 float64_t lambda,
 float64_t a_lo,
 float64_t a_hi,
 float64_t initial_fval,
 SGVector<float64_t>& initial_solution,
 SGVector<float64_t>& search_dir,
 float64_t wolfe_c1,
 float64_t wolfe_c2,
 float64_t init_lgrad,
 float64_t f_lo,
 float64_t g_lo,
 float64_t f_hi,
 float64_t g_hi
)
{
    line_search_res ls_res;
    ls_res.solution.resize_vector(initial_solution.vlen);
    ls_res.gradient.resize_vector(initial_solution.vlen);

    SGVector<float64_t> cur_solution(initial_solution.vlen);
    cur_solution.zero();
    SGVector<float64_t> cur_grad(initial_solution.vlen);

    uint32_t iter = 0;
    while (1)
    {
        float64_t d1 = g_lo+g_hi - 3*(f_lo-f_hi)/(a_lo-a_hi);
        float64_t d2 = CMath::sqrt(d1*d1 - g_lo*g_hi);
        float64_t a_j = a_hi -(a_hi-a_lo)*(g_hi+d2-d1)/(g_hi-g_lo+2*d2);

        if (a_lo < a_hi)
        {
            if ((a_j < a_lo) || (a_j > a_hi))
            {
                a_j = 0.5*(a_lo+a_hi);
            }
        }
        else
        {
            if ((a_j > a_lo) || (a_j < a_hi))
            {
                a_j = 0.5*(a_lo+a_hi);
            }
        }

        cur_solution.add(cur_solution.vector, 1.0,
                initial_solution.vector, a_j,
                search_dir.vector, cur_solution.vlen);

        float64_t cur_fval = machine->risk(cur_grad.vector, cur_solution.vector);
        float64_t cur_reg
            = 0.5*lambda*CMath::dot(cur_solution.vector,
                    cur_solution.vector, cur_solution.vlen);
        cur_fval += cur_reg;

        cur_grad.vec1_plus_scalar_times_vec2(cur_grad.vector, lambda, cur_solution.vector, cur_grad.vlen);

        if
            (
             (cur_fval > (initial_fval+wolfe_c1*a_j*init_lgrad))
             ||
             (cur_fval > f_lo)
            )
        {
            a_hi = a_j;
            f_hi = cur_fval;
            g_hi = CMath::dot(cur_grad.vector, search_dir.vector, cur_grad.vlen);
        }
        else
        {
            float64_t cur_lgrad
                = CMath::dot(cur_grad.vector, search_dir.vector, cur_grad.vlen);

            if (CMath::abs(cur_lgrad) < -wolfe_c2*init_lgrad)
            {
                ls_res.a = a_j;
                ls_res.fval = cur_fval;
                ls_res.reg = cur_reg;
                ls_res.gradient = cur_grad;
                ls_res.solution = cur_solution;
//              SG_SPRINT("in zoom (wolfe2): %f\n", cur_fval)
                return ls_res;
            }

            if (cur_lgrad*(a_hi - a_lo) >= 0)
            {
                a_hi = a_lo;
                f_hi = f_lo;
                g_hi = g_lo;
            }
            a_lo = a_j;
            f_lo = cur_fval;
            g_lo = cur_lgrad;
        }

        if
            (
             (CMath::abs(a_lo - a_hi) <= 0.01*a_lo)
             ||
             (iter >= 5)
            )
        {
            ls_res.a = a_j;
            ls_res.fval = cur_fval;
            ls_res.reg = cur_reg;
            ls_res.gradient = cur_grad;
            ls_res.solution = cur_solution;
//          SG_SPRINT("in zoom iter: %d %f\n", iter, cur_fval)
            return ls_res;
        }
        iter++;
    }
}

