hidden markov modelsweb.eecs.umich.edu/~jjcorso/t/2011s_555/files/platt_hmm.pdf · hidden markov...
TRANSCRIPT
![Page 1: Hidden Markov Modelsweb.eecs.umich.edu/~jjcorso/t/2011S_555/files/platt_hmm.pdf · Hidden Markov models are everywhere! Used to model time-series data: - Robot perception/control](https://reader030.vdocuments.site/reader030/viewer/2022041107/5f09e6b67e708231d4290bb6/html5/thumbnails/1.jpg)
Hidden Markov Models
Robert PlattResearch Scientist
Computer Science and Artificial Intelligence LabMIT
Copyright Robert Platt Jr.
![Page 2: Hidden Markov Modelsweb.eecs.umich.edu/~jjcorso/t/2011S_555/files/platt_hmm.pdf · Hidden Markov models are everywhere! Used to model time-series data: - Robot perception/control](https://reader030.vdocuments.site/reader030/viewer/2022041107/5f09e6b67e708231d4290bb6/html5/thumbnails/2.jpg)
Hidden Markov models are everywhere!
Used to model time-series data:- Robot perception/control- Speech recognition- Video understanding- ...
MIT DARPA grand challenge vehicle Human speech
![Page 3: Hidden Markov Modelsweb.eecs.umich.edu/~jjcorso/t/2011S_555/files/platt_hmm.pdf · Hidden Markov models are everywhere! Used to model time-series data: - Robot perception/control](https://reader030.vdocuments.site/reader030/viewer/2022041107/5f09e6b67e708231d4290bb6/html5/thumbnails/3.jpg)
A Markov Process
N states:
At time t, state is:
Discrete time steps:
Stochastic transitions once per time step:
![Page 4: Hidden Markov Modelsweb.eecs.umich.edu/~jjcorso/t/2011S_555/files/platt_hmm.pdf · Hidden Markov models are everywhere! Used to model time-series data: - Robot perception/control](https://reader030.vdocuments.site/reader030/viewer/2022041107/5f09e6b67e708231d4290bb6/html5/thumbnails/4.jpg)
State transitions
Example
Suppose:Transition probabilities:
![Page 5: Hidden Markov Modelsweb.eecs.umich.edu/~jjcorso/t/2011S_555/files/platt_hmm.pdf · Hidden Markov models are everywhere! Used to model time-series data: - Robot perception/control](https://reader030.vdocuments.site/reader030/viewer/2022041107/5f09e6b67e708231d4290bb6/html5/thumbnails/5.jpg)
State transitions
Notation:
- conditional probability table (CPT)- (process dynamics)
![Page 6: Hidden Markov Modelsweb.eecs.umich.edu/~jjcorso/t/2011S_555/files/platt_hmm.pdf · Hidden Markov models are everywhere! Used to model time-series data: - Robot perception/control](https://reader030.vdocuments.site/reader030/viewer/2022041107/5f09e6b67e708231d4290bb6/html5/thumbnails/6.jpg)
State transitions
Notation:
- conditional probability table (CPT)- (process dynamics)
Notice:
![Page 7: Hidden Markov Modelsweb.eecs.umich.edu/~jjcorso/t/2011S_555/files/platt_hmm.pdf · Hidden Markov models are everywhere! Used to model time-series data: - Robot perception/control](https://reader030.vdocuments.site/reader030/viewer/2022041107/5f09e6b67e708231d4290bb6/html5/thumbnails/7.jpg)
Markov Property
Implicit in CPT ...
![Page 8: Hidden Markov Modelsweb.eecs.umich.edu/~jjcorso/t/2011S_555/files/platt_hmm.pdf · Hidden Markov models are everywhere! Used to model time-series data: - Robot perception/control](https://reader030.vdocuments.site/reader030/viewer/2022041107/5f09e6b67e708231d4290bb6/html5/thumbnails/8.jpg)
Markov Property
Implicit in CPT ...
