Documentation DE-Sampler / nrChains

[stefan-paul]

It seems that the documentation of the DE sampler is a bit confusing.
According to the documentation the parameter "startValue" can be used to define the number of chains.
However, the following code ends in an insufficient starting population:

## Example Code (based on ?runMCMC)
## Generate a test likelihood function. 
ll <- generateTestDensityMultiNormal(sigma = "no correlation")

## Create a BayesianSetup object from the likelihood 
## is the recommended way of using the runMCMC() function.
bayesianSetup <- createBayesianSetup(likelihood = ll, lower = rep(-10, 3), upper = rep(10, 3))


## Run the sampler with number of chains defined
settings = list(iterations = 5000, startValue = 3)
out <- runMCMC(bayesianSetup = bayesianSetup, sampler = "DE", settings = settings)
plot(out)  

Additionally, the parameter "nrChains" is not listed as a parameter of the DE sampler.

[florianhartig]

The meaning of startValue changes between samplers - the behaviour for the DE is explained in ?DE.

I can run the code above without problems - what do you mean by "ends in an insufficient starting population"?

nrChains is one of the arguments that is not sampler-specific, but part of runMCMC, which is the reason why it is not documented in ?DE but in ?runMCMC. I agree that which arguments are documented with the sampler and which in runMCMC could be improved in the help.

[stefan-paul]

Hey Florian,
thanks for the reply!
Sorry, I didn't properly describe the (my?) problem and mixed two things...

About startValue: I know that that the meaning is different for the samplers, yet there are two arguments that can be used for DE which are both described as "number of chains". This is a bit confusing. I guess it would already help to add the explanation in ?runMCMC to ?DE(zs).

About the rest: I don't know if "insufficient" is the correct word.
According to the paper N should be > d. In the example above, the sampler is stuck at the starting population:

For all values > 3 (number of parameters) the sampler works fine.

[florianhartig]

Hi Stefan,

sorry for the late reply.

1. about the documentation - I agree that there is continuous confusion about parameters of the population MCMCs - so for the DEzs, there are the internal chains, and then there is the nrChains, which is the number of separate MCMC runs that you do. My logic in writing this was that an DEzs run, although having formally 3 chains, is one correlated MCMC chain, therefore I treat it as one chain, which, however, in the plots shows 3 (internal) chains. Maybe I should change this although I think it also makes sense to distinguish between completely independent runs and runs that share the same z Matrix if you see what I mean.

2. About your example: the convergence failure is because the DE algorithm is not meant to be set up with such a low number of chains, as it doesn't have memory and mixes proposals only from the current settings - if you change to DEzs, your code works fine. Alternatively, try DE with a larger number of chains -

settings = list(iterations = 5000, startValue = 50)
out <- runMCMC(bayesianSetup = bayesianSetup, sampler = "DE", settings = settings)
plot(out)

which results in