Set control parameters for a lazy learning object.
lazy.control(conIdPar=NULL, linIdPar=1, quaIdPar=NULL,
distance=c("manhattan","euclidean"), metric=NULL,
cmbPar=1, lambda=1e+06)The output of lazy.control is a list containing the
following components: conIdPar, linIdPar, quaIdPar,
distance, metric, cmbPar, lambda.
Parameter controlling the number of neighbors to be used
for identifying and validating constant models. conIdPar can assume
different forms:
conIdPar=c(idm0,idM0,valM0):In this case,
idm0:idM0 is the range in which the best number of
neighbors is searched when identifying the local polynomial
models of degree 0 and where valM0 is the maximum
number of neighbors used for their validation. This means
that the constant models identified with k neighbors,
are validated on the first v neighbors, where
v=min(k,valM0). If valM0=0, valM0 is set
to idMO: see next case for details.
conIdPar=c(idm0,idM0):Here idm0 and
idM0 have the same role as in previous case, and
valM0 is by default set to idM0: each model is
validated on all the neighbors used in identification.
conIdPar=p:Here idmO and idMO are
obtained according to the following formulas: idm0=3
and idMX=5*p. Recommended choice: p=1. As far as
the quantity valM0 is concerned, it gets the default
value as in previous case.
conIdPar=NULL:No constant model is considered.
Parameter controlling the number of neighbors to be used
for identifying and validating linear models. linIdPar can assume
different forms:
linIdPar=c(idm1,idM1,valM1):In this case,
idm1:idM1 is the range in which the best number of
neighbors is searched when identifying the local polynomial
models of degree 1 and where valM1 is the maximum
number of neighbors used for their validation. This means
that the linear models identified with k neighbors, are
validated on the first v neighbors, where
v=min(k,valM1). If valM1=0, valM1 is set
to idM1: see next case for details.
linIdPar=c(idm1,idM1):Here
idm1 and idM1 have the same role as in previous
case, and valM1 is by default set to idM1: each
model is validated on all the neighbors used in identification.
linIdPar=p:Here idmO and idMO are
obtained according to the following formulas:
idm1=3*noPar and idM1=5*p*noPar, where
noPar=nx+1 is the number of parameter of the polynomial
model of degree 1, and nx is the dimensionality of the
input space. Recommended choice: p=1. As far as the
quantity valM1 is concerned, it gets the default value
as in previous case.
linIdPar=NULL:No linear model is considered.
Parameter controlling the number of neighbors to be
used for identifying and validating quadratic
models. quaIdPar can assume different forms:
quaIdPar=c(idm2,idM2,valM2):In this case,
idm2:idM2 is the range in which the best number of
neighbors is searched when identifying the local polynomial
models of degree 2 and where valM2 is the maximum
number of neighbors used for their validation. This means
that the quadratic models identified with k neighbors, are
validated on the first v neighbors, where
v=min(k,valM2). If valM2=0, valM2 is set
to idM2: see next case for details.
quaIdPar=c(idm2,idM2):Here
idm2 and idM2 have the same role as in previous
case, and valM2 is by default set to idM2: each
model is validated on all the neighbors used in identification.
quaIdPar=p:Here idmO and idMO are
obtained according to the following formulas:
idm2=3*noPar and idM2=5*p*noPar, where in this
case the number of parameters is
noPar=(nx+1)*(nx+2)/2, and nx is the
dimensionality of the input space. Recommended choice:
p=1. As far as the quantity valM2 is concerned,
it gets the default value as in previous case.
quaIdPar=NULL:No quadratic model is considered.
The distance metric: can be manhattan or
euclidean.
Vector of n elements. Weights used to evaluate
the distance between query point and neighbors.
Parameter controlling the local combination of
models. cmbPar can assume different forms:
cmbPar=c(cmb0,cmb1,cmb2):In this case, cmbX
is the number of polynomial models of degree X that will
be included in the local combination. Each local model will be
therfore a combination of the best cmb0
models of degree 0, the best cmb1 models of
degree 1, and the best cmb2 models of degree 2
identified as specified by idPar.
cmbPar=cmb:Here cmb is the number of models
that will be combined, disregarding any constraint on the
degree of the models that will be considered. Each local model
will be therfore a combination of the best cmb
models, identified as specified by id_par.
Initialization of the diagonal elements of the local variance/covariance matrix for Ridge Regression.
Mauro Birattari and Gianluca Bontempi
lazy, predict.lazy