Bayes net for a Markov model:
![Page 9: Hidden Markov Modelsweb.eecs.umich.edu/~jjcorso/t/2011S_555/files/platt_hmm.pdf · Hidden Markov models are everywhere! Used to model time-series data: - Robot perception/control](https://reader030.vdocuments.site/reader030/viewer/2022041107/5f09e6b67e708231d4290bb6/html5/thumbnails/9.jpg)
Markov Property
Implicit in CPT ...
Bayes net for a Markov model:
What does the Bayes net look like if theMarkov property does not hold?
![Page 10: Hidden Markov Modelsweb.eecs.umich.edu/~jjcorso/t/2011S_555/files/platt_hmm.pdf · Hidden Markov models are everywhere! Used to model time-series data: - Robot perception/control](https://reader030.vdocuments.site/reader030/viewer/2022041107/5f09e6b67e708231d4290bb6/html5/thumbnails/10.jpg)
Gridworld Example
Robot on a 4x5 grid …
Define transition dynamics:- 5 possible transitions
(up, down, left, right, stay)
P(right) = 0.4P(left) = 0.15P(up) = 0.15P(down) = 0.15P(stay) = 0.15
- different dynamics at walls...
![Page 11: Hidden Markov Modelsweb.eecs.umich.edu/~jjcorso/t/2011S_555/files/platt_hmm.pdf · Hidden Markov models are everywhere! Used to model time-series data: - Robot perception/control](https://reader030.vdocuments.site/reader030/viewer/2022041107/5f09e6b67e708231d4290bb6/html5/thumbnails/11.jpg)
Warm up question
If robot starts in state
What is probability of landing on stateat time ?
![Page 12: Hidden Markov Modelsweb.eecs.umich.edu/~jjcorso/t/2011S_555/files/platt_hmm.pdf · Hidden Markov models are everywhere! Used to model time-series data: - Robot perception/control](https://reader030.vdocuments.site/reader030/viewer/2022041107/5f09e6b67e708231d4290bb6/html5/thumbnails/12.jpg)
Warm up question
If robot starts in state
What is probability of landing on stateat time ?
i.e. find:
![Page 13: Hidden Markov Modelsweb.eecs.umich.edu/~jjcorso/t/2011S_555/files/platt_hmm.pdf · Hidden Markov models are everywhere! Used to model time-series data: - Robot perception/control](https://reader030.vdocuments.site/reader030/viewer/2022041107/5f09e6b67e708231d4290bb6/html5/thumbnails/13.jpg)
Naïve solution
Find:
![Page 14: Hidden Markov Modelsweb.eecs.umich.edu/~jjcorso/t/2011S_555/files/platt_hmm.pdf · Hidden Markov models are everywhere! Used to model time-series data: - Robot perception/control](https://reader030.vdocuments.site/reader030/viewer/2022041107/5f09e6b67e708231d4290bb6/html5/thumbnails/14.jpg)
Naïve solution
Find:
Sol'n:
![Page 15: Hidden Markov Modelsweb.eecs.umich.edu/~jjcorso/t/2011S_555/files/platt_hmm.pdf · Hidden Markov models are everywhere! Used to model time-series data: - Robot perception/control](https://reader030.vdocuments.site/reader030/viewer/2022041107/5f09e6b67e708231d4290bb6/html5/thumbnails/15.jpg)
Naïve solution
Find:
Sol'n:
Cost:
![Page 16: Hidden Markov Modelsweb.eecs.umich.edu/~jjcorso/t/2011S_555/files/platt_hmm.pdf · Hidden Markov models are everywhere! Used to model time-series data: - Robot perception/control](https://reader030.vdocuments.site/reader030/viewer/2022041107/5f09e6b67e708231d4290bb6/html5/thumbnails/16.jpg)
Dynamic Programming Sol'n
Find:
Let:
![Page 17: Hidden Markov Modelsweb.eecs.umich.edu/~jjcorso/t/2011S_555/files/platt_hmm.pdf · Hidden Markov models are everywhere! Used to model time-series data: - Robot perception/control](https://reader030.vdocuments.site/reader030/viewer/2022041107/5f09e6b67e708231d4290bb6/html5/thumbnails/17.jpg)
Dynamic Programming Sol'n
Find:
Sol'n:
1.
2.
3.
