I am using an Bayesian sensor to detect if we are all sleeping or not, but I am running into the problem of figuring out the optimal probability of observations.
So these are the observations I use to detect if everyone in the house is sleeping and it’s probability:
- Shutters down: 0.9
- All phones plugged in: 0.7
- Last motion at Master Bedroom: 0.7
And then there are some negating observations:
- Ground floor activity: 0.1
- Xbox on: 0.1
- Laptop on: 0.1
- Nobody home: 0
I tampered with these probabilities so that only when all these activities are correct (so shutters down, all phones plugged in, last motion at Master Bedroom, no devices on) the probability will rise above the 0.95 threshold and thus the sensor telling me that we are sleeping.
It feels like I am doing something wrong with the Bayesian sensor since I picked these probabilities to match the result I want. If I am being fair and ask my self: “How probable is it that all phone are plugged in when we are sleeping”, I would say that that would be very very high. More 0.9 than the 0.7.
But when I change it to 0.9, it can happen that only covers down and phones plugged in together with an device turned on would still carry the total probability over the 0.95 threshold when I know this can’t be true if a device is turned on.
Then the question I ask myself is, why use an Bayesian sensor at all? Why not simply check whether the shutters are down and phones are plugged in, etc.
Can someone shed some more light on this?