Ejemplo n.º 1
0
// Cov returns the covariance between a set of data points based on the current
// GP fit.
func (g *GP) Cov(m *mat64.SymDense, x mat64.Matrix) *mat64.SymDense {
	if m != nil {
		// TODO(btracey): Make this k**
		panic("resuing m not coded")
	}
	// The joint covariance matrix is
	// K(x_*, k_*) - k(x_*, x) k(x,x)^-1 k(x, x*)
	nSamp, nDim := x.Dims()
	if nDim != g.inputDim {
		panic(badInputLength)
	}

	// Compute K(x_*, x) K(x, x)^-1 K(x, x_*)
	kstar := g.formKStar(x)
	var tmp mat64.Dense
	tmp.SolveCholesky(g.cholK, kstar)
	var tmp2 mat64.Dense
	tmp2.Mul(kstar.T(), &tmp)

	// Compute k(x_*, x_*) and perform the subtraction.
	kstarstar := mat64.NewSymDense(nSamp, nil)
	for i := 0; i < nSamp; i++ {
		for j := i; j < nSamp; j++ {
			v := g.kernel.Distance(mat64.Row(nil, i, x), mat64.Row(nil, j, x))
			if i == j {
				v += g.noise
			}
			kstarstar.SetSym(i, j, v-tmp2.At(i, j))
		}
	}
	return kstarstar
}
Ejemplo n.º 2
0
// StdDevBatch predicts the standard deviation at a set of locations of x.
func (g *GP) StdDevBatch(std []float64, x mat64.Matrix) []float64 {
	r, c := x.Dims()
	if c != g.inputDim {
		panic(badInputLength)
	}
	if std == nil {
		std = make([]float64, r)
	}
	if len(std) != r {
		panic(badStorage)
	}
	// For a single point, the stddev is
	// 		sigma = k(x,x) - k_*^T * K^-1 * k_*
	// where k is the vector of kernels between the input points and the output points
	// For many points, the formula is:
	// 		nu_* = k(x_*, k_*) - k_*^T * K^-1 * k_*
	// This creates the full covariance matrix which is an rxr matrix. However,
	// the standard deviations are just the diagonal of this matrix. Instead, be
	// smart about it and compute the diagonal terms one at a time.
	kStar := g.formKStar(x)
	var tmp mat64.Dense
	tmp.SolveCholesky(g.cholK, kStar)

	// set k(x_*, x_*) into std then subtract k_*^T K^-1 k_* , computed one row at a time
	var tmp2 mat64.Vector
	row := make([]float64, c)
	for i := range std {
		for k := 0; k < c; k++ {
			row[k] = x.At(i, k)
		}
		std[i] = g.kernel.Distance(row, row)
		tmp2.MulVec(kStar.ColView(i).T(), tmp.ColView(i))
		rt, ct := tmp2.Dims()
		if rt != 1 && ct != 1 {
			panic("bad size")
		}
		std[i] -= tmp2.At(0, 0)
		std[i] = math.Sqrt(std[i])
	}
	// Need to scale the standard deviation to be in the same units as y.
	floats.Scale(g.std, std)
	return std
}
Ejemplo n.º 3
0
// ConditionNormal returns the Normal distribution that is the receiver conditioned
// on the input evidence. The returned multivariate normal has dimension
// n - len(observed), where n is the dimension of the original receiver. The updated
// mean and covariance are
//  mu = mu_un + sigma_{ob,un}^T * sigma_{ob,ob}^-1 (v - mu_ob)
//  sigma = sigma_{un,un} - sigma_{ob,un}^T * sigma_{ob,ob}^-1 * sigma_{ob,un}
// where mu_un and mu_ob are the original means of the unobserved and observed
// variables respectively, sigma_{un,un} is the unobserved subset of the covariance
// matrix, sigma_{ob,ob} is the observed subset of the covariance matrix, and
// sigma_{un,ob} are the cross terms. The elements of x_2 have been observed with
// values v. The dimension order is preserved during conditioning, so if the value
// of dimension 1 is observed, the returned normal represents dimensions {0, 2, ...}
// of the original Normal distribution.
//
// ConditionNormal returns {nil, false} if there is a failure during the update.
// Mathematically this is impossible, but can occur with finite precision arithmetic.
func (n *Normal) ConditionNormal(observed []int, values []float64, src *rand.Rand) (*Normal, bool) {
	if len(observed) == 0 {
		panic("normal: no observed value")
	}
	if len(observed) != len(values) {
		panic("normal: input slice length mismatch")
	}
	for _, v := range observed {
		if v < 0 || v >= n.Dim() {
			panic("normal: observed value out of bounds")
		}
	}

	ob := len(observed)
	unob := n.Dim() - ob
	obMap := make(map[int]struct{})
	for _, v := range observed {
		if _, ok := obMap[v]; ok {
			panic("normal: observed dimension occurs twice")
		}
		obMap[v] = struct{}{}
	}
	if len(observed) == n.Dim() {
		panic("normal: all dimensions observed")
	}
	unobserved := make([]int, 0, unob)
	for i := 0; i < n.Dim(); i++ {
		if _, ok := obMap[i]; !ok {
			unobserved = append(unobserved, i)
		}
	}
	mu1 := make([]float64, unob)
	for i, v := range unobserved {
		mu1[i] = n.mu[v]
	}
	mu2 := make([]float64, ob) // really v - mu2
	for i, v := range observed {
		mu2[i] = values[i] - n.mu[v]
	}

	n.setSigma()

	var sigma11, sigma22 mat64.SymDense
	sigma11.SubsetSym(n.sigma, unobserved)
	sigma22.SubsetSym(n.sigma, observed)

	sigma21 := mat64.NewDense(ob, unob, nil)
	for i, r := range observed {
		for j, c := range unobserved {
			v := n.sigma.At(r, c)
			sigma21.Set(i, j, v)
		}
	}

	var chol mat64.Cholesky
	ok := chol.Factorize(&sigma22)
	if !ok {
		return nil, ok
	}

	// Compute sigma_{2,1}^T * sigma_{2,2}^-1 (v - mu_2).
	v := mat64.NewVector(ob, mu2)
	var tmp, tmp2 mat64.Vector
	err := tmp.SolveCholeskyVec(&chol, v)
	if err != nil {
		return nil, false
	}
	tmp2.MulVec(sigma21.T(), &tmp)

	// Compute sigma_{2,1}^T * sigma_{2,2}^-1 * sigma_{2,1}.
	// TODO(btracey): Should this be a method of SymDense?
	var tmp3, tmp4 mat64.Dense
	err = tmp3.SolveCholesky(&chol, sigma21)
	if err != nil {
		return nil, false
	}
	tmp4.Mul(sigma21.T(), &tmp3)

	for i := range mu1 {
		mu1[i] += tmp2.At(i, 0)
	}

	// TODO(btracey): If tmp2 can constructed with a method, then this can be
	// replaced with SubSym.
	for i := 0; i < len(unobserved); i++ {
		for j := i; j < len(unobserved); j++ {
			v := sigma11.At(i, j)
			sigma11.SetSym(i, j, v-tmp4.At(i, j))
		}
	}
	return NewNormal(mu1, &sigma11, src)
}