Let:
![Page 18: Hidden Markov Modelsweb.eecs.umich.edu/~jjcorso/t/2011S_555/files/platt_hmm.pdf · Hidden Markov models are everywhere! Used to model time-series data: - Robot perception/control](https://reader030.vdocuments.site/reader030/viewer/2022041107/5f09e6b67e708231d4290bb6/html5/thumbnails/18.jpg)
Dynamic Programming Sol'n
Find:
Sol'n:
1.
2.
3.
Let:
![Page 19: Hidden Markov Modelsweb.eecs.umich.edu/~jjcorso/t/2011S_555/files/platt_hmm.pdf · Hidden Markov models are everywhere! Used to model time-series data: - Robot perception/control](https://reader030.vdocuments.site/reader030/viewer/2022041107/5f09e6b67e708231d4290bb6/html5/thumbnails/19.jpg)
Dynamic Programming Sol'n
Find:
Sol'n:
1.
2.
3.
Cost:
Let:
![Page 20: Hidden Markov Modelsweb.eecs.umich.edu/~jjcorso/t/2011S_555/files/platt_hmm.pdf · Hidden Markov models are everywhere! Used to model time-series data: - Robot perception/control](https://reader030.vdocuments.site/reader030/viewer/2022041107/5f09e6b67e708231d4290bb6/html5/thumbnails/20.jpg)
Dynamic Programming Sol'n
Find:
Another way of writing the DP sol'n:
1.
2.
Let:
![Page 21: Hidden Markov Modelsweb.eecs.umich.edu/~jjcorso/t/2011S_555/files/platt_hmm.pdf · Hidden Markov models are everywhere! Used to model time-series data: - Robot perception/control](https://reader030.vdocuments.site/reader030/viewer/2022041107/5f09e6b67e708231d4290bb6/html5/thumbnails/21.jpg)
Stationary distribution
Given a Markov process with:
The process ultimately converges to a particularstationary distribution:
How do you calculate ?
![Page 22: Hidden Markov Modelsweb.eecs.umich.edu/~jjcorso/t/2011S_555/files/platt_hmm.pdf · Hidden Markov models are everywhere! Used to model time-series data: - Robot perception/control](https://reader030.vdocuments.site/reader030/viewer/2022041107/5f09e6b67e708231d4290bb6/html5/thumbnails/22.jpg)
A Hidden Markov Model
N states:
M observations:
Discrete time steps:
At time t, state is:
At time t, observation is:
You never observe system state ...
![Page 23: Hidden Markov Modelsweb.eecs.umich.edu/~jjcorso/t/2011S_555/files/platt_hmm.pdf · Hidden Markov models are everywhere! Used to model time-series data: - Robot perception/control](https://reader030.vdocuments.site/reader030/viewer/2022041107/5f09e6b67e708231d4290bb6/html5/thumbnails/23.jpg)
Observations
Stochastic observations once per time step:
- when you first arrive in a state
What does the Bayes net for an HMM look like?
Assume:
![Page 24: Hidden Markov Modelsweb.eecs.umich.edu/~jjcorso/t/2011S_555/files/platt_hmm.pdf · Hidden Markov models are everywhere! Used to model time-series data: - Robot perception/control](https://reader030.vdocuments.site/reader030/viewer/2022041107/5f09e6b67e708231d4290bb6/html5/thumbnails/24.jpg)
HMM
What does the Bayes net for an HMM look like?
![Page 25: Hidden Markov Modelsweb.eecs.umich.edu/~jjcorso/t/2011S_555/files/platt_hmm.pdf · Hidden Markov models are everywhere! Used to model time-series data: - Robot perception/control](https://reader030.vdocuments.site/reader030/viewer/2022041107/5f09e6b67e708231d4290bb6/html5/thumbnails/25.jpg)
HMM
Notation:- observation dynamics
Notice:
![Page 26: Hidden Markov Modelsweb.eecs.umich.edu/~jjcorso/t/2011S_555/files/platt_hmm.pdf · Hidden Markov models are everywhere! Used to model time-series data: - Robot perception/control](https://reader030.vdocuments.site/reader030/viewer/2022041107/5f09e6b67e708231d4290bb6/html5/thumbnails/26.jpg)
HMM Definition
N states:
M observations:
Process dynamics:
Observation dynamics:
Prior distribution: or
or
![Page 27: Hidden Markov Modelsweb.eecs.umich.edu/~jjcorso/t/2011S_555/files/platt_hmm.pdf · Hidden Markov models are everywhere! Used to model time-series data: - Robot perception/control](https://reader030.vdocuments.site/reader030/viewer/2022041107/5f09e6b67e708231d4290bb6/html5/thumbnails/27.jpg)
Games you can play with an HMM
1. Inference:- where is the robot now?- where was the robot?- what path did it take?
2. Learning:- what are the robot
process and observation dynamics?
