I'm interested in computing of cross-covariance between state vectors at two different times;
Cov{xk,xk-m}. For example considering xk & Pk, the state vector and covariance matrix at time step of k, and xk-m & Pk-m the state vector at time step of k-m, I want to express Cov{xk,xk-m} w.r.t Pk. I appreciate any help.
I think this question would be better on math.stackexchange.com or dsp.stackexchange.com (where I found help with a Kalman filter question).
I don't have the time to do this calculation, but in principal I think it should be possible. However, the answer will depend on more than just the xi and the Pi, because there will be dependencies on all the other Kalman inputs (in particular the measurements and their covariances etc). If I was going to do the calculation, I would start with the simplest m=1 case, and generalise from there.
Related
How process noise covariance and measurement noise covariance are helping better functioning of Kalman filter ?
Can someone explain intuitively without significant equations and math please.
Well, its difficult to explain mathematical things (like kalman filters) without mathematics, but here's my attempt:
There are two parts to a kalman filter, a time update part and a measurement part. In the time update part we estimate the state at the time of observation; in the measurement part we combine (via least squares) our 'predictions' (ie the estimate from the time update) with the measurements to get a new estimate of the state.
So far, no mention of noise. There are two sources of noise: one in the time update part (sometimes called process noise) and one in the measurement part (observation noise). In each case what we need is a measure of the 'size' of that noise, ie the covariance matrix. These are used when we combine the
predictions with the measurements. When we view our predictions as very uncertain (that is, they have a large covariance matrix) the combination will be closer to the measurements than to the predictions; on the other hand when we view our predictions as very good (small covariance) the combination will be closer to the predictions than to the measurements.
So you could look upon the process and observation noise covariances as saying how much to trust the (parts of) the predictions and observations. Increasing, say, the variance of a particular component of the predictions is to say: trust this prediction less; while increasing the variance of a particular measurement is to say: trust this measurement less. This is mostly an analogy but it can be made more precise. A simple case is when the covariance matrices are diagonal. In that case the cost, ie the contrinution to what we are trying to minimise, of a difference between an measurement and the computed value is te square of that difference, divided by the observations variance. So the higher an observations variance, the lower the cost.
Note that out of the measurement part we also get a new state covariance matrix; this is used (along with the process noise and the dynamics) in the next time update when we compute the predicted state covariance.
I think the question of why the covariance is the appropriate measure of the size of the noise is rather a deep one, as is why least squares is the appropriate way to combine the predictions and the measurements. The shallow answer is that kalman filtering and least squares have been found, over decades (centuries in the case of least squares), to work well in many application areas. In the case of kalman filtering I find the derivation of it from hidden markobv models (From Hidden Markov Models to Linear Dynamical Systems by T.Minka, though this is rather mathematical) convincing. In Hidden markov models we seek to find the (conditional) probability of the states given the measurements so far; Minka shows that if the measurements are linear functions of the states and the dynamics are linear and all probability distributions are Gaussian, then we get the kalman filter.
I'm building a Neural Network from scratch that categorizes values of x into 21 possible estimates for sin(x). I plan to use MSE as my loss function.
MSE of each minibatch = ||y(x) - a||^2, where y(x) is the vector of network outputs for x-
values in the minibatch. a is the vector of expected outputs that correspond to each x.
After finding the loss, the column vector of all weights in the network is recalculated. Column vector of delta w's ~= column vector of partial derivatives of C with respect to each weight.
∇C≡(∂C/∂w1,∂C/∂w2...).T and Δw =−η∇C where η is the (positive) learn rate.
The problem is, to find the gradient of C, you have to differentiate with respect to each weight. What does that function even look like? It's not just the previously stated MSE right?
Any help is appreciated. Also, apologies in advance if this question is misplaced, I wasn't sure if it belonged here or in a math forum.
Thank you.
(I might add that I have tried to find an answer to this online, but few examples exist that either don't use libraries to do the dirty work or present the information clearly.)
http://neuralnetworksanddeeplearning.com/chap2.html
I had found this a while ago but only now realized it's significance. The link describes δ(j,l) as an intermediary value to arrive at the partial derivative of C with respect to weights. I will post back here with a full answer if the link above answers my question, as I've seen a few posts similar to mine that have yet to be answered.
