示例#1
0
文件: kkt.go 项目: sguzwf/algorithm
//    Solution of KKT equations by reduction to a 2 x 2 system, a QR
//    factorization to eliminate the equality constraints, and a dense
//    Cholesky factorization of order n-p.
//
//    Computes the QR factorization
//
//        A' = [Q1, Q2] * [R; 0]
//
//    and returns a function that (1) computes the Cholesky factorization
//
//        Q_2^T * (H + GG^T * W^{-1} * W^{-T} * GG) * Q2 = L * L^T,
//
//    given H, Df, W, where GG = [Df; G], and (2) returns a function for
//    solving
//
//        [ H    A'   GG'    ]   [ ux ]   [ bx ]
//        [ A    0    0      ] * [ uy ] = [ by ].
//        [ GG   0    -W'*W  ]   [ uz ]   [ bz ]
//
//    H is n x n,  A is p x n, Df is mnl x n, G is N x n where
//    N = dims['l'] + sum(dims['q']) + sum( k**2 for k in dims['s'] ).
//
func kktChol(G *matrix.FloatMatrix, dims *sets.DimensionSet, A *matrix.FloatMatrix, mnl int) (kktFactor, error) {

	p, n := A.Size()
	cdim := mnl + dims.Sum("l", "q") + dims.SumSquared("s")
	cdim_pckd := mnl + dims.Sum("l", "q") + dims.SumPacked("s")

	QA := A.Transpose()
	tauA := matrix.FloatZeros(p, 1)
	lapack.Geqrf(QA, tauA)

	Gs := matrix.FloatZeros(cdim, n)
	K := matrix.FloatZeros(n, n)
	bzp := matrix.FloatZeros(cdim_pckd, 1)
	yy := matrix.FloatZeros(p, 1)
	checkpnt.AddMatrixVar("tauA", tauA)
	checkpnt.AddMatrixVar("Gs", Gs)
	checkpnt.AddMatrixVar("K", K)

	factor := func(W *sets.FloatMatrixSet, H, Df *matrix.FloatMatrix) (KKTFunc, error) {
		// Compute
		//
		//     K = [Q1, Q2]' * (H + GG' * W^{-1} * W^{-T} * GG) * [Q1, Q2]
		//
		// and take the Cholesky factorization of the 2,2 block
		//
		//     Q_2' * (H + GG^T * W^{-1} * W^{-T} * GG) * Q2.

		var err error = nil
		minor := 0
		if !checkpnt.MinorEmpty() {
			minor = checkpnt.MinorTop()
		}
		// Gs = W^{-T} * GG in packed storage.
		if mnl > 0 {
			Gs.SetSubMatrix(0, 0, Df)
		}
		Gs.SetSubMatrix(mnl, 0, G)
		checkpnt.Check("00factor_chol", minor)
		scale(Gs, W, true, true)
		pack2(Gs, dims, mnl)
		//checkpnt.Check("10factor_chol", minor)

		// K = [Q1, Q2]' * (H + Gs' * Gs) * [Q1, Q2].
		blas.SyrkFloat(Gs, K, 1.0, 0.0, la.OptTrans, &la.IOpt{"k", cdim_pckd})
		if H != nil {
			K.SetSubMatrix(0, 0, matrix.Plus(H, K.GetSubMatrix(0, 0, H.Rows(), H.Cols())))
		}
		//checkpnt.Check("20factor_chol", minor)
		symm(K, n, 0)
		lapack.Ormqr(QA, tauA, K, la.OptLeft, la.OptTrans)
		lapack.Ormqr(QA, tauA, K, la.OptRight)
		//checkpnt.Check("30factor_chol", minor)

		// Cholesky factorization of 2,2 block of K.
		lapack.Potrf(K, &la.IOpt{"n", n - p}, &la.IOpt{"offseta", p * (n + 1)})
		checkpnt.Check("40factor_chol", minor)

		solve := func(x, y, z *matrix.FloatMatrix) (err error) {
			// Solve
			//
			//     [ 0          A'  GG'*W^{-1} ]   [ ux   ]   [ bx        ]
			//     [ A          0   0          ] * [ uy   ] = [ by        ]
			//     [ W^{-T}*GG  0   -I         ]   [ W*uz ]   [ W^{-T}*bz ]
			//
			// and return ux, uy, W*uz.
			//
			// On entry, x, y, z contain bx, by, bz.  On exit, they contain
			// the solution ux, uy, W*uz.
			//
			// If we change variables ux = Q1*v + Q2*w, the system becomes
			//
			//     [ K11 K12 R ]   [ v  ]   [Q1'*(bx+GG'*W^{-1}*W^{-T}*bz)]
			//     [ K21 K22 0 ] * [ w  ] = [Q2'*(bx+GG'*W^{-1}*W^{-T}*bz)]
			//     [ R^T 0   0 ]   [ uy ]   [by                           ]
			//
			//     W*uz = W^{-T} * ( GG*ux - bz ).
			minor := 0
			if !checkpnt.MinorEmpty() {
				minor = checkpnt.MinorTop()
			}