Today
Some other time
![Page 28: Hidden Markov Modelsweb.eecs.umich.edu/~jjcorso/t/2011S_555/files/platt_hmm.pdf · Hidden Markov models are everywhere! Used to model time-series data: - Robot perception/control](https://reader030.vdocuments.site/reader030/viewer/2022041107/5f09e6b67e708231d4290bb6/html5/thumbnails/28.jpg)
Problem #1: Filtering
If the robot starts in prior distribution,
and make a sequence of observations,
Then, where is the robot at time T?
In other words, what is: ?
![Page 29: Hidden Markov Modelsweb.eecs.umich.edu/~jjcorso/t/2011S_555/files/platt_hmm.pdf · Hidden Markov models are everywhere! Used to model time-series data: - Robot perception/control](https://reader030.vdocuments.site/reader030/viewer/2022041107/5f09e6b67e708231d4290bb6/html5/thumbnails/29.jpg)
Naïve solution
? ?
![Page 30: Hidden Markov Modelsweb.eecs.umich.edu/~jjcorso/t/2011S_555/files/platt_hmm.pdf · Hidden Markov models are everywhere! Used to model time-series data: - Robot perception/control](https://reader030.vdocuments.site/reader030/viewer/2022041107/5f09e6b67e708231d4290bb6/html5/thumbnails/30.jpg)
Naïve solution
? ?
Not cool …
![Page 31: Hidden Markov Modelsweb.eecs.umich.edu/~jjcorso/t/2011S_555/files/platt_hmm.pdf · Hidden Markov models are everywhere! Used to model time-series data: - Robot perception/control](https://reader030.vdocuments.site/reader030/viewer/2022041107/5f09e6b67e708231d4290bb6/html5/thumbnails/31.jpg)
Filtering: dynamic programming sol'n
Find:
Define:
![Page 32: Hidden Markov Modelsweb.eecs.umich.edu/~jjcorso/t/2011S_555/files/platt_hmm.pdf · Hidden Markov models are everywhere! Used to model time-series data: - Robot perception/control](https://reader030.vdocuments.site/reader030/viewer/2022041107/5f09e6b67e708231d4290bb6/html5/thumbnails/32.jpg)
Filtering: dynamic programming sol'n
Find:
Define:
Or ...
![Page 33: Hidden Markov Modelsweb.eecs.umich.edu/~jjcorso/t/2011S_555/files/platt_hmm.pdf · Hidden Markov models are everywhere! Used to model time-series data: - Robot perception/control](https://reader030.vdocuments.site/reader030/viewer/2022041107/5f09e6b67e708231d4290bb6/html5/thumbnails/33.jpg)
Filtering: dynamic programming sol'n
Find:
Define:
1.
2.
3.
Also called the forward algorithm
![Page 34: Hidden Markov Modelsweb.eecs.umich.edu/~jjcorso/t/2011S_555/files/platt_hmm.pdf · Hidden Markov models are everywhere! Used to model time-series data: - Robot perception/control](https://reader030.vdocuments.site/reader030/viewer/2022041107/5f09e6b67e708231d4290bb6/html5/thumbnails/34.jpg)
Filtering: dynamic programming sol'n
Find:
Define:
1.
2.
3.
