CSCI-561 Foundations of Artificial Intelligence - Homework 3: Temporal Reasoning in Artificial Intelligence
Overview
CSCI-561 - Fall 2023 - Foundations of Artificial Intelligence
Homework 3
Due Time: Monday, November 20th, 2023, 23:59:59 PST
This homework explores the applications of Temporal Reasoning in Artificial Intelligence. In general, the solution for a temporal reasoning task involves taking a sequence of actions/observations on an Partially Observable Markov Decision Process (POMDP Environment) , applying a temporal-reasoning algorithm that you learned from this class, and returning the most probable sequence of the hidden states that the POMDP most-likely went through when experiencing the given sequence of actions/observations.
More specifically, this assignment provides you with two versions of temporal data: a base version involving the “Little Prince” Environment and an advanced version that revolves around speech recognition and text prediction.
Scenario 1 : Little Prince Model
The setup in this version is very similar to the “Little Prince” environment shown in Figure 1, presented in the lecture notes and in the optional reading textbook (ALFE).
Figure 1: The Little Prince POMDP (lecture 08-09 and 22).
You will be given a list of available percepts, actions and states and the corresponding initial state weightages, transition and observation weight values in that environment (More on the input structure will be covered in the sections below) .<rose, forward, none, ..., turn, rose, backward, volcano, ...>
Then, your program should return a sequence of hidden-states that this POMDP is most-likely going through (The following sequence is an example of the final state sequence that one might encounter):
<s3, s4, s2, s3, ...>
Figure 2: The inputs and outputs of a temporal-reasoning task (lecture 22).
More Details on solving this state sequence prediction problem can be found in the lecture slides. The input file format for this environment will be the same as the Speech Recognition environment. You will be given weight values instead of probabilities and the process of converting weights to probabilities is explained in detail in the following sections.
Scenario 2 : Simplified Speech Recognition
This scenario deals with a more sophisticated environment of speech recognition - without the hassle of going through audio signal processing. This model will primarily focus on resolving the ambiguity introduced by multiple plausible texts that could correspond to a single spoken utterance. For every word pronunciation, instead of dealing with audio signals, you will be given a set of phonemes which phonetically represent the word, and will produce a series of text fragments the same length as the sequence of phonemes. The following table provides some sample words and their phoneme / fragment mapping for better understanding.
|
Example Words |
Phoneme Mapping |
Fragment Mapping |
|
water |
W | AO1 | T | ER0 |
w | a | t | er |
|
human |
Y | UW1 | M | AH0 | N |
h|u|m|a|n |
|
ocean |
OW1 | SH | AH0 | N |
o | c | ea | n |
As evident in the table, there is a 1:1 correspondence with phoneme and fragment for a given word. For
the word water, fragment “A” corresponds to the phoneme “AO1” and fragment “er” corresponds to
“ER0”.
Because this project builds off prior work from the CMU Pronouncing Dictionary, we use a dialect of
English known as North American English, where e.g. human is pronounced with a “Y” sound at the start.
The ultimate goal of this environment is to find the best sequence of text fragments for a given sequence of phonemes. Formally, we will represent the process as a POMDP, with the text fragments corresponding to states and the phonemes corresponding to observations. For convenience, we will use a single null action “N”. This will ensure that we are using a Partially Observable Markov Decision Process instead of just a Partially Observable Markov Process. This will also allow you to re-use parsing code between the Little Prince environment and the Speech Recognition environment.
As part of the input, you will be provided with a dataset containing a list of fragment to phoneme pairs, along with a weight value for each pair. You will also be given fragment-to-fragment transition pairs with their weight values. The following example explains the procedure to compute the probability tables in more detail. These weights correspond to un-normalized total probabilities P(o, s) and P(s, s’) (using the null action “N”), which were computed by counting (observation, state) and (state, state) pairs from a set of approximately 300k Wikipedia articles. Note that you will need to normalize these weights into the appropriate probability distributions P(o | s) and P(s’ | s, N).
Consider the following dataset:
Fragment to Phoneme Mapping Fragment to Fragment Transition Mapping
|
Fragment |
Phoneme |
Weight |
|
s |
S |
100 |
|
er |
ER0 |
10 |
|
o |
AH0 |
10 |
|
e |
AH0 |
20 |
|
Fragment |
Fragment |
Weight |
|
s |
er |
80 |
|
s |
o |
10 |
|
er |
o |
5 |
|
o |
e |
8 |
Construction of Probability Tables :
● Initial State Probability: Computing the initial state / prior state distribution only involves
dividing each weight in the state probability table by the total weight in the table.
Given a state table as follows:
|
State |
s |
er |
o |
e |
|
Weight |
1 |
1 |
1 |
1 |
The initial probability P(s) is:
● State Transition Probability: This can be determined by looking at the weights of the transitions from one fragment to another and calculating the probability through normalization for each (state, action) pair. This produces P(s|s,a). In this example, there is only a single valid action, so we don’t show it in the table.
The state transition probability for the sample dataset would be:
|
State |
s |
er |
o |
e |
|
Prob |
0.25 |
0.25 |
0.25 |
0.25 |
|
States |
s |
er |
o |
e |
|
s |
0.0 |
0.8 |
0.1 |
0.1 |
|
er |
0 |
0 |
1 |
0 |
|
o |
0 |
0 |
0 |
1 |
|
e |
0 |
0 |
0 |
0 |
● Appearance Probabilities: This can be inferred from the fragment-phoneme pairs through normalization over all weights for a given state, producing the conditional distribution P(o|s). The appearance probability for the above example would be as follows:
|
S |
Z |
ER0 |
AH0 |
|
|
s |
0.667 |
0.333 |
0 |
0 |
|
er |
0 |
0 |
1 |
0 |
|
o |
0 |
0 |
0 |
1 |
|
e |
0 |
0 |
0 |
1 |
Note how the row for state “s” sums to 1 over the different possible observations / phonemes “S” and “Z”.