			// bzp := W^{-T} * bz in packed storage
			scale(z, W, true, true)
			pack(z, bzp, dims, &la.IOpt{"mnl", mnl})

			// x := [Q1, Q2]' * (x + Gs' * bzp)
			//    = [Q1, Q2]' * (bx + Gs' * W^{-T} * bz)
			blas.GemvFloat(Gs, bzp, x, 1.0, 1.0, la.OptTrans, &la.IOpt{"m", cdim_pckd})
			lapack.Ormqr(QA, tauA, x, la.OptLeft, la.OptTrans)

			// y := x[:p]
			//    = Q1' * (bx + Gs' * W^{-T} * bz)
			blas.Copy(y, yy)
			blas.Copy(x, y, &la.IOpt{"n", p})

			// x[:p] := v = R^{-T} * by
			blas.Copy(yy, x)
			lapack.Trtrs(QA, x, la.OptUpper, la.OptTrans, &la.IOpt{"n", p})

			// x[p:] := K22^{-1} * (x[p:] - K21*x[:p])
			//        = K22^{-1} * (Q2' * (bx + Gs' * W^{-T} * bz) - K21*v)
			blas.GemvFloat(K, x, x, -1.0, 1.0, &la.IOpt{"m", n - p}, &la.IOpt{"n", p},
				&la.IOpt{"offseta", p}, &la.IOpt{"offsety", p})
			lapack.Potrs(K, x, &la.IOpt{"n", n - p}, &la.IOpt{"offseta", p * (n + 1)},
				&la.IOpt{"offsetb", p})

			// y := y - [K11, K12] * x
			//    = Q1' * (bx + Gs' * W^{-T} * bz) - K11*v - K12*w
			blas.GemvFloat(K, x, y, -1.0, 1.0, &la.IOpt{"m", p}, &la.IOpt{"n", n})

			// y := R^{-1}*y
			//    = R^{-1} * (Q1' * (bx + Gs' * W^{-T} * bz) - K11*v
			//      - K12*w)
			lapack.Trtrs(QA, y, la.OptUpper, &la.IOpt{"n", p})

			// x := [Q1, Q2] * x
			lapack.Ormqr(QA, tauA, x, la.OptLeft)

			// bzp := Gs * x - bzp.
			//      = W^{-T} * ( GG*ux - bz ) in packed storage.
			// Unpack and copy to z.
			blas.GemvFloat(Gs, x, bzp, 1.0, -1.0, &la.IOpt{"m", cdim_pckd})
			unpack(bzp, z, dims, &la.IOpt{"mnl", mnl})

			checkpnt.Check("90solve_chol", minor)
			return nil
		}
		return solve, err
	}
	return factor, nil
}
示例#2
0
// Computes analytic center of A*x <= b with A m by n of rank n.
// We assume that b > 0 and the feasible set is bounded.
func acent(A, b *matrix.FloatMatrix, niters int) (x *matrix.FloatMatrix, ntdecrs []float64, err error) {

	err = nil
	if niters <= 0 {
		niters = MAXITERS
	}
	ntdecrs = make([]float64, 0, niters)

	if A.Rows() != b.Rows() {
		return nil, nil, errors.New("A.Rows() != b.Rows()")
	}

	m, n := A.Size()
	x = matrix.FloatZeros(n, 1)
	H := matrix.FloatZeros(n, n)
	// Helper m*n matrix
	Dmn := matrix.FloatZeros(m, n)

	for i := 0; i < niters; i++ {

		// Gradient is g = A^T * (1.0/(b - A*x)). d = 1.0/(b - A*x)
		// d is m*1 matrix, g is n*1 matrix
		d := matrix.Minus(b, matrix.Times(A, x)).Inv()
		g := matrix.Times(A.Transpose(), d)

		// Hessian is H = A^T * diag(1./(b-A*x))^2 * A.
		// in the original python code expression d[:,n*[0]] creates
		// a m*n matrix where each column is copy of column 0.
		// We do it here manually.
		for i := 0; i < n; i++ {
			Dmn.SetColumn(i, d)
		}

