Given a dataset and the desired Markovian order, this function returns a DBN structure ready to be fitted. Original algorithm at https://link.springer.com/chapter/10.1007/978-3-030-86271-8_14
natPsoho(
dt,
size,
f_dt = NULL,
n_inds = 50,
n_it = 50,
in_cte = 1,
gb_cte = 0.5,
lb_cte = 0.5,
v_probs = c(10, 65, 25),
r_probs = c(-0.5, 1.5),
score = "bge",
p = 0.06,
cte = TRUE
)A 'dbn' object with the structure of the best network found
a data.table with the data of the network to be trained
Maximum number of timeslices of the DBN allowed. Markovian order 1 equals size 2, and so on
previously folded dataset, in case some specific rows have to be removed after the folding
Number of particles used in the algorithm
Maximum number of iterations that the algorithm can perform
parameter that varies the effect of the inertia
parameter that varies the effect of the global best
parameter that varies the effect of the local best
vector that defines the random velocity initialization probabilities
vector that defines the range of random variation of gb_cte and lb_cte
bnlearn score function used
parameter of the truncated geometric distribution for sampling edges
a boolean that determines whether the inertia, global best and local best parameters remain constant or vary as the algorithm progresses. Inertia and local best values decrease as the global best increases, to favor exploration at first and exploitation at the end