Description
Problem 1: Robot Localization (30 points)
A robot is wandering around a room with some obstacles, labeled as # in the grid below. It can
occupy any of the free cells labeled with a letter, but we are uncertain about its true location and
thus keep a belief distribution over its current location. At each timestep it moves from its current
cell to a neighboring free cell in one of the four cardinal directions with uniform probability; it
cannot stay in the same cell.
For example, from A the robot can move to either B or C with
probability 1
2
, while from D it can move to B, C, E, or F, each with probability 1
4
.
A B #
C D E
# F #
The robot may also make an observation after each transition, returning what it sees in a random
cardinal direction. Possibilities include observing #, “wall”, or “empty” (for a free cell). For
example, in B the robot observes “wall”, # (each with probability 1
4
), or “empty” (probability 1
2
).
(a) Suppose the robot wanders around forever without making any observations. What is the stationary distribution π over the robot’s predicted location? Hint: You can use numpy.linalg.eig
in Python. The first return value is a 1D array of eigenvalues; the second return value is a 2D
array, where each column is a corresponding eigenvector. Remember that eigenvectors may not
sum to 1 by default.
(b) Now suppose that we know that the robot is in state D; i.e., Pr(X0 = D) = 1. Starting from
this state, the robot makes one transition and observes e1 = #. What is the updated belief
distribution Pr(X1 | e1)?
(c) The robot makes a second transition and observes e2 = “empty”. What is the updated belief
distribution Pr(X2 | e1, e2)?
(d) Compute the joint distribution Pr(X1, X2 | e1, e2). You do not need to explicitly list the values
that have probability 0. What is the most likely state sequence(s)?
(e) Compute b = Pr(e2 | X1). Briefly explain what this quantity represents.
(f) Compute the smoothed distribution Pr(X1 | e1, e2) by multiplying f = Pr(X1 | e1) with
b = Pr(e2 | X1) and normalizing the result. Confirm that the distribution is the same as that
obtained from marginalization of Pr(X1, X2 | e1, e2).
Problem 2: Descendants of Effects (20 points)
We will investigate the absence of conditional independence guarantees between two random variables when an arbitrary descendant of a common effect is observed. We will consider the simple
case of a causal chain of descendants:
Suppose that all random variables are binary. The marginal distributions of A and B are both
uniform (0.5, 0.5), and the CPTs of the common effect D0 and its descendants are as follows:
A B Pr(+d0 | A, B)
+a +b 1.0
+a −b 0.5
−a +b 0.5
−a −b 0.0
Di−1 Pr(+di
| Di−1)
+di−1 1.0
−di−1 0.0
(a) Give an analytical expression for the joint distribution Pr(D0, D1, · · · , Dn). Your expression
should only contain CPTs from the Bayes net parameters. What is the size of the full joint
distribution, and how many entries are nonzero?
(b) Suppose we observe Dn = +dn. Numerically compute the CPT Pr(+dn|D0). Please show how
you can solve for it using the joint distribution in (a), even if you do not actually use it.
(c) Let’s turn our attention to A and B. Give a minimal analytical expression for Pr(A, B, D0, +dn).
Your expression should only contain CPTs from the Bayes net parameters or the CPT you found
in part (b) above.
(d) Lastly, compute Pr(A, B | +dn). Show that A and B are not independent conditioned on Dn.
Problem 3: Part-of-Speech Tagging
In this problem you will explore part-of-speech (POS) tagging, a standard task in natural language
processing. The goal is to identify parts of speech and related labels for each word in a given
corpus. Hidden Markov models are well suited for this problem, with parts of speech being hidden
states and the words themselves being observations.
We will be using data from the English EWT
treebank from Universal Dependencies, which uses 17 POS tags. We are providing clean versions
of training and test data for you. The data format is such that each line contains a word and
associated tag, and an empty line signifies the end of a sentence. Feel free to open the files in a
text editor to get an idea.
The provided Python file contains a couple of functions that can be used to read a data file (you
do not need to call these yourself). The global variable POS contains a list of all possible parts
of speech. You will be filling in the remaining functions in the file and running the code in main
where instructed.
3.1: Supervised Learning (12 points)
Your first task is to learn the three sets of HMM parameters from the training data. The initial
distribution Pr(X0) will be stored in a 1D array of size 17. The transition probabilities will be
stored in a 2D array of size 17 × 17. The observation probabilities will be stored in a dictionary,
where each key is a word and the value is a 1D array (size 17) of probabilities Pr(word | POS).
These probabilities should follow the same order as in the POS list.