inline std::vector<line_search_res> line_search_with_strong_wolfe
(
        CDualLibQPBMSOSVM *machine,
        float64_t lambda,
        float64_t initial_val,
        SGVector<float64_t>& initial_solution,
        SGVector<float64_t>& initial_grad,
        SGVector<float64_t>& search_dir,
        float64_t astart,
        float64_t amax = 1.1,
        float64_t wolfe_c1 = 1E-4,
        float64_t wolfe_c2 = 0.9,
        float64_t max_iter = 5
)
{
    /* NOTE:
     * model->risk returns only the risk as it's name says
     * have to cur_fval = model.risk + (lambda*0.5*w*w')
     *
     * subgrad should be: subgrad + (lambda*w)
     *
     */

    uint32_t iter = 0;

    initial_grad.vec1_plus_scalar_times_vec2(initial_grad.vector, lambda, initial_solution.vector, initial_grad.vlen);

    float64_t initial_lgrad = CMath::dot(initial_grad.vector, search_dir.vector, initial_grad.vlen);
    float64_t prev_lgrad = initial_lgrad;
    float64_t prev_fval = initial_val;

    float64_t prev_a = 0;
    float64_t cur_a = astart;

    std::vector<line_search_res> ret;
    while (1)
    {
        SGVector<float64_t> x(initial_solution.vlen);
        SGVector<float64_t> cur_subgrad(initial_solution.vlen);

        x.add(x.vector, 1.0, initial_solution.vector, cur_a, search_dir.vector, x.vlen);
        float64_t cur_fval = machine->risk(cur_subgrad.vector, x.vector);
        float64_t cur_reg = 0.5*lambda*CMath::dot(x.vector, x.vector, x.vlen);
        cur_fval += cur_reg;

        cur_subgrad.vec1_plus_scalar_times_vec2(cur_subgrad.vector, lambda, x.vector, x.vlen);
        if (iter == 0)
        {
            line_search_res initial_step;
            initial_step.fval = cur_fval;
            initial_step.reg = cur_reg;
            initial_step.gradient = cur_subgrad;
            initial_step.solution = x;
            ret.push_back(initial_step);
        }

        float64_t cur_lgrad
            = CMath::dot(cur_subgrad.vector, search_dir.vector,
                    cur_subgrad.vlen);
        if
            (
             (cur_fval > initial_val+wolfe_c1*cur_a*initial_lgrad)
             ||
             (cur_fval >= prev_fval && iter > 0)
            )
        {
            ret.push_back(
                    zoom(machine, lambda, prev_a, cur_a, initial_val,
                        initial_solution, search_dir, wolfe_c1, wolfe_c2,
                        initial_lgrad, prev_fval, prev_lgrad, cur_fval, cur_lgrad)
                    );
            return ret;
        }

        if (CMath::abs(cur_lgrad) <= -wolfe_c2*initial_lgrad)
        {
            line_search_res ls_res;
            ls_res.a = cur_a;
            ls_res.fval = cur_fval;
            ls_res.reg = cur_reg;
            ls_res.solution = x;
            ls_res.gradient = cur_subgrad;
            ret.push_back(ls_res);
            return ret;
        }

        if (cur_lgrad >= 0)
        {
            ret.push_back(
                    zoom(machine, lambda, cur_a, prev_a, initial_val,
                        initial_solution, search_dir, wolfe_c1, wolfe_c2,
                        initial_lgrad, cur_fval, cur_lgrad, prev_fval, prev_lgrad)
                    );
            return ret;
        }
        iter++;
        if ((CMath::abs(cur_a - amax) <= 0.01*amax) || (iter >= max_iter))
        {
            line_search_res ls_res;
            ls_res.a = cur_a;
            ls_res.fval = cur_fval;
            ls_res.reg = cur_reg;
            ls_res.solution = x;
            ls_res.gradient = cur_subgrad;
            ret.push_back(ls_res);
            return ret;
        }

        prev_a = cur_a;
        prev_fval = cur_fval;
        prev_lgrad = cur_lgrad;

        cur_a = (cur_a + amax)*0.5;
    }
}

inline void update_H(BmrmStatistics& ncbm,
        bmrm_ll* head,
        bmrm_ll* tail,
        SGMatrix<float64_t>& H,
        SGVector<float64_t>& diag_H,
        float64_t lambda,
        uint32_t maxCP,
        int32_t w_dim)
{
    float64_t* a_2 = get_cutting_plane(tail);
    bmrm_ll* cp_ptr=head;

    for (uint32_t i=0; i < ncbm.nCP; ++i)
    {
        float64_t* a_1 = get_cutting_plane(cp_ptr);
        cp_ptr=cp_ptr->next;

        float64_t dot_val = CMath::dot(a_2, a_1, w_dim);

