activeSet {isotone}

Active Set Methods for Isotone Optimization

Description

Isotone optimization can be formulated as a convex programming problem with simple linear constraints. This functions offers active set strategies for a collection of isotone optimization problems pre-specified in the package.

Usage

activeSet(isomat, mySolver = "LS", x0 = NULL, ups = 1e-12, check = TRUE, 
maxiter = 100, ...)

Arguments

isomat

Matrix with 2 columns that contains isotonicity conditions, i.e. for row i it holds that fitted value i column 1 <= fitted value i column 2 (see examples)

mySolver

Various functions are pre-defined (see details). Either to funtction name or the corresponding string equivalent can be used. For user-specified functions fSolver with additional arguments can be used (see details as well).

x0

Feasible starting solution. If NULL the null-vector is used internally.

ups

Upper boundary

check

If TRUE, KKT feasibility checks for isotonicity of the solution are performed

maxiter

Iteration limit

...

Additional arguments for the various solvers (see details)

Details

The following solvers are specified. Note that y as the vector of observed values and weights as the vector of weights need to provided through ... for each solver (except for fSolver() and sSolver()). Some solvers need additional arguments as described in the corresponding solver help files. More technical details can be found in the package vignette.

The pre-specified solvers are the following (we always give the corresponding string equivalent in brackets): lsSolver() ("LS") for least squares with diagonal weights, aSolver() ("asyLS") for asymmetric least squares, dSolver() ("L1") for the least absolute value, eSolver() ("L1eps") minimizes l1-approximation. hSolver() ("huber") for Huber loss function, iSolver() ("SILF") for SILF loss (support vector regression), lfSolver() ("GLS") for general least squares with non-diagonal weights, mSolver() ("chebyshev") for Chebyshev L-inf norm, oSolver() ("Lp") for L-p power norm, pSolver() ("quantile") for quantile loss function, and finally sSolver() ("poisson") for Poisson likelihood.

fSolver() for user-specified arbitrary differentiable functions. The arguments fobj (target function ) and gobj (first derivative) must be provided plus any additional arguments used in the definition of fobj.

Value

Generates an object of class activeset.

x

Vector containing the fitted values

y

Vector containing the observed values

lambda

Vector with Lagrange multipliers

fval

Value of the target function

constr.vals

Vector with the values of isotonicity constraints

Alambda

Constraint matrix multiplied by lambda (should be equal to gradient)

gradient

Gradient

isocheck

List containing the KKT checks for stationarity, primal feasibility, dual feasibility, and complementary slackness (>= 0 means feasible)

niter

Number of iterations

call

Matched call

Author(s)

Jan de Leeuw, Kurt Hornik, Patrick Mair

References

de Leeuw, J., Hornik, K., Mair, P. (2009). Isotone optimization in R: Active Set methods and pool-adjacent-violators algorithm. Journal of Statistical Software, 32(5), 1-24.

See Also

gpava, lsSolver, dSolver, mSolver, fSolver, pSolver, lfSolver, oSolver, aSolver, eSolver, sSolver, hSolver, iSolver

Examples

## Data specification
set.seed(12345)
y <- rnorm(9)               ##normal distributed response values
w1 <- rep(1,9)              ##unit weights
Atot <- cbind(1:8, 2:9)     ##Matrix defining isotonicity (total order)
Atot


## Least squares solver (pre-specified and user-specified)
fit.ls1 <- activeSet(Atot, "LS", y = y, weights = w1)
fit.ls1
summary(fit.ls1)
fit.ls2 <- activeSet(Atot, fSolver, fobj = function(x) sum(w1*(x-y)^2), 
gobj = function(x) 2*drop(w1*(x-y)), y = y, weights = w1)

## LS vs. GLS solver (needs weight matrix)
set.seed(12345)
wvec <- 1:9
wmat <- crossprod(matrix(rnorm(81),9,9))/9  
fit.wls <- activeSet(Atot, "LS", y = y, weights = wvec)
fit.gls <- activeSet(Atot, "GLS", y = y, weights = wmat)


## Quantile regression
fit.qua <- activeSet(Atot, "quantile", y = y, weights = wvec, aw = 0.3, bw = 0.7)


## Mean absolute value norm
fit.abs <- activeSet(Atot, "L1", y = y, weights = w1)

## Lp norm
fit.pow <- activeSet(Atot, "Lp", y = y, weights = w1, p = 1.2)

## Chebyshev norm
fit.che <- activeSet(Atot, "chebyshev", y = y, weights = w1)

## Efron's asymmetric LS
fit.asy <- activeSet(Atot, "asyLS", y = y, weights = w1, aw = 2, bw = 1)

## Huber and SILF loss
fit.hub <- activeSet(Atot, "huber", y = y, weights = w1, eps = 1)
fit.svm <- activeSet(Atot, "SILF", y = y, weights = w1, beta = 0.8, eps = 0.2)


## Negative Poisson log-likelihood
set.seed(12345)
yp <- rpois(9,5)
x0 <- 1:9
fit.poi <- activeSet(Atot, "poisson", x0 = x0, y = yp)

## LS on tree ordering
Atree <- matrix(c(1,1,2,2,2,3,3,8,2,3,4,5,6,7,8,9),8,2)
Atree
fit.tree <- activeSet(Atree, "LS", y = y, weights = w1)


## LS on loop ordering
Aloop <- matrix(c(1,2,3,3,4,5,6,6,7,8,3,3,4,5,6,6,7,8,9,9),10,2)
Aloop
fit.loop <- activeSet(Aloop, "LS", y = y, weights = w1)


## LS on block ordering
Ablock <- cbind(c(rep(1,3),rep(2,3),rep(3,3),rep(4,3),rep(5,3),rep(6,3)),c(rep(c(4,5,6),3),
rep(c(7,8,9),3)))
Ablock
fit.block <- activeSet(Ablock, "LS", y = y, weights = w1)

## Isotone LS regression using gpava and active set (same results)
pava.fitted <- gpava(y = y)$x
aset.fitted <- activeSet(Atot, "LS", weights = w1, y = y)$x
mse <- mean((pava.fitted - aset.fitted)^2)
mse