evidence_samples
The number of samples used in Monte Carlo approximation of the model evidence.
Specification
Alias: None
Arguments: INTEGER
Description
The mc_approx keyword for model evidence indicates that
sample values will be generated from the prior distribution,
and then the simulation model will be evaluated at these
sample values to obtain corresponding likelihood values.
The average of the likelihood weighted by the prior is the
model evidence. The accuracy of this approximation depends on the
number of samples taken, which is specified by
the evidence_samples keyword.
Note that each sample specified by the evidence_samples keyword
will require an evaluation of the simulation model to compute the
corresponding likelihood. So, this may become costly for expensive
simulations. Additionally, many prior samples will have very low
(near zero) likelihood, so millions of samples may be required for
accurate computation of the integral which defines model evidence.
Default Behavior
If evidence_samples is not specified with mc_approx, Dakota
uses the number of chain samples from the MCMC ( chain_samples)
as the number of samples to use for calculating the model evidence.
Expected Output Currently, the model evidence will be printed in the screen output with prefacing text indicating if it is calculated by Monte Carlo sampling.
Usage Tips