I have AR(1) model with data samples $N=500$ that is driven by a random input sequence x. THe observation y is corrupted with measurement noise $v$ of zero mean. The model is
y(t) = 0.195y(t-1) + x(t) + v(t) where x(t) is generated as randn(). I am unsure how to represent this as a state space model and how to estimate the parameters $a$ and the states. I tried the state space representation would be
d(t) = \mathbf{a^T} d(t) + x(t)
y(t) = \mathbf{h^T}d(t) + sigma*v(t)
sigma =2.
I cannot understand how to perform parameter and state estimation. Using the toolbox mentioned below, I checked the Equations of KF to be matching with those in textbooks. However, the approach for parameter estimation is different. I will appreciate a recommendation for the implementation procedure.
Implementation 1:
I am following the implementation here : Learning Kalman Filter. This implementation does not use Expectation Maximization to estimate the parameters of AR model and it finds out the Covariance of the process noise. In my case, I don't have a process noise, but an input $x$.
Implementation 2: Kalman Filter by Kevin Murphy is another toolbox which uses EM for parameter estimation of AR model. Now, it is confusing since both the implementations uses different approach for parameter estimation.
I am having a tough time figuring out the correct approach, the state space model representation and the code. Shall appreciate recommendations on how to proceed.
I ran the first implementation for the KalmanARSquareRoot technique and the result is completely different. There is Exponential Moving Average Smoothing being performed and a MA filer of length 30 being used. The toolbox runs fine if I run the Demo examples. But on changing the model, the result is extremely poor. Maybe I am doing something wrong. Do I need to change the equations of KF for my time series?
In the second implementation, I cannot figure out what and where to change the Equations.
In general, if I have to use these tools, then do I need to change the KF equations for every time series model? How do I write the Equations myself if these toolboxes are inappropriate for all the time series model - AR, MA, ARMA?
I only have a bit of experience with Kalman Filters, so take this with a grain of salt.
It seems you shouldn't need to change the equations at all. Working with the second package (learn_kalman), you can create an A0 matrix of size [length(d(t)) length(d(t))]. C0 is the same, and in your case the initial state probably makes sense to be the Identity matrix (unlike your A0. All you need to do is choose a good initial condition.
However, I took a look at your system (plotted an example) and it seems noise dominates your system. KF is an optimal estimator but I have not known it to reject that much noise. It only guarantees a reduced covariance...which means that if your system is mostly dominated by noise, you will calculate a bad model that estimates your system given the noise!
Try plotting [d f] where d is the initial data and f is calculated using the regressive formula f(t) = C * A * f(t-1) :
f(t) = A * f(t-1)
; y(t) = C * f(t)
That is, pretend as if there is no noise but using the estimated AR model. You will see that it rejects all the noise and 'technically' models the system well (since the only unique behaviour is at the very beginning).
For example, if you have a system with A = 0.195, Q=R=0.l then you will converge to an A = 0.2207 but it still isn't good enough. Here the problem is that your initial state is so low, that within a few steps of data and you are essentially at 0 accounting for noise. Naturally KF can converge to a LOT of model solutions that are similar. Any noise will throw off even the best initial condition.
If you increase the resolution of your data in some way (e.g. larger initial condition, more refined timesteps) you will see a good fit. Ex, changing your initial condition to 110 and you'll find the two curves similar, though the model is still fairly different.
I am not aware of any approach to model your data well. If the noise variance is in fact 1 and your system converges to 0 that quickly, it seems doomed to not be effectively modelled since you just don't capture any unique behaviour in the dataset.
I have two data sets (t,y1) and (t,y2). These data sets visually look same but their is some time delay or magnitude shift. i want to find the similarity between the two curves (giving the score of similarity 1 for approximately similar curves and 0 for not similar curves). Some curves are seem to be different because of oscillation in data. so, i am searching for the method to find the similarity between the curves. i already tried gradient command in Matlab to find the slope of the curve at each time step and compared it. but it is not giving me satisfactory results. please anybody suggest me the method to find the similarity between the curves.
Thanks in Advance
This answer assumes your y1 and y2 are signals rather than curves. The latter I would try to parametrise with POLYFIT.