        H.matrix[LIBBMRM_INDEX(ncbm.nCP, i, maxCP)]
            = H.matrix[LIBBMRM_INDEX(i, ncbm.nCP, maxCP)]
            = dot_val/lambda;
    }

    /* set the diagonal element, i.e. subgrad_i*subgrad_i' */
    float64_t dot_val = CMath::dot(a_2, a_2, w_dim);
    H[LIBBMRM_INDEX(ncbm.nCP, ncbm.nCP, maxCP)]=dot_val/lambda;

    diag_H[ncbm.nCP]=H[LIBBMRM_INDEX(ncbm.nCP, ncbm.nCP, maxCP)];

    ncbm.nCP++;
}


BmrmStatistics svm_ncbm_solver(
        CDualLibQPBMSOSVM *machine,
        float64_t         *w,
        float64_t         TolRel,
        float64_t         TolAbs,
        float64_t         _lambda,
        uint32_t          _BufSize,
        bool              cleanICP,
        uint32_t          cleanAfter,
        bool              is_convex,
        bool              line_search,
        bool              verbose
        )
{
    BmrmStatistics ncbm;
    libqp_state_T qp_exitflag={0, 0, 0, 0};

    CStructuredModel* model = machine->get_model();
    int32_t w_dim = model->get_dim();

    maxCPs = _BufSize;
    BufSize = _BufSize;

    ncbm.nCP=0;
    ncbm.nIter=0;
    ncbm.exitflag=0;

    /* variables for timing the algorithm*/
    CTime ttime;
    float64_t tstart, tstop;
    tstart=ttime.cur_time_diff(false);

    /* matrix of subgradiends */
    SGMatrix<float64_t> A(w_dim, maxCPs);

    /* bias vector */
    SGVector<float64_t> bias(maxCPs);
    bias.zero();

    /* Matrix for storing H = A*A' */
    SGMatrix<float64_t> H(maxCPs,maxCPs);
    HMatrix = H.matrix;

    /* diag_H */
    SGVector<float64_t> diag_H(maxCPs);
    diag_H.zero();

    /* x the solution vector of the dual problem:
     * 1/lambda x*H*x' - x*bias
     */
    SGVector<float64_t> x(maxCPs);
    x.zero();

    /* for sum_{i in I_k} x[i] <= b[k] for all k such that S[k] == 1 */
    float64_t b = 1.0;
    uint8_t S = 1;
    SGVector<uint32_t> I(maxCPs);
    I.set_const(1);

    /* libqp paramteres */
    uint32_t QPSolverMaxIter = 0xFFFFFFFF;
    float64_t QPSolverTolRel = 1E-9;

    /* variables for maintaining inactive planes */
    SGVector<bool> map(maxCPs);
    map.set_const(true);
    ICP_stats icp_stats;
    icp_stats.maxCPs = maxCPs;
    icp_stats.ICPcounter = (uint32_t*) LIBBMRM_CALLOC(maxCPs, uint32_t);
    icp_stats.ICPs = (float64_t**) LIBBMRM_CALLOC(maxCPs, float64_t*);
    icp_stats.ACPs = (uint32_t*) LIBBMRM_CALLOC(maxCPs, uint32_t);
    icp_stats.H_buff = (float64_t*) LIBBMRM_CALLOC(maxCPs*maxCPs,float64_t);
    if
    (
        icp_stats.ICPcounter == NULL || icp_stats.ICPs == NULL
        || icp_stats.ACPs == NULL || icp_stats.H_buff == NULL
    )
    {
        ncbm.exitflag=-2;
        LIBBMRM_FREE(icp_stats.ICPcounter);
        LIBBMRM_FREE(icp_stats.ICPs);
        LIBBMRM_FREE(icp_stats.ACPs);
        LIBBMRM_FREE(icp_stats.H_buff);
        return ncbm;
    }

