/**
* Constructs a synthetic oracle for data generated from a Bayesian network.
* @param vector_var_order vector variables for dataset features
* @param class_variable finite variable for class
* @param var_type_order order of variable types
*/
explicit syn_oracle_bayes_net(const bayesian_network<Factor>& bn,
parameters params = parameters())
: base(bn.arguments(), vector_var_vector(),
std::vector<variable::variable_typenames>
(bn.arguments().size(), variable::FINITE_VARIABLE)),
params(params), bn(bn),
current_rec(finite_numbering_ptr_, vector_numbering_ptr_, dvector) {
rng.seed(static_cast<unsigned>(params.random_seed));
}
//! Throws an assertion violation if the DBN is not valid
void check_valid() const {
// The arguments at the current and next time step
domain_type vars_t = arguments(processes(), current_step);
domain_type vars_t1 = arguments(processes(), next_step);
// Check the prior and the transition model
assert(prior.arguments() == vars_t);
assert(is_superset(transition.arguments(),vars_t1));
assert(is_subset(transition.arguments(), set_union(vars_t, vars_t1)));
}
/**
* Adds a new CPD to the transition model.
* @param factor
* A factor that represents the conditional probability distribution.
* The arguments of this factor must be either time-t or time-t+1
* arguments of one or more timed processes.
* @param p
* The process, for which the CPD is being added. The argument
* factor must contain the t+1-step argument of process p.
*/
void add_factor(process_type p, const F& factor) {
assert(factor.arguments().count(p->next()) > 0);
for (argument_type v : factor.arguments()) {
int t = boost::any_cast<int>(v.index());
assert(t == current_step || t == next_step);
}
processes_.insert(p);
transition.add_factor(p->next(), factor);
}
//! Initialize the oracle
void init() {
assert(params.valid());
// Initialize the random number generator
rng.seed(static_cast<unsigned>(params.random_seed));
uniform_prob = boost::uniform_real<double>(0,1);
// Construct the graph and make CPTs
if (num_finite() > 2 * params.num_parents) {
for (size_t j = 2 * params.num_parents; j < num_finite(); ++j) {
// Choose NUM_PARENTS random parents
finite_domain parents;
std::vector<size_t> r(randperm(j, rng));
for (size_t k = 0; k < params.num_parents; ++k)
parents.insert(finite_seq[r[k]]);
parents.insert(finite_seq[j]);
tablef f(parents.plus(finite_seq[j]));
// RIGHT HERE NOW: MAKE FACTOR
bn.add_factor(parents, finite_seq[j]);
}
}
}
/**
* Unrolls the dynamic Bayesian network over steps 0, ..., n.
* The unrolled network contains
*/
bayesian_network<F> unroll(std::size_t n) const {
// Initialize the prior
std::map<argument_type, argument_type> prior_var_map
= make_process_var_map(processes(), current_step, 0);
bayesian_network<F> bn;
for (process_type p : processes()) {
F factor = prior.factor(p->current());
factor.subst_args(prior_var_map);
bn.add_factor(p->at(0), factor);
}
// Add the n transition models
for(std::size_t t = 0; t < n; t++) {
std::map<argument_type, argument_type> var_map
= map_union(make_process_var_map(processes(), current_step, t),
make_process_var_map(processes(), next_step, t+1));
for (process_type p : processes()) {
F cpd = transition[p->next()];
cpd.subst_args(var_map);
bn.add_factor(p->at(t+1), cpd);
}
}
return bn;
}
/**
* Returns the ancestors of the t+1-time arguments in the transition model.
* \param procs The set of processes whose ancestors are being sought
*/
domain_type ancestors(const std::set<process_type>& procs) const {
return transition.ancestors(arguments(procs, next_step));
}
/**
* Removes all processes and factors from this DBN.
*/
void clear() {
prior.clear();
transition.clear();
processes_.clear();
}
/**
* Adds a new factor to the prior distribution.
* @param factor
* A factor that represents the conditional probability distribution.
* The arguments of this factor must be time-t arguments of one or
* more timed processes.
* @param v
* The head of the conditional probability distribution.
*/
void add_factor(argument_type head, const F& factor) {
assert(factor.arguments().count(head) > 0);
check_index(factor.arguments(), current_step);
prior.add_factor(head, factor);
}
