PDA Updater

[jmbarr]

Provides an updater based on probabilistic data association. See the documentation in the Updaters section, or have a look at this gist: https://gist.github.com/jmbarr/92dc83e28c04026136d4f8706a1159c1

[nperree-dstl]

Do these lines work if line 162 is False? I can't work out where posterior_mean, posterior_covariance or kalman_gain would come from in that case.

[nperree-dstl]

Ignore me, I think I've got it now - the first hypothesis always a MissedDetection hypothesis right?

[jmbarr]

the first hypothesis always a MissedDetection hypothesis right?

It seems to be but that's a matter for the associator.associate() method. I think this should work whatever the order; it exploits the fact that not hypothesis returns True for a missed detection (not sure why that is) and calculates the required mean and covariance, and attaches a 0-innovation.

[sdhiscocks]

Set these two sum variables outside the loop to 0?

[sdhiscocks]

Wasn't too sure if the proposed would work, but doing a simple test, looks like it should.

>>> a = 0
>>> a += np.array([[-5], [2]])
>>> a
array([[-5],
       [ 2]])

[jmbarr]

So,
sum_of_innovations = 0
sum_of_innovations += hypothesis.probability * innovation
?

[jmbarr]

So, sum_of_innovations = 0 sum_of_innovations += hypothesis.probability * innovation ?

But won't matter if :class:Probability is corrected.

[sdhiscocks]

diff --git a/stonesoup/updater/probability.py b/stonesoup/updater/probability.py
index c471e514..c6495111 100644
--- a/stonesoup/updater/probability.py
+++ b/stonesoup/updater/probability.py
@@ -156,6 +156,8 @@ class PDAUpdater(ExtendedKalmanUpdater):
             The covariance of the reduced/single Gaussian
         """
 
+        sum_of_innovations = 0
+        sum_of_weighted_cov = 0
         for n, hypothesis in enumerate(hypotheses):
             # Check for the existence of an associated measurement. Because of the way the
             # hypothesis is constructed, you can do this:
@@ -174,15 +176,10 @@ class PDAUpdater(ExtendedKalmanUpdater):
                 innovation = hypothesis.measurement.state_vector - \
                              hypothesis.measurement_prediction.state_vector
 
-            # probably exists a less clunky way of doing this using exists() or overwritten +=
             # All these floats should be redundant if/when the bug in Probability.__mult__() is
             # fixed.
-            if n == 0:
-                sum_of_innovations = float(hypothesis.probability) * innovation
-                sum_of_weighted_cov = float(hypothesis.probability) * (innovation @ innovation.T)
-            else:
-                sum_of_innovations += float(hypothesis.probability) * innovation
-                sum_of_weighted_cov += float(hypothesis.probability) * (innovation @ innovation.T)
+            sum_of_innovations += float(hypothesis.probability) * innovation
+            sum_of_weighted_cov += float(hypothesis.probability) * (innovation @ innovation.T)
 
         posterior_mean += kalman_gain @ sum_of_innovations
         posterior_covariance += \

[sglvladi]

Looking at this made me think whether we could make this more modular by using a KalmanUpdater (or subclass) via composition, rather than inheritance:

• PDAUpdater inherits from Updater and accepts a KalmanUpdater as an updater property, which it then uses to either update() or predict_measurement(), as per lines 127 and 164
• The class could also be renamed to KalmanPDAUpdater/GaussianPDAUpdater to reflect the fact that it applies specifically to Kalman/Gaussian cases.

The above would pave the way for a ParticlePDAUpdater/ELPFUpdater that could be structured similarly to perform the PDA update for particle filters.

[sdhiscocks]

Looking at this made me think whether we could make this more modular by using a KalmanUpdater (or subclass) via composition, rather than inheritance:

• PDAUpdater inherits from Updater and accepts a KalmanUpdater as an updater property, which it then uses to either update() or predict_measurement(), as per lines 127 and 164
• The class could also be renamed to KalmanPDAUpdater/GaussianPDAUpdater to reflect the fact that it applies specifically to Kalman/Gaussian cases.

The above would pave the way for a ParticlePDAUpdater/ELPFUpdater that could be structured similarly to perform the PDA update for particle filters.

Something similar was done for the DIEKF in #891: https://github.com/dstl/Stone-Soup/blob/e27fc3639d791d2441b6c4b8572c57932651deb3/stonesoup/updater/iterated.py

[sdhiscocks]

Per @sglvladi comment, I think it would be good to change structure to support more flexible structure, but I'll merge this for now.