    /* best */
    float64_t best_Fp = CMath::INFTY;
    float64_t best_risk = CMath::INFTY;
    SGVector<float64_t> best_w(w_dim);
    best_w.zero();
    SGVector<float64_t> best_subgrad(w_dim);
    best_subgrad.zero();

    /* initial solution */
    SGVector<float64_t> cur_subgrad(w_dim);
    SGVector<float64_t> cur_w(w_dim);
    memcpy(cur_w.vector, w, sizeof(float64_t)*w_dim);

    float64_t cur_risk = machine->risk(cur_subgrad.vector, cur_w.vector);
    bias[0] = -cur_risk;
    best_Fp = 0.5*_lambda*CMath::dot(cur_w.vector, cur_w.vector, cur_w.vlen) + cur_risk;
    best_risk = cur_risk;
    memcpy(best_w.vector, cur_w.vector, sizeof(float64_t)*w_dim);
    memcpy(best_subgrad.vector, cur_subgrad.vector, sizeof(float64_t)*w_dim);

    /* create a double-linked list over the A the subgrad matrix */
    bmrm_ll *CPList_head, *CPList_tail, *cp_ptr, *cp_list=NULL;
    cp_list = (bmrm_ll*) SG_CALLOC(bmrm_ll, 1);
    if (cp_list==NULL)
    {
        ncbm.exitflag=-2;
        return ncbm;
    }
    /* save the subgradient */
    memcpy(A.matrix, cur_subgrad.vector, sizeof(float64_t)*w_dim);
    map[0] = false;
    cp_list->address=&A[0];
    cp_list->idx=0;
    cp_list->prev=NULL;
    cp_list->next=NULL;
    CPList_head=cp_list;
    CPList_tail=cp_list;

    update_H(ncbm, CPList_head, CPList_tail, H, diag_H, _lambda, maxCPs, w_dim);
    tstop=ttime.cur_time_diff(false);
    if (verbose)
        SG_SPRINT("%4d: tim=%.3lf, Fp=%lf, Fd=%lf, R=%lf\n",
                ncbm.nIter, tstop-tstart, ncbm.Fp, ncbm.Fd, cur_risk);

    float64_t astar = 0.01;

    SG_SPRINT("clean icps: %d\n", cleanICP)
    while (ncbm.exitflag==0)
    {
        tstart=ttime.cur_time_diff(false);
        ncbm.nIter++;

        //diag_H.display_vector();
        //bias.display_vector();

        /* solve the dual of the problem, namely:
         *
         */
        qp_exitflag =
            libqp_splx_solver(&get_col, diag_H.vector, bias.vector, &b, I.vector, &S, x.vector,
                    ncbm.nCP, QPSolverMaxIter, 0.0, QPSolverTolRel, -LIBBMRM_PLUS_INF, 0);

        ncbm.Fd = -qp_exitflag.QP;

        ncbm.qp_exitflag=qp_exitflag.exitflag;

        /* Update ICPcounter (add one to unused and reset used)
         * + compute number of active CPs */
        ncbm.nzA=0;

        for (uint32_t i=0; i < ncbm.nCP; ++i)
        {
            if (x[i] > epsilon)
            {
                /* cp was active => reset counter */
                ++ncbm.nzA;
                icp_stats.ICPcounter[i]=0;
            }
            else
            {
                icp_stats.ICPcounter[i]++;
            }
        }

        /* Inactive Cutting Planes (ICP) removal */
        if (cleanICP)
        {
            clean_icp(&icp_stats, ncbm, &CPList_head, &CPList_tail,
                    H.matrix, diag_H.vector, x.vector,
                    map.vector, cleanAfter, bias.vector, I.vector);
        }

