No constraints on the parameters of a sum node

[AlexandreDubray]

Hi,

It seems that there are no constraints on the parameters (log probs for the children) of a sum node. It means that an input (adapted from the doc)

using ProbabilisticCircuits;

X1, X2, X3    = [InputNode(i, Indicator(true))  for i=1:3]
X1_, X2_, X3_ = [InputNode(i, Indicator(false)) for i=1:3]

pc = 1.3 * (X1_ *
             (0.2 * X2_ + 0.8 * X3)) +
     0.7 * (X1 *
             (0.4 * X2 + 0.6 * X3_));

data = Matrix{Union{Missing, Bool}}([missing missing true])
probs = loglikelihoods(pc, data; batch_size=3)
println(exp(first(probs)))

does not raise any error (or warning) and gives a probability of 1.58. Is it allowed by design ? Otherwise, a little check that the weights of the mixture sum up to 1.0 adds almost no overhead at creation time and avoid such situation.

[khosravipasha]

Hi, yes we currently do not enforce constraints when user specifies the circuit manually. This is mostly because we almost always learn both the structure and paramters of the circuit from data which in those cases we do ensure paramters are valid and normalized.