gibbs_sampler: Gibbs sampler for DS convex polytopes

In pierrejacob/montecarlodsm: Gibbs sampler for Dempster-Shafer inference in Categorical distributions

Description

This is the main function of the package. It runs the proposed Gibbs sampler for a desired number of iterations, for a given vector of counts. It generates a convex polytope at each iteration, in the form of a matrix "eta". Below the number of categories is denoted by K, and corresponds to the length of the input vector 'counts'. Each category k has count N_k, possibly equal to zero. The zeros are removed from the counts when performing the Gibbs iterations, and "added back" using extend_us.

Usage

gibbs_sampler(niterations, counts)

Arguments

niterations	a number of iterations to perform; each iteration is a full sweep of Gibbs updates for each category. For help on choosing the number of iterations to perform, see the function meeting_times.
counts	a vector of non-negative integers containing the count data; its length defines K, the number of categories. It can include zeros.

Value

A list with the following entries:

• "etas": an array of dimension niterations x K x K. For iteration 'iteration', etasiteration,, is a K x K matrix representing a convex polytope. The feasible set at that iteration, is the set of all vectors theta such that, theta_l/theta_k < etasiteration,k,l for all k,l in 1,...,K.

• "Us": a list of K arrays. For iteration 'iteration', and category 'k' in 1,...,K, if N_k = 0 then Us[k] is NA. If N_k > 0 then Us[k]iteration,, is a matrix with N_k rows, and K columns, representing the auxiliary variables "U"'s generated at that iteration.

Examples

## Not run: 
gibbs_results <- gibbs_sampler(niterations = 5, counts = c(1,2,0,3))
gibbs_results$etas[5,,]

## End(Not run)

pierrejacob/montecarlodsm documentation built on June 16, 2021, 1:06 p.m.