        /* calculate the new w
         * w[i] = -1/lambda*A[i]*x[i]
         */
        cur_w.zero();
        cp_ptr=CPList_head;
        for (uint32_t i=0; i < ncbm.nCP; ++i)
        {
            float64_t* A_1 = get_cutting_plane(cp_ptr);
            cp_ptr=cp_ptr->next;
            SGVector<float64_t>::vec1_plus_scalar_times_vec2(cur_w.vector, -x[i]/_lambda, A_1, w_dim);
        }

        bool calc_gap = false;
        if (calc_gap)
        {
            SGVector<float64_t> scores(ncbm.nCP);
            cp_ptr=CPList_head;

            for (uint32_t i=0; i < ncbm.nCP; ++i)
            {
                float64_t* a_1 = get_cutting_plane(cp_ptr);
                cp_ptr = cp_ptr->next;
                scores[i] = CMath::dot(cur_w.vector, a_1, w_dim);
            }
            scores.vec1_plus_scalar_times_vec2(scores.vector, -1.0, bias.vector, scores.vlen);

            float64_t w_norm = CMath::dot(cur_w.vector, cur_w.vector, cur_w.vlen);
            float64_t PO = 0.5*_lambda*w_norm + CMath::max(scores.vector, scores.vlen);
            float64_t QP_gap = PO - ncbm.Fd;

            SG_SPRINT("%4d: primal:%f dual:%f QP_gap:%f\n", ncbm.nIter, PO, ncbm.Fd, QP_gap)
        }

        /* Stopping conditions */
        if ((best_Fp - ncbm.Fd) <= TolRel*LIBBMRM_ABS(best_Fp))
            ncbm.exitflag = 1;

        if ((best_Fp - ncbm.Fd) <= TolAbs)
            ncbm.exitflag = 2;

        if (ncbm.nCP >= maxCPs)
            ncbm.exitflag = -1;

        // next CP would exceed BufSize/maxCPs
        if (ncbm.nCP+3 >= maxCPs)
            ncbm.exitflag=-1;

        tstop=ttime.cur_time_diff(false);

        /* Verbose output */
        if (verbose)
            SG_SPRINT("%4d: tim=%.3lf, Fp=%lf, Fd=%lf, (Fp-Fd)=%lf, (Fp-Fd)/Fp=%lf, R=%lf, nCP=%d, nzA=%d, QPexitflag=%d, best_fp=%f, gap=%f\n",
                    ncbm.nIter, tstop-tstart, ncbm.Fp, ncbm.Fd, ncbm.Fp-ncbm.Fd,
                    (ncbm.Fp-ncbm.Fd)/ncbm.Fp, cur_risk, ncbm.nCP, ncbm.nzA, qp_exitflag.exitflag, best_Fp, (best_Fp-ncbm.Fd)/best_Fp);

        if (ncbm.exitflag!=0)
            break;

        std::vector<line_search_res> wbest_candidates;
        if (!line_search)
        {
            cur_risk = machine->risk(cur_subgrad.vector, cur_w.vector);

            add_cutting_plane(&CPList_tail, map, A.matrix,
                    find_free_idx(map, maxCPs), cur_subgrad.vector, w_dim);

            bias[ncbm.nCP] = CMath::dot(cur_w.vector, cur_subgrad.vector, cur_w.vlen) - cur_risk;

            update_H(ncbm, CPList_head, CPList_tail, H, diag_H, _lambda, maxCPs, w_dim);

            // add as a new wbest candidate
            line_search_res ls;
            ls.fval = cur_risk+0.5*_lambda*CMath::dot(cur_w.vector, cur_w.vector, cur_w.vlen);
            ls.solution = cur_w;
            ls.gradient = cur_subgrad;

            wbest_candidates.push_back(ls);
        }
        else
        {
            tstart=ttime.cur_time_diff(false);
            /* do line searching */
            SGVector<float64_t> search_dir(w_dim);
            search_dir.add(search_dir.vector, 1.0, cur_w.vector, -1.0, best_w.vector, w_dim);

            float64_t norm_dir = search_dir.twonorm(search_dir.vector, search_dir.vlen);
            float64_t astart;
            uint32_t cp_min_approx = 0;
            if (cp_min_approx || (ncbm.nIter == 1))
            {
                astart = 1.0;
            }
            else
            {
                astart = CMath::min(astar/norm_dir,1.0);
                if (astart == 0)
                    astart = 1.0;
            }

