0
$\begingroup$

I am using the Forward Algorithm to compute the most likely state at a given time, given a series of observations (Pointwise Maximum A Posteriori Estimation). While studying the performance of this estimation algorithm, I noticed that the emission probabilities play a crucial role in how well the algorithm performs. If the emission probabilities are very similar to each other, for example, given 2 observation symbols and 3 hidden states, if the emission probability matrix is \begin{equation} \begin{bmatrix} 0.6 &0.6 &0.6 \\ 0.4 &0.4 &0.4 \end{bmatrix} \end{equation} the algorithm performs very badly and the estimates are simply based on which states occurs more frequently (based on the transition probability matrix). This is expected because we are not getting any information from the observations.

My Question: How do I justify this behavior theoretically? How does a difference in emission probabilities affect the accuracy of a PMAP estimator? Any insights will be deeply appreciated.

$\endgroup$

1 Answer 1

1
$\begingroup$

This answer assumes you have the right emission probabilities. It's not that you don't know them, but rather that they are not ideal for your application.

Let $K_t$ be the $3x3$ state transition matrix. Let $p(x_{t-1}|y_{1:t-1})$ be a row vector. Let $p(y_t|x_t)$ be the emission probabilities for different $x_t$ values given a specific value of $y_t$. Generally,

$$ p(x_t|y_{1:t}) \propto p(x_{t-1}|y_{1:t-1}) K_t \circ p(y_t|x_t). $$

In your case $p(y_t|x_t)$ is constant in $x_t$ given a specific $y_t$. Which means

$$ p(x_t|y_{1:t}) \propto p(x_{t-1}|y_{1:t-1}) K_t. $$

So for each time point, your emission probabilities will not play a role at all.

PS: $\circ$ is the element-by-element multiplication or Hadamard product.

$\endgroup$
3
  • $\begingroup$ That makes sense! How could this reasoning be extended to quantify the effects of different emission probabilties? For example I have noticed that a emission matrix $\begin{bmatrix} 0.63 &0.58 &0.6 \\ 0.57 &0.62 &0.4 \end{bmatrix}$ leads to much worse performance than $\begin{bmatrix} 0.7 & 0.1 &0.6 \\ 0.3 &0.9 &0.4 \end{bmatrix}$. I guess what I am fishing for is a formal way I can approach proving that one is better than the other. Should I think regarding entropy (information supplied) or likelihood probabilities or something like that? $\endgroup$
    – sid
    Nov 17, 2016 at 0:55
  • $\begingroup$ I have a specific application where the effect of changing a design parameter changes the emission probabilities and hence affects the performance of the estimation. Hence I want to study this in further detail. $\endgroup$
    – sid
    Nov 17, 2016 at 0:57
  • $\begingroup$ @sid what do you mean by worse? usually parameters like the ones in your emission matrix need to be estimated. How are you estimating them? Maximum likelihood? $\endgroup$
    – Taylor
    Nov 17, 2016 at 1:15

Your Answer

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

Not the answer you're looking for? Browse other questions tagged or ask your own question.