Generate a probability list for GMJMCMC (Genetically Modified MJMCMC)
gen.probs.gmjmcmc(transforms)A named list with eight elements:
largeThe probability of a large jump kernel in the MJMCMC algorithm. With this probability, a large jump proposal will be made; otherwise, a local Metropolis-Hastings proposal will be used. One needs to consider good mixing around and between modes when specifying this parameter.
large.kernA numeric vector of length 4 specifying the probabilities for different types of large jump kernels. The four components correspond to:
Random change with random neighborhood size
Random change with fixed neighborhood size
Swap with random neighborhood size
Swap with fixed neighborhood size
These probabilities will be automatically normalized if they do not sum to 1.
localopt.kernA numeric vector of length 2 specifying the probabilities for different local optimization methods during large jumps. The first value represents the probability of using simulated annealing, while the second corresponds to the greedy optimizer. These probabilities will be normalized if needed.
random.kernA numeric vector of length 2 specifying the probabilities
of first two randomization kernels applied after local optimization. These correspond
to the same kernel types as in large.kern but are used for local proposals
where type and 2 only are allowed.
mhA numeric vector specifying the probabilities of different standard Metropolis-Hastings kernels, where the first four as the same as for other kernels, while fifths and sixes components are uniform addition/deletion of a covariate.
filterA numeric value controlling the filtering of features with low posterior probabilities in the current population. Features with posterior probabilities below this threshold will be removed with a probability proportional to \(1 - P(\text{feature} \mid \text{population})\).
genA numeric vector of length 4 specifying the probabilities of different feature generation operators. These determine how new nonlinear features are introduced. The first entry gives the probability for an interaction, followed by modification, nonlinear projection, and a mutation operator, which reintroduces discarded features. If these probabilities do not sum to 1, they are automatically normalized.
transA numeric vector of length equal to the number of elements in transforms,
specifying the probabilities of selecting each nonlinear transformation from \(\mathcal{G}\).
By default, a uniform distribution is assigned, but this can be modified by providing a specific
transforms argument.
A list of the transformations used (to get the count).
gen.probs.gmjmcmc(c("p0", "exp_dbl"))
Run the code above in your browser using DataLab