		// Function mul creates element wise product of matrices.
		Asc := matrix.Mul(Dmn, A)
		blas.SyrkFloat(Asc, H, 1.0, 0.0, linalg.OptTrans)

		// Newton step is v = H^-1 * g.
		v := matrix.Scale(g, -1.0)
		PosvFloat(H, v)

		// Directional derivative and Newton decrement.
		lam := blas.DotFloat(g, v)
		ntdecrs = append(ntdecrs, math.Sqrt(-lam))
		if ntdecrs[len(ntdecrs)-1] < TOL {
			return x, ntdecrs, err
		}

		// Backtracking line search.
		// y = d .* A*v
		y := matrix.Mul(d, matrix.Times(A, v))
		step := 1.0
		for 1-step*y.Max() < 0 {
			step *= BETA
		}

	search:
		for {
			// t = -step*y + 1 [e.g. t = 1 - step*y]
			t := matrix.Scale(y, -step).Add(1.0)

			// ts = sum(log(1-step*y))
			ts := t.Log().Sum()
			if -ts < ALPHA*step*lam {
				break search
			}
			step *= BETA
		}
		v.Scale(step)
		x.Plus(v)
	}
	// no solution !!
	err = errors.New(fmt.Sprintf("Iteration %d exhausted\n", niters))
	return x, ntdecrs, err
}
示例#3
0
文件: kkt.go 项目: sguzwf/algorithm
// Solution of KKT equations with zero 1,1 block, by eliminating the
// equality constraints via a QR factorization, and solving the
// reduced KKT system by another QR factorization.
//
// Computes the QR factorization
//
//        A' = [Q1, Q2] * [R1; 0]
//
// and returns a function that (1) computes the QR factorization
//
//        W^{-T} * G * Q2 = Q3 * R3
//
// (with columns of W^{-T}*G in packed storage), and (2) returns a function for solving
//
//        [ 0    A'   G'    ]   [ ux ]   [ bx ]
//        [ A    0    0     ] * [ uy ] = [ by ].
//        [ G    0   -W'*W  ]   [ uz ]   [ bz ]
//
// A is p x n and G is N x n where N = dims['l'] + sum(dims['q']) +
// sum( k**2 for k in dims['s'] ).
//
func kktQr(G *matrix.FloatMatrix, dims *sets.DimensionSet, A *matrix.FloatMatrix, mnl int) (kktFactor, error) {

	p, n := A.Size()
	cdim := dims.Sum("l", "q") + dims.SumSquared("s")
	cdim_pckd := dims.Sum("l", "q") + dims.SumPacked("s")

	QA := A.Transpose()
	tauA := matrix.FloatZeros(p, 1)
	lapack.Geqrf(QA, tauA)

	Gs := matrix.FloatZeros(cdim, n)
	tauG := matrix.FloatZeros(n-p, 1)
	u := matrix.FloatZeros(cdim_pckd, 1)
	vv := matrix.FloatZeros(n, 1)
	w := matrix.FloatZeros(cdim_pckd, 1)
	checkpnt.AddMatrixVar("tauA", tauA)
	checkpnt.AddMatrixVar("tauG", tauG)
	checkpnt.AddMatrixVar("Gs", Gs)
	checkpnt.AddMatrixVar("qr_u", u)
	checkpnt.AddMatrixVar("qr_vv", vv)

	factor := func(W *sets.FloatMatrixSet, H, Df *matrix.FloatMatrix) (KKTFunc, error) {
		var err error = nil
		minor := 0
		if !checkpnt.MinorEmpty() {
			minor = checkpnt.MinorTop()
		}

		// Gs = W^{-T}*G, in packed storage.
		blas.Copy(G, Gs)
		//checkpnt.Check("00factor_qr", minor)
		scale(Gs, W, true, true)
		//checkpnt.Check("01factor_qr", minor)
		pack2(Gs, dims, 0)
		//checkpnt.Check("02factor_qr", minor)

		// Gs := [ Gs1, Gs2 ]
		//     = Gs * [ Q1, Q2 ]
		lapack.Ormqr(QA, tauA, Gs, la.OptRight, &la.IOpt{"m", cdim_pckd})
		//checkpnt.Check("03factor_qr", minor)

