Initialize augmented rankings at random with pairwise preference data in SMC

[osorensen]

Currently in SMC, when data arrive as pairwise preferences, all particles get the same initial value. In contrast, with partial ranks, they are randomly assigned.

Relevant code lines are here. With pairwise preference data, missing_indicator is always empty and the data generate for setup_rank_data are used for each particle.

BayesMallows/src/particles.cpp

Lines 16 to 49 in 6f6b85d

	std::vector<Particle> initialize_particles(
	const Rcpp::List& initial_values,
	unsigned int n_particles,
	const SMCData& dat
	) {
	vec alpha_samples(initial_values["alpha_init"]);
	mat rho_samples(initial_values["rho_init"]);
	Rcpp::Nullable<cube> aug_init(initial_values["aug_init"]);
	if(rho_samples.n_rows != dat.n_items) {
	Rcpp::stop("Wrong format for initial values for rho.");
	}
	std::vector<Particle> pvec;
	pvec.reserve(n_particles);
	for(size_t i{}; i < n_particles; i++) {
	uvec particle_consistent;
	mat augmented_data;
	if(dat.any_missing || dat.augpair) {
	if(aug_init.isNotNull()) {
	particle_consistent = uvec(dat.n_assessors - dat.num_new_obs, fill::ones);
	augmented_data = Rcpp::as<cube>(aug_init).slice(i);
	} else {
	augmented_data = initialize_missing_ranks(dat.rankings, dat.missing_indicator);
	}
	}
	pvec.emplace_back(
	Particle(alpha_samples(i), rho_samples.col(i), augmented_data,
	dat.n_assessors, particle_consistent));
	}
	return pvec;
	}