If you have a probability vector, you would typically take the max, not model the action on the probability distribution.

e.g. If an image (from the forward camera) returns the probability vector: {"stop_sign": .9, "green_light": .1} You stop the car 10/10 times.