            /* line search */
            std::vector<line_search_res> ls_res
                = line_search_with_strong_wolfe(machine, _lambda, best_Fp, best_w, best_subgrad, search_dir, astart);

            if (ls_res[0].fval != ls_res[1].fval)
            {
                ls_res[0].gradient.vec1_plus_scalar_times_vec2(ls_res[0].gradient.vector, -_lambda, ls_res[0].solution.vector, w_dim);

                add_cutting_plane(&CPList_tail, map, A.matrix,
                        find_free_idx(map, maxCPs), ls_res[0].gradient, w_dim);

                bias[ncbm.nCP]
                    = CMath::dot(ls_res[0].solution.vector, ls_res[0].gradient, w_dim)
                    - (ls_res[0].fval - ls_res[0].reg);

                update_H(ncbm, CPList_head, CPList_tail, H, diag_H, _lambda, maxCPs, w_dim);

                wbest_candidates.push_back(ls_res[0]);
            }

            ls_res[1].gradient.vec1_plus_scalar_times_vec2(ls_res[1].gradient.vector, -_lambda, ls_res[1].solution.vector, w_dim);

            add_cutting_plane(&CPList_tail, map, A.matrix,
                    find_free_idx(map, maxCPs), ls_res[1].gradient.vector, w_dim);

            bias[ncbm.nCP]
                = CMath::dot(ls_res[1].solution.vector, ls_res[1].gradient.vector, w_dim)
                - (ls_res[1].fval - ls_res[1].reg);

            update_H(ncbm, CPList_head, CPList_tail, H, diag_H, _lambda, maxCPs, w_dim);

            wbest_candidates.push_back(ls_res[1]);

            if ((best_Fp <= ls_res[1].fval) && (astart != 1))
            {
                cur_risk = machine->risk(cur_subgrad.vector, cur_w.vector);

                add_cutting_plane(&CPList_tail, map, A.matrix,
                            find_free_idx(map, maxCPs), cur_subgrad.vector, w_dim);

                bias[ncbm.nCP]
                    =  CMath::dot(cur_w.vector, cur_subgrad.vector, cur_w.vlen) - cur_risk;

                update_H(ncbm, CPList_head, CPList_tail, H, diag_H, _lambda, maxCPs, w_dim);

                /* add as a new wbest candidate */
                line_search_res ls;
                ls.fval = cur_risk+0.5*_lambda*CMath::dot(cur_w.vector, cur_w.vector, cur_w.vlen);
                ls.solution = cur_w;
                ls.gradient = cur_subgrad;
                SG_SPRINT("%lf\n", ls.fval)

                wbest_candidates.push_back(ls);
            }

            astar = ls_res[1].a * norm_dir;

            tstop=ttime.cur_time_diff(false);
            SG_SPRINT("\t\tline search time: %.5lf\n", tstop-tstart)
        }

        /* search for the best w among the new candidates */
        if (verbose)
            SG_SPRINT("\t searching for the best Fp:\n")
        for (size_t i = 0; i < wbest_candidates.size(); i++)
        {
            if (verbose)
                SG_SPRINT("\t\t %d fcurrent: %.16lf\n", i, wbest_candidates[i].fval)

            if (wbest_candidates[i].fval < best_Fp)
            {
                best_Fp = wbest_candidates[i].fval;
                best_risk = wbest_candidates[i].fval - wbest_candidates[i].reg;
                memcpy(best_w, wbest_candidates[i].solution.vector, sizeof(float64_t)*w_dim);
                memcpy(best_subgrad.vector, wbest_candidates[i].gradient.vector, sizeof(float64_t)*w_dim);

                ncbm.Fp = best_Fp;

                if (verbose)
                    SG_SPRINT("\t\t new best norm: %f\n",
                            best_w.twonorm(best_w.vector, w_dim));
            }

            if (!is_convex)
            {
                index_t cp_idx = ncbm.nCP-(wbest_candidates.size()-i);