If they really look the same, but are shifted in time (and not wrapped around) then you can:
y1n=y1/norm(y1);
y2n=y2/norm(y2);
normratio=norm(y1)/norm(y2);
c=conv2(y1n,y2n,'same');
[val ind]=max(c);
ind will indicate the time shift and normratio the difference in magnitude.
Both can be used as features for your similarity metric. I assume however your signals actually vary by more than just timeshift or magnitude in which case some sort of signal parametrisation may be a better choice and then building a metric on those parameters.
Without knowing anything about your data I would first try with AR (assuming things as typical as FFT or PRINCOMP won't work).
For time series data similarity measurement, one traditional solution is DTW (Dynamic Time Warpping)
Kolmongrov Smirnov Test (kstest2 function in Matlab)
Chi Square Test
to measure similarity there is a measure called MIC: Maximal information coefficient. It quantifies the information shared between 2 data or curves.
The dv and dc distance in the following paper may solve your problem.
http://bioinformatics.oxfordjournals.org/content/27/22/3135.full
How can I do K-means clustering of time series data?
I understand how this works when the input data is a set of points, but I don't know how to cluster a time series with 1XM, where M is the data length. In particular, I'm not sure how to update the mean of the cluster for time series data.
I have a set of labelled time series, and I want to use the K-means algorithm to check whether I will get back a similar label or not. My X matrix will be N X M, where N is number of time series and M is data length as mentioned above.
Does anyone know how to do this? For example, how could I modify this k-means MATLAB code so that it would work for time series data? Also, I would like to be able to use different distance metrics besides Euclidean distance.
To better illustrate my doubts, here is the code I modified for time series data:
% Check if second input is centroids
if ~isscalar(k)
c=k;
k=size(c,1);
else
c=X(ceil(rand(k,1)*n),:); % assign centroid randomly at start
end
% allocating variables
g0=ones(n,1);
gIdx=zeros(n,1);
D=zeros(n,k);
% Main loop converge if previous partition is the same as current
while any(g0~=gIdx)
% disp(sum(g0~=gIdx))
g0=gIdx;
% Loop for each centroid
for t=1:k
% d=zeros(n,1);
% Loop for each dimension
for s=1:n
D(s,t) = sqrt(sum((X(s,:)-c(t,:)).^2));
end
end
% Partition data to closest centroids
[z,gIdx]=min(D,[],2);
% Update centroids using means of partitions
for t=1:k
% Is this how we calculate new mean of the time series?
c(t,:)=mean(X(gIdx==t,:));
end
end
Time series are usually high-dimensional. And you need specialized distance function to compare them for similarity. Plus, there might be outliers.
k-means is designed for low-dimensional spaces with a (meaningful) euclidean distance. It is not very robust towards outliers, as it puts squared weight on them.
Doesn't sound like a good idea to me to use k-means on time series data. Try looking into more modern, robust clustering algorithms. Many will allow you to use arbitrary distance functions, including time series distances such as DTW.
It's probably too late for an answer, but:
k-means can be used to cluster longitudinal data
Anony-Mousse is right, DWT distance is the way to go for time series
The methods above use R. You'll find more methods by looking, e.g., for "Iterative Incremental Clustering of Time Series".
I have recently come across the kml R package which claims to implement k-means clustering for longitudinal data. I have not tried it out myself.
Also the Time-series clustering - A decade review paper by S. Aghabozorgi, A. S. Shirkhorshidi and T. Ying Wah might be useful to you to seek out alternatives. Another nice paper although somewhat dated is Clustering of time series data-a survey by T. Warren Liao.
If you did really want to use clustering, then dependent on your application you could generate a low dimensional feature vector for each time series. For example, use time series mean, standard deviation, dominant frequency from a Fourier transform etc. This would be suitable for use with k-means, but whether it would give you useful results is dependent on your specific application and the content of your time series.
I don't think k-means is the right way for it either. As #Anony-Mousse suggested you can utilize DTW. In fact, I had the same problem for one of my projects and I wrote my own class for that in Python. The logic is;
Create your all cluster combinations. k is for cluster count and n is for number of series. The number of items returned should be n! / k! / (n-k)!. These would be something like potential centers.
For each series, calculate distances for each center in each cluster groups and assign it to the minimum one.
For each cluster groups, calculate total distance within individual clusters.
Choose the minimum.
And, the Python implementation is here if you're interested.