		// QR factorization Gs2 := [ Q3, Q4 ] * [ R3; 0 ]
		lapack.Geqrf(Gs, tauG, &la.IOpt{"n", n - p}, &la.IOpt{"m", cdim_pckd},
			&la.IOpt{"offseta", Gs.Rows() * p})
		checkpnt.Check("10factor_qr", minor)

		solve := func(x, y, z *matrix.FloatMatrix) (err error) {
			// On entry, x, y, z contain bx, by, bz.  On exit, they
			// contain the solution x, y, W*z of
			//
			//     [ 0         A'  G'*W^{-1} ]   [ x   ]   [bx       ]
			//     [ A         0   0         ] * [ y   ] = [by       ].
			//     [ W^{-T}*G  0   -I        ]   [ W*z ]   [W^{-T}*bz]
			//
			// The system is solved in five steps:
			//
			//       w := W^{-T}*bz - Gs1*R1^{-T}*by
			//       u := R3^{-T}*Q2'*bx + Q3'*w
			//     W*z := Q3*u - w
			//       y := R1^{-1} * (Q1'*bx - Gs1'*(W*z))
			//       x := [ Q1, Q2 ] * [ R1^{-T}*by;  R3^{-1}*u ]

			minor := 0
			if !checkpnt.MinorEmpty() {
				minor = checkpnt.MinorTop()
			}

			// w := W^{-T} * bz in packed storage
			scale(z, W, true, true)
			pack(z, w, dims)
			//checkpnt.Check("00solve_qr", minor)

			// vv := [ Q1'*bx;  R3^{-T}*Q2'*bx ]
			blas.Copy(x, vv)
			lapack.Ormqr(QA, tauA, vv, la.OptTrans)
			lapack.Trtrs(Gs, vv, la.OptUpper, la.OptTrans, &la.IOpt{"n", n - p},
				&la.IOpt{"offseta", Gs.Rows() * p}, &la.IOpt{"offsetb", p})
			//checkpnt.Check("10solve_qr", minor)

			// x[:p] := R1^{-T} * by
			blas.Copy(y, x)
			lapack.Trtrs(QA, x, la.OptUpper, la.OptTrans, &la.IOpt{"n", p})
			//checkpnt.Check("20solve_qr", minor)

			// w := w - Gs1 * x[:p]
			//    = W^{-T}*bz - Gs1*by
			blas.GemvFloat(Gs, x, w, -1.0, 1.0, &la.IOpt{"n", p}, &la.IOpt{"m", cdim_pckd})
			//checkpnt.Check("30solve_qr", minor)

			// u := [ Q3'*w + v[p:];  0 ]
			//    = [ Q3'*w + R3^{-T}*Q2'*bx; 0 ]
			blas.Copy(w, u)
			lapack.Ormqr(Gs, tauG, u, la.OptTrans, &la.IOpt{"k", n - p},
				&la.IOpt{"offseta", Gs.Rows() * p}, &la.IOpt{"m", cdim_pckd})
			blas.AxpyFloat(vv, u, 1.0, &la.IOpt{"offsetx", p}, &la.IOpt{"n", n - p})
			blas.ScalFloat(u, 0.0, &la.IOpt{"offset", n - p})
			//checkpnt.Check("40solve_qr", minor)

			// x[p:] := R3^{-1} * u[:n-p]
			blas.Copy(u, x, &la.IOpt{"offsety", p}, &la.IOpt{"n", n - p})
			lapack.Trtrs(Gs, x, la.OptUpper, &la.IOpt{"n", n - p},
				&la.IOpt{"offset", Gs.Rows() * p}, &la.IOpt{"offsetb", p})
			//checkpnt.Check("50solve_qr", minor)

			// x is now [ R1^{-T}*by;  R3^{-1}*u[:n-p] ]
			// x := [Q1 Q2]*x
			lapack.Ormqr(QA, tauA, x)
			//checkpnt.Check("60solve_qr", minor)

			// u := [Q3, Q4] * u - w
			//    = Q3 * u[:n-p] - w
			lapack.Ormqr(Gs, tauG, u, &la.IOpt{"k", n - p}, &la.IOpt{"m", cdim_pckd},
				&la.IOpt{"offseta", Gs.Rows() * p})
			blas.AxpyFloat(w, u, -1.0)
			//checkpnt.Check("70solve_qr", minor)

			// y := R1^{-1} * ( v[:p] - Gs1'*u )
			//    = R1^{-1} * ( Q1'*bx - Gs1'*u )
			blas.Copy(vv, y, &la.IOpt{"n", p})
			blas.GemvFloat(Gs, u, y, -1.0, 1.0, &la.IOpt{"m", cdim_pckd},
				&la.IOpt{"n", p}, la.OptTrans)
			lapack.Trtrs(QA, y, la.OptUpper, &la.IOpt{"n", p})
			//checkpnt.Check("80solve_qr", minor)

			unpack(u, z, dims)
			checkpnt.Check("90solve_qr", minor)
			return nil
		}
		return solve, err
	}
	return factor, nil
}