Implement learn model so that it compiles and returns these three structures. The data input is
a list of sentences, each of which is a list of (word, POS) pairs. Your method should iterate over
each sentence in the training data, counting the POS appearances in the first word, the number
of POS to POS transitions, and the number of POS to word observations. Treat each sentence
independently; do not count transitions between different sentences.
Be sure to correctly normalize all of these distributions. Make sure that the quantities P
i Pr(X0)i
P ,
i Pr(Xt+1 = i | Xt = j), and P
k Pr(Et = k | Xt) are all equal to 1.
3.2: Viterbi Algorithm (12 points)
The next task is to implement the Viterbi algorithm to predict the most likely sequence of states
given a sequence of observations. We will break this into two pieces: viterbi forward and
viterbi backward. viterbi forward takes in four parameters: initial distribution X0 (1D array),
transition matrix Tprobs (2D array), observation probabilities Oprobs (dictionary), and observation sequence obs (list of word observations) of length T.
viterbi forward should compute and return two quantities: maxx1,…,xT −1 Pr(x1, …, xT −1, XT , e1:T )
as a 1D array of size 17, and a T ×17 2D array of pointer indices. Row i of the pointer array should
contain argmaxxi−1
Pr(x1, …, xi−1, Xi
, e1:i).
For simplicity, pointers will be the indices of the POS
in the POS array rather than strings. Note that it is possible for an observation e to not exist in
the Oprobs dictionary if it was not present when the model was trained. If this occurs, you may
simply take Pr(e | x) = 1 for all POS x; this is equivalent to skipping the observation step.
viterbi backward takes the two quantities returned by viterbi forward as parameters. It should
start with the most likely POS according to maxx1,…,xT −1 Pr(x1, …, xT −1, XT , e1:T ) and then follow
the pointers backward to reconstruct the entire POS sequence argmaxx1,…,xT
Pr(x1, …, xT | e1:T ).
The returned object should be a list of POS (strings) from state 1 to state T. Note that the predicted
state for X0 should not be included, so your list should be length T.
3.3: Model Evaluation (8 points)
Your Viterbi implementation can now be used for prediction. evaluate viterbi takes in the HMM
parameters, along with a data list in the same format as in learn model. Complete the function so
that it runs Viterbi on each sentence separately on the data set. Then compare all returned POS
predictions with the true POS in data. Compute and return the accuracy rate as the proportion
of correct predictions (this should be a number between 0 and 1). After implementing this function,
run the associated code in main and answer the following questions.
(a) Report the accuracies of your Viterbi implementation on each data set. Why is the accuracy
on the test data set lower than that of the training set?
(b) Why can we not expect 100% accuracy on the training set, despite defining the HMM parameters to maximize the likelihood on the training data set?
3.4: Forward and Backward Algorithms (12 points)
Next, implement the the forward, backward, and forward-backward algorithms (ideally fewer than
5 lines of code each). forward takes in the same four parameters as viterbi forward above, and
it computes and returns Pr(Xk, e1:k) (no normalization). backward also takes in four parameters,
dropping X0 but newly including the state index k. It should compute and return Pr(ek+1:T | Xk).
Note that k follows Python indexing; in other words, k=0 corresponds to X1 and backward should
return Pr(e2:T | X1). Again, to deal with the scenario in which a word is not in the observation
model, you should explicitly check when this occurs and if so simply use Pr(e | x) = 1.
Once you have both forward and backward, it should be straightforward to call these to implement
forward backward, which computes the smoothed state distribution Pr(Xk | e1:T ). Note that the
call to forward should only use the observations up to (and including) the kth one. Don’t forget
to normalize the smoothed distribution.
3.5: Inference Comparisons (6 points)
Now that you have all of these inference algorithms implemented, come up with a short English
phrase p, ideally around 10 words or fewer. You should then identify a word w within this phrase,
such that (i) the most likely POS of w according to the result of the forward algorithm on the
first portion of p up to word w, and (ii) the most likely POS of w according to the result of the
forward-backward algorithm on the entirety of p, are different.
Leave the code that you used to obtain these observations in the main function of the code file. In
your writeup, give the phrase that you used and the word that you evaluated. Indicate the POS
returned by each of the results as described above. Give a brief explanation about why each of the
methods returns something different.
Submission
You should have one PDF document containing your solutions for problems 1-2, as well as your
responses to 3.3 and 3.5. You should also have a completed Python file implementing problem 3;
make sure that all provided function headers and the filename are unchanged. Submit the document
and .py code file to the respective assignment bins on Gradescope. For full credit, you must tag
your pages for each given problem on the former.