Cost:
![Page 35: Hidden Markov Modelsweb.eecs.umich.edu/~jjcorso/t/2011S_555/files/platt_hmm.pdf · Hidden Markov models are everywhere! Used to model time-series data: - Robot perception/control](https://reader030.vdocuments.site/reader030/viewer/2022041107/5f09e6b67e708231d4290bb6/html5/thumbnails/35.jpg)
Filtering: dynamic programming sol'n
Find:
Define:
Last step:
![Page 36: Hidden Markov Modelsweb.eecs.umich.edu/~jjcorso/t/2011S_555/files/platt_hmm.pdf · Hidden Markov models are everywhere! Used to model time-series data: - Robot perception/control](https://reader030.vdocuments.site/reader030/viewer/2022041107/5f09e6b67e708231d4290bb6/html5/thumbnails/36.jpg)
Problem #2: Smoothing
If the robot starts in prior distribution,
and make a sequence of observations,
Then, where is the robot at time t<T?
In other words, what is: ?
![Page 37: Hidden Markov Modelsweb.eecs.umich.edu/~jjcorso/t/2011S_555/files/platt_hmm.pdf · Hidden Markov models are everywhere! Used to model time-series data: - Robot perception/control](https://reader030.vdocuments.site/reader030/viewer/2022041107/5f09e6b67e708231d4290bb6/html5/thumbnails/37.jpg)
Smoothing: DP solution
Notice:
why?
Find:
![Page 38: Hidden Markov Modelsweb.eecs.umich.edu/~jjcorso/t/2011S_555/files/platt_hmm.pdf · Hidden Markov models are everywhere! Used to model time-series data: - Robot perception/control](https://reader030.vdocuments.site/reader030/viewer/2022041107/5f09e6b67e708231d4290bb6/html5/thumbnails/38.jpg)
Smoothing: DP solution
Notice:
Find:
Therefore:
![Page 39: Hidden Markov Modelsweb.eecs.umich.edu/~jjcorso/t/2011S_555/files/platt_hmm.pdf · Hidden Markov models are everywhere! Used to model time-series data: - Robot perception/control](https://reader030.vdocuments.site/reader030/viewer/2022041107/5f09e6b67e708231d4290bb6/html5/thumbnails/39.jpg)
Smoothing: DP solution
Find recursion for:
![Page 40: Hidden Markov Modelsweb.eecs.umich.edu/~jjcorso/t/2011S_555/files/platt_hmm.pdf · Hidden Markov models are everywhere! Used to model time-series data: - Robot perception/control](https://reader030.vdocuments.site/reader030/viewer/2022041107/5f09e6b67e708231d4290bb6/html5/thumbnails/40.jpg)
Smoothing: DP solution
Find recursion for:
![Page 41: Hidden Markov Modelsweb.eecs.umich.edu/~jjcorso/t/2011S_555/files/platt_hmm.pdf · Hidden Markov models are everywhere! Used to model time-series data: - Robot perception/control](https://reader030.vdocuments.site/reader030/viewer/2022041107/5f09e6b67e708231d4290bb6/html5/thumbnails/41.jpg)
Smoothing: DP solution
![Page 42: Hidden Markov Modelsweb.eecs.umich.edu/~jjcorso/t/2011S_555/files/platt_hmm.pdf · Hidden Markov models are everywhere! Used to model time-series data: - Robot perception/control](https://reader030.vdocuments.site/reader030/viewer/2022041107/5f09e6b67e708231d4290bb6/html5/thumbnails/42.jpg)
Smoothing: DP solution
![Page 43: Hidden Markov Modelsweb.eecs.umich.edu/~jjcorso/t/2011S_555/files/platt_hmm.pdf · Hidden Markov models are everywhere! Used to model time-series data: - Robot perception/control](https://reader030.vdocuments.site/reader030/viewer/2022041107/5f09e6b67e708231d4290bb6/html5/thumbnails/43.jpg)
Smoothing: DP solution
Find:
Define:
1.
2.
3.
![Page 44: Hidden Markov Modelsweb.eecs.umich.edu/~jjcorso/t/2011S_555/files/platt_hmm.pdf · Hidden Markov models are everywhere! Used to model time-series data: - Robot perception/control](https://reader030.vdocuments.site/reader030/viewer/2022041107/5f09e6b67e708231d4290bb6/html5/thumbnails/44.jpg)
Smoothing: DP solution
Find:
Define:
1.
2.
3.
