Weighted relations and neural predicates

A weighted relation is an ordinary operation whose answers carry their weights as annotations. There is no dedicated machinery: op registers the callable, each answer names its class as the value and its weight as k, and annotations states the semiring. top orders the answers, (annotation) reads each weight beside its class, and the in-language measure library consumes (weight value) pairs you build with that bridge.

python

from metta import Answer, S, space

m = space()

def mood(day, chosen=None):
    yield Answer(value=S.calm, k=0.25)
    yield Answer(value=S.tense, k=0.75)

m.op(mood, name="mood", effect="nondeterministicReadOnly")
m.annotations("mood", "prob")

m.run("!(collapse (mood today))")                 # (calm tense)
m.run("!(collapse (top 1 (mood today)))")         # (tense)
m.run("!(collapse (let $c (mood today) (pair (annotation) $c)))")
# ((pair 0.25 calm) (pair 0.75 tense))

The callable can be a lookup table, a heuristic, or a model. The torch instance is pettorch.neural_predicate, which lives in the pettorch repository beside this one: it softmaxes a network's forward pass and answers each class with its probability as the annotation, DeepProbLog's nn predicate reading, built entirely on the surface above. Its docs travel with that repository.

Weighted relations and neural predicates ​

Weighted relations and neural predicates