Rmixmod (version 2.0.3)

mixmodStrategy: Create an instance of [Strategy] class

Description

This class will contain all the parameters needed by the estimation algorithms.

Usage

mixmodStrategy(algo = "EM", nbTry = 1, initMethod = "smallEM", nbTryInInit = 50, nbIterationInInit = 5, nbIterationInAlgo = 200, epsilonInInit = 0.001, epsilonInAlgo = 0.001, seed = NULL)

Arguments

algo

list of character string with the estimation algorithm. Possible values: "EM", "SEM", "CEM", c("EM","SEM"). Default value is "EM".

nbTry

integer defining the number of tries. nbTry must be a positive integer. Option available only if init is "random" or "smallEM" or "CEM" or "SEMMax". Default value: 1.

initMethod

a character string with the method of initialization of the algorithm specified in the algo argument. Possible values: "random", "smallEM", "CEM", "SEMMax". Default value: "smallEM".

nbTryInInit

integer defining number of tries in initMethod algorithm. nbTryInInit must be a positive integer. Option available only if init is "smallEM" or "CEM". Default value: 50.

nbIterationInInit

integer defining the number of "EM" or "SEM" iterations in initMethod. nbIterationInInit must be a positive integer. Only available if initMethod is "smallEM" or "SEMMax". Default values: 5 if initMethod is "smallEM" and 100 if initMethod is "SEMMax".

nbIterationInAlgo

list of integers defining the number of iterations if you want to use nbIteration as rule to stop the algorithm(s). Default value: 200.

epsilonInInit

real defining the epsilon value in the initialization step. Only available if initMethod is "smallEM". Default value: 0.001.

epsilonInAlgo

list of reals defining the epsilon value for the algorithm. Warning: epsilonInAlgo doesn't have any sens if algo is SEM, so it needs to be set as NaN in that case. Default value: 0.001.

seed

a positive integer defining the seed of the random number generator. Setting a particular seed allows the user to (re)-generate a particular serie of random numbers. NULL or negative value for a random seed.

Value

a [Strategy] object

Details

There are different ways to initialize an algorithm :

random

Initialization from a random position is a standard way to initialize an algorithm. This random initial position is obtained by choosing at random centers in the data set. This simple strategy is repeated $5$ times (the user can choose the number of times) from different random positions and the position that maximises the likelihood is selected.

smallEM

A maximum of $50$ iterations of the EM algorithm according to the process : $n_i$ numbers of iterations of EM are done (with random initialization) until the smallEM stop criterion value has been reached. This action is repeated until the sum of $n_i$

reaches $50$ iterations (or if in one action $50$ iterations are reached before the stop criterion value).\ It appears that repeating runs of EM is generally profitable since using a single run of EM can often lead to suboptimal solutions.

CEM

$10$ repetitions of $50$ iterations of the CEM algorithm are done. One advantage of initializing an algorithm with CEM lies in the fact that CEM converges generally in a small number of iterations. Thus, without consuming a large amount of CPU times, several runs of CEM are performed. Then EM is run with the best solution among the $10$ repetitions.

SEMMax

A run of $500$ iterations of SEM. The idea is that an SEM sequence is expected to enter rapidly in the neighbourhood of the global maximum of the likelihood function.

Defining the algorithms used in the strategy, the stopping rule and when to stop.

• Algorithms :

  EM

  Expectation Maximisation

  CEM

  Classification EM

SEM

Stochastic EM

Stopping rules for the algorithm :

nbIterationInAlgo

Sets the maximum number of iterations

epsilonInAlgo

Sets relative increase of the log-likelihood criterion

Default values are $200$ nbIterationInAlgo of EM with an epsilonInAlgo value of $10-3$.

References

Biernacki, C., Celeux, G., Govaert, G., 2003. "Choosing starting values for the EM algorithm for getting the highest likelihood in multivariate gaussian mixture models". Computational Statistics and Data Analysis 41, 561-575.

Examples

mixmodStrategy()
   mixmodStrategy(algo="CEM",initMethod="random",nbTry=10,epsilonInInit=0.00001)
   mixmodStrategy(algo=c("SEM","EM"), nbIterationInAlgo=c(200,100), epsilonInAlgo=c(NA,0.000001))