I'm trying to learn a random walk process as described at section 7 of http://www.cs.cmu.edu/~epxing/papers/SDM08_Ahmed.pdf

I have a set of points over N epochs.

Given a set of clusters $K$, every point is generated as follow

  • pick $k \in K$ following a Chinese restaurant process
  • sample $ x \sim N(\theta_{k,t} , \Sigma)$, where $\theta_{k,t}$ is the mean of the cluster $k$ at time $t$

The mean of the clusters evolves like a markov chain: more specifically it's a random walk.

$\theta_{k,0} \sim N(0, \sigma)$

$\theta_{k,t+1} \sim N(\theta_{k,t}, \rho)$

$\sigma, \rho, \Sigma,$ N are known parameters.

Here is an example of generated points (N=2)

enter image description here enter image description here

Since I generated the points, for every point I know which cluster generated it.

Given this information, the number of clusters and all the points coordinates, the paper specifies I can use a RTS Smoother to learn the $\theta_k$ chain.

I found this example of Rauch Tung Striebel Smoother, but it doesn't look to me like I can apply it to this problem ( as the link I found has the points with a random noise added to them, while I have points generated from a distribution of which I need to find the parameters).

Can you please explain how I can apply a smoother algorithm to discover the parameters of the distribution which generated the points?


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.