This first post is about using Bayes’ theorem to assign devices to rooms probabilistically, based on noisy measurements. I like it because it’s a very practical application of Bayes’ rule. Examples like these always resonate more with me than the artificial examples on Wikipedia about things like drug use.
Here’s the problem. In the meeting room module in BlueSense, we’re collecting data from a Wi-Fi sensor system that measures the presence of Wi-Fi client devices. The systems measures the MAC address, which uniquely identifies a Wi-Fi device, and the strength of the received Wi-Fi signal. There is a sensor in every room, but there aren’t enough sensors to apply an advanced reconstruction technique such as trilateration.
A naive approach to determine in which room a device is, involves ranking the measurements on signal strength, taking the highest signal, finding out in which room the corresponding sensor is, and assigning the device to that room. Simple, right? This actually works ok-ish, because the walls between the rooms dampen the signal a bit, leading the room that the device is in to usually come out on top. But every now and then the system measures a signal that’s way off, which causes us to assign the device to the wrong room. We need some way to deal with the noise. Bayes’ theorem is well suited to these kinds of problems, as we shall see.
Bayes’ rule is stated mathematically as follows:
Let’s assume for a minute that there are only two rooms named A and B, and a “no room” option named N. Let’s create the equations for calculating the probability that the device is in A. We get:
where should be read as “is in A” and as “measured in A”. This makes the probability that a device is in A and the probability that a device gets measured in A.
The conditional probability describes the probability that a device gets measured in A if it is indeed in A, and describes the probability that we’re interested in: whether the device is actually in A, given that it was measured in A. This is called the posterior probability.
The denominator can be expanded to:
Two new probabilities popped up in the denominator. is the probability that we see a measurement in A, given that the device is actually in B. Likewise, is the probability that we see a measurement in N (no room at all) even though the device is in B.
What if we get a measurement in room B? The posterior that the device is actually in A is given by:
Only the numerator changes on the right side. The denominator is basically used to normalize the output between 0 and 1.
The equation for the event that a measurement is seen in N can be constructed in the same way.
That’s it for room A. The probabilities for B and N are calculated in the same way, just make the appropriate substitutions for on the left side of the equation and in the numerator on the right side.
In order to evaluate these equations, we need to quantify our prior beliefs for , and . Let’s assume for now that it’s equally likely for a device to be in any of the three locations, so where is the amount of rooms + 1 (to account for the “no room” case). So, for all three priors.
This initial prior value is only used the very first time we see a device. For subsequent measurements, we use the posterior calculated previously as the new prior.
Finally, we need to think about the probability that a device gets measured in room A, if it is indeed in room A: . This is obviously based on multiple factors such as room size, sensor placement, etc. Empirically, we have found this to be approximately equal to .
For simplicity, we’ll assign to the probabilities that we get measurements in B or N while the device is in A. That would be in this example with . To improve this further, we could make this inversely proportional to the distance from room A to the other locations (i.e. the probability decreases if the room is further away), but I like to keep things simple as long as possible. Additional complexity can always be introduced later later if our simple assumptions don’t seem to work.
A Python-based simulation
We have 1 sensor in the middle of each room and a bunch in the hallway (N). There are 7 rooms in total, so .
Let’s start by simulating measurements for a device that’s actually in room B. After each measurement, the probabilities for each room are computed. Remember that measurements are noisy, so even though the device is in B, occasionally we’ll get something in A or C. After 10 measurements, the device moves across the hall to room E and we simulate 10 more.
Code for Bayesian updates
Dealing with posteriors that tend to zero
As we gather more and more evidence in favor of a specific room, at some point the probabilities for the rooms that are far away will tend to 0. It will take a very long time or be practically impossible for the model to ‘drop’ the belief that a device is in a specific room. This is not the desired behaviour; after all we’re still in a dynamic situation in which a device can actually move to a different room at some point in time. To compensate for this, we’ve put a minimum bound on probabilities of 0.01. It’s not entirely pure from a statistical point of view, but it works well in this case.
The green star shows the actual device location, the purple room shows where the measurement was and the text inside the room indicates the room name and the probability that the device is in that room.
Alright, this shows the desired behavior: we don’t immediately assign a device to the latest known room if there’s a strong prior belief that this device is somewhere else. But we can still convince it that it did in fact move, after a couple of measurements have been received. Much better than the naive solution we’ve been using so far.