Cost:
![Page 45: Hidden Markov Modelsweb.eecs.umich.edu/~jjcorso/t/2011S_555/files/platt_hmm.pdf · Hidden Markov models are everywhere! Used to model time-series data: - Robot perception/control](https://reader030.vdocuments.site/reader030/viewer/2022041107/5f09e6b67e708231d4290bb6/html5/thumbnails/45.jpg)
Smoothing: DP solution
Find:
1. Calculate all alphas and betas
2.
Also called: forward-backward algorithm
![Page 46: Hidden Markov Modelsweb.eecs.umich.edu/~jjcorso/t/2011S_555/files/platt_hmm.pdf · Hidden Markov models are everywhere! Used to model time-series data: - Robot perception/control](https://reader030.vdocuments.site/reader030/viewer/2022041107/5f09e6b67e708231d4290bb6/html5/thumbnails/46.jpg)
Viterbi: finding the most likely path
Smoothing gives us for any t.
- most likely state is not quite the same as the most likely path …
Viterbi will give us most likely path:
![Page 47: Hidden Markov Modelsweb.eecs.umich.edu/~jjcorso/t/2011S_555/files/platt_hmm.pdf · Hidden Markov models are everywhere! Used to model time-series data: - Robot perception/control](https://reader030.vdocuments.site/reader030/viewer/2022041107/5f09e6b67e708231d4290bb6/html5/thumbnails/47.jpg)
Viterbi
Find:
Define:
Compare with:
![Page 48: Hidden Markov Modelsweb.eecs.umich.edu/~jjcorso/t/2011S_555/files/platt_hmm.pdf · Hidden Markov models are everywhere! Used to model time-series data: - Robot perception/control](https://reader030.vdocuments.site/reader030/viewer/2022041107/5f09e6b67e708231d4290bb6/html5/thumbnails/48.jpg)
Filtering: dynamic programming sol'n
Find:
Define:
1.
2.
3.
![Page 49: Hidden Markov Modelsweb.eecs.umich.edu/~jjcorso/t/2011S_555/files/platt_hmm.pdf · Hidden Markov models are everywhere! Used to model time-series data: - Robot perception/control](https://reader030.vdocuments.site/reader030/viewer/2022041107/5f09e6b67e708231d4290bb6/html5/thumbnails/49.jpg)
Viterbi
Find:
1.
2.
3.
4.
Define:
![Page 50: Hidden Markov Modelsweb.eecs.umich.edu/~jjcorso/t/2011S_555/files/platt_hmm.pdf · Hidden Markov models are everywhere! Used to model time-series data: - Robot perception/control](https://reader030.vdocuments.site/reader030/viewer/2022041107/5f09e6b67e708231d4290bb6/html5/thumbnails/50.jpg)
Viterbi
Define:
Shorthand:
![Page 51: Hidden Markov Modelsweb.eecs.umich.edu/~jjcorso/t/2011S_555/files/platt_hmm.pdf · Hidden Markov models are everywhere! Used to model time-series data: - Robot perception/control](https://reader030.vdocuments.site/reader030/viewer/2022041107/5f09e6b67e708231d4290bb6/html5/thumbnails/51.jpg)
Viterbi recursion
probability of most likely path to followed by
Recursion:
![Page 52: Hidden Markov Modelsweb.eecs.umich.edu/~jjcorso/t/2011S_555/files/platt_hmm.pdf · Hidden Markov models are everywhere! Used to model time-series data: - Robot perception/control](https://reader030.vdocuments.site/reader030/viewer/2022041107/5f09e6b67e708231d4290bb6/html5/thumbnails/52.jpg)
Kevin Murphy, Kalman filter toolbox
Filtering smoothing example: Kalman filter
![Page 53: Hidden Markov Modelsweb.eecs.umich.edu/~jjcorso/t/2011S_555/files/platt_hmm.pdf · Hidden Markov models are everywhere! Used to model time-series data: - Robot perception/control](https://reader030.vdocuments.site/reader030/viewer/2022041107/5f09e6b67e708231d4290bb6/html5/thumbnails/53.jpg)
Kevin Murphy, Kalman filter toolbox
Filtering smoothing example: Kalman filter