                /* conflict */
                float64_t score
                    = CMath::dot(best_w.vector,
                            wbest_candidates[i].gradient.vector, w_dim)
                    + (-1.0*bias[cp_idx]);
                if (score > best_risk)
                {
                    float64_t U
                        = best_risk
                        - CMath::dot(best_w.vector,
                                wbest_candidates[i].gradient.vector, w_dim);

                    float64_t L
                        = best_Fp - wbest_candidates[i].reg
                        - CMath::dot(wbest_candidates[i].solution.vector,
                                wbest_candidates[i].gradient.vector, w_dim);

                    if (verbose)
                        SG_SPRINT("CONFLICT Rbest=%.6lg score=%g L=%.6lg U=%.6lg\n", best_risk, score, L, U)
                    if (L <= U)
                    {
                        if (verbose)
                            SG_SPRINT("%.6lf < %.6lf => changing bias[%d]=%g\n", L, U, cp_idx, L)
                        bias[cp_idx]= -L;
                    }
                    else
                    {
                        wbest_candidates[i].gradient.zero();
                        SGVector<float64_t>::vec1_plus_scalar_times_vec2(wbest_candidates[i].gradient.vector, -_lambda, best_w.vector, w_dim);

                        cp_ptr = CPList_tail;
                        for (size_t j = wbest_candidates.size()-1; i < j; --j)
                        {
                            cp_ptr = cp_ptr->prev;
                            SG_SPRINT("tail - %d\n (%d)", j, i)
                        }

                        float64_t* cp = get_cutting_plane(cp_ptr);
                        LIBBMRM_MEMCPY(cp, wbest_candidates[i].gradient.vector, w_dim*sizeof(float64_t));

                        /* update the corresponding column and row in H */
                        cp_ptr = CPList_head;
                        for (uint32_t j = 0; j < ncbm.nCP-1; ++j)
                        {
                            float64_t* a = get_cutting_plane(cp_ptr);
                            cp_ptr = cp_ptr->next;
                            float64_t dot_val
                                = CMath::dot(a, wbest_candidates[i].gradient.vector, w_dim);

                            H.matrix[LIBBMRM_INDEX(cp_idx, j, maxCPs)]
                                = H.matrix[LIBBMRM_INDEX(j, cp_idx, maxCPs)]
                                = dot_val/_lambda;
                        }

                        diag_H[LIBBMRM_INDEX(cp_idx, cp_idx, maxCPs)]
                            = CMath::dot(wbest_candidates[i].gradient.vector,
                                    wbest_candidates[i].gradient.vector, w_dim);


                        bias[cp_idx]
                            = best_Fp - wbest_candidates[i].reg
                            - CMath::dot(wbest_candidates[i].solution.vector,
                                    wbest_candidates[i].gradient.vector, w_dim);

                        if (verbose)
                            SG_SPRINT("solved by changing nCP=%d bias:%g (%g)\n", cp_idx, bias[cp_idx], L)
                    }
                }
            }
        }

        /* Inactive Cutting Planes (ICP) removal
        if (cleanICP)
        {
            clean_icp(&icp_stats, ncbm, &CPList_head, &CPList_tail,
                    H.matrix, diag_H.vector, x.vector,
                    map.vector, cleanAfter, bias.vector, I.vector);
        }
        */
    }

    memcpy(w, best_w.vector, sizeof(float64_t)*w_dim);

    /* free ICP_stats variables */
    LIBBMRM_FREE(icp_stats.ICPcounter);
    LIBBMRM_FREE(icp_stats.ICPs);
    LIBBMRM_FREE(icp_stats.ACPs);
    LIBBMRM_FREE(icp_stats.H_buff);

        cp_ptr=CPList_head;
        while(cp_ptr!=NULL)
        {
                bmrm_ll * cp_ptr_this=cp_ptr;
                cp_ptr=cp_ptr->next;
                LIBBMRM_FREE(cp_ptr_this);
                cp_ptr_this=NULL;
        }

    SG_UNREF(model);

    return ncbm;
}

} /* namespace shogun */