exp.min and exp.max will control the range of
densities function before taking logarithm. If the density values were no
in the range, they would be rescaled. The scaling factor will be also
recorded for post adjustment for observed data log likelihood.
This will provide more accurate posterior probabilities and observed data
log likelihood. Also, U.min and U.max will control the output of
chol when decomposing SIGMA in every
E-steps. If the diagonal terms were out of the range, a PARAM$U.check
would be set to FALSE. Only the components with TRUE
U.check will estimate and update the dispersions in M-steps
for the rest of iterations.
These problems may cause wrong posteriors and log likelihood due to
the degenerate and inflated components. Usually, this is a sign of
overestimate the number of components K, or the initialization
do not provide good estimations for parameters.
See e.step for more information about computing.
.pmclustEnv$CONTROL are
max.iter maximum number of iterations (1000)
abs.err absolute error for convergence (1e-4)
rel.err relative error for convergence (1e-6)
debug debugging flag (0)
RndEM.iter number of RndEM iterations (10)
exp.min minimum exponent (log(.Machine$double.xmin))
exp.max maximum exponent (log(.Machine$double.xmax))
U.min minimum of diagonal of chol
U.max maximum of diagonal of chol
}
These elements govern the computing including number of iterations,
convergent criteria, ill conditions, and numerical issues.
Some of them are machine dependent. Programming with Big Data in R Website:
set.global.# Use set.global() to generate one of this.
# X.spmd should be pre-specified before calling set.global().Run the code above in your browser using DataLab