![Page 54: Hidden Markov Modelsweb.eecs.umich.edu/~jjcorso/t/2011S_555/files/platt_hmm.pdf · Hidden Markov models are everywhere! Used to model time-series data: - Robot perception/control](https://reader030.vdocuments.site/reader030/viewer/2022041107/5f09e6b67e708231d4290bb6/html5/thumbnails/54.jpg)
Kevin Murphy, Kalman filter toolbox
Filtering smoothing example: Kalman filter
![Page 55: Hidden Markov Modelsweb.eecs.umich.edu/~jjcorso/t/2011S_555/files/platt_hmm.pdf · Hidden Markov models are everywhere! Used to model time-series data: - Robot perception/control](https://reader030.vdocuments.site/reader030/viewer/2022041107/5f09e6b67e708231d4290bb6/html5/thumbnails/55.jpg)
HMM Learning (overview)
Easy once you know filtering/smoothing
- given a series of observations, learnthe parameters of the HMM,
Find the parameters, ,that maximize:
![Page 56: Hidden Markov Modelsweb.eecs.umich.edu/~jjcorso/t/2011S_555/files/platt_hmm.pdf · Hidden Markov models are everywhere! Used to model time-series data: - Robot perception/control](https://reader030.vdocuments.site/reader030/viewer/2022041107/5f09e6b67e708231d4290bb6/html5/thumbnails/56.jpg)
HMM Learning (overview)
1. Given, , do forward/backward to calculate
![Page 57: Hidden Markov Modelsweb.eecs.umich.edu/~jjcorso/t/2011S_555/files/platt_hmm.pdf · Hidden Markov models are everywhere! Used to model time-series data: - Robot perception/control](https://reader030.vdocuments.site/reader030/viewer/2022041107/5f09e6b67e708231d4290bb6/html5/thumbnails/57.jpg)
HMM Learning (overview)
2. Given , calculate expected :
- expected value of :
- expected value of :
![Page 58: Hidden Markov Modelsweb.eecs.umich.edu/~jjcorso/t/2011S_555/files/platt_hmm.pdf · Hidden Markov models are everywhere! Used to model time-series data: - Robot perception/control](https://reader030.vdocuments.site/reader030/viewer/2022041107/5f09e6b67e708231d4290bb6/html5/thumbnails/58.jpg)
HMM Learning (overview)
3. Goto step 1, using new values of
This procedure is an application of EM(expectation maximization) known as theBaum-Welch algorithm.
- procedure converges to local optimum- not necessarily globally optimal
![Page 59: Hidden Markov Modelsweb.eecs.umich.edu/~jjcorso/t/2011S_555/files/platt_hmm.pdf · Hidden Markov models are everywhere! Used to model time-series data: - Robot perception/control](https://reader030.vdocuments.site/reader030/viewer/2022041107/5f09e6b67e708231d4290bb6/html5/thumbnails/59.jpg)
HMM Learning (overview)
Baum-Welch can be viewed as coordinateascent:
Roweis, Ghahramani, 2001
![Page 60: Hidden Markov Modelsweb.eecs.umich.edu/~jjcorso/t/2011S_555/files/platt_hmm.pdf · Hidden Markov models are everywhere! Used to model time-series data: - Robot perception/control](https://reader030.vdocuments.site/reader030/viewer/2022041107/5f09e6b67e708231d4290bb6/html5/thumbnails/60.jpg)
Summary
1. Markov Processes
2. Hidden Markov Models
3. Filtering (forward algorithm)
4. Smoothing (backward algorithm)
5. Viterbi (most likely path)
6. Parameter learning
![Page 61: Hidden Markov Modelsweb.eecs.umich.edu/~jjcorso/t/2011S_555/files/platt_hmm.pdf · Hidden Markov models are everywhere! Used to model time-series data: - Robot perception/control](https://reader030.vdocuments.site/reader030/viewer/2022041107/5f09e6b67e708231d4290bb6/html5/thumbnails/61.jpg)
References
1. Rabiner, A tutorial on hidden Markov models and selected applications in speech Recognition, Proc. of the IEEE, 1989
2. Bishop, Pattern recognition and machineLearning, Springer, Ch 13, 2006