Your algorithm will be tested on a list of phonemes for which it should provide the most probable sequence of fragments.
Input / Output Format
As mentioned in the problem statement above, there will be two types of input to indicate two modes / stages : Little Prince environment and Simplified Speech Recognition.
In both cases, you will be given a set of input files, containing the table of weights described above. You will need to parse and normalize these tables into the appropriate conditional probabilities. weight. Then, each file contains a sequence of entries, where states, observations, and actions are wrapped in double quotes, and the weights are specified as integers. At the end of the file, there will be one final newline.
There are three weight table files:
The first of these files contains weights for every state, and describes the prior probability of each state P(s). (The default weight value is present in this file just to ensure format consistency across all weight files - this will not be required to be used, as all states will be present in the state weights file)
state_weights.txt:
The second of these files contains weights for (state, action, state) triples, and describes the probability of state transitions P(s, a, s). Triples not specified in the table should be given the weight default weight specified on the second line. Note that you will need to normalize these weights into the appropriate probability distribution P(s’ | a, s).
state_action_state_weights.txt:
The third of these files contains weights for every (state, observation) pair, and describes the probability of each state observation pair P(s, o). Pairs not specified in the table should be given the default weight specified on the second line. Note that you will need to normalize these weights into the appropriate probability distribution P(o | s).
state_observation_weights.txt:
state_weights
<number of states> <default weight>
“state1” <weight of state1>
“state2” <weight of state2>
etc...
state_action_state_weights
<number of triples in file> <number of unique states> <number of unique
actions> <default weight>
“state1” “action1” “next state1” <weight of (state1, action1, next state1)>
“state2” “action2” “next state2” <weight of (state2, action2, next state2)>
etc...
state_observation_weights
<number of pairs in file> <number of unique states> <number of unique
observations> <default weight>
“state1” “observation1” <weight of (state1, observation1)>
“state2” “observation2” <weight of (state1, observation1)>
etc...
Lastly, you will receive a file containing the sequence of (observation, action) pairs, on which you should run the Viterbi algorithm.
observation_actions.txt:
Your code will be expected to produce a file containing the predicted state sequence.
states.txt:
Sample Test Case
The following sample test case corresponds to the Little Prince Environment (The format matches that used for Simplified Speech Recognition Environment)
Little Prince Environment Test case:
state_weights.txt state_observation_weights.txt
observation_actions <number of pairs in file> “observation1” “action1” “observation2” “action2” etc...
states
<length of state sequence>
“state1”
“state2”
"S0" "Apple" 2 "S1" "Volcano" 5 "S1" "Grass" 5 "S1" "Apple" 2 "S2" "Volcano" 3 "S2" "Grass" 5 "S2" "Apple" 2
state_weights
30
"S0" 2
"S1" 5
"S2" 5
state_action_state_weights.txt
state_action_state_weights
27 3 3 0
"S0" "Forward" "S0" 3
"S0" "Forward" "S1" 3
"S0" "Forward" "S2" 2
"S1" "Forward" "S0" 4
"S1" "Forward" "S1" 5
"S1" "Forward" "S2" 1
"S2" "Forward" "S0" 1
"S2" "Forward" "S1" 5
"S2" "Forward" "S2" 4
"S0" "Backward" "S0" 5
"S0" "Backward" "S1" 5
"S0" "Backward" "S2" 5
"S1" "Backward" "S0" 3
"S1" "Backward" "S1" 4
"S1" "Backward" "S2" 1
"S2" "Backward" "S0" 2
"S2" "Backward" "S1" 3
"S2" "Backward" "S2" 3
"S0" "Turnaround" "S0" 5
"S0" "Turnaround" "S1" 3
"S0" "Turnaround" "S2" 5
"S1" "Turnaround" "S0" 3
"S1" "Turnaround" "S1" 4
"S1" "Turnaround" "S2" 2
"S2" "Turnaround" "S0" 1
"S2" "Turnaround" "S1" 2
"S2" "Turnaround" "S2" 2
observation_actions.txt
Output:
states.txt
observation_actions
4
"Apple" "Turnaround"
"Apple" "Backward"
"Apple" "Forward"
"Volcano"
states 4 "S2" "S2" "S2" "S1"
Input Constraints & Time Limits Little Prince Environment:
|
Maximum Number of States in the Environment |
10 states |
|
Maximum Number of Percepts in the Environment |
10 percepts |
|
Number of Actions Possible in the Environment (Fixed) |
3 (“Forward”, “Backward”, “Turnaround”) |
|
Maximum Length of Observation Action Sequence |
20 |
|
Time Limit allowed per test case |
1 second |
|
Number of test cases |
20 ( 5 Preliminary and 15 hidden) |
Simplified Speech Recognition Environment:
|
Maximum Number of States in the Environment |
682 |
|
Maximum Number of Observations in the Environment |
69 |
|
Number of Actions Possible in the Environment (Fixed) |
“N” ( Indicates Null Action, Refer Scenario Description ) |
|
Maximum Length of Observation Action Sequence |
100 steps |
|
Time Limit allowed per test case |
1 minute |
|
Number of test cases |
30 ( 5 Preliminary and 25 hidden) |
Grading Criteria
There will be 50 test cases in total : 20 cases with the Little Prince environment and 30 cases with Speech Recognition. Each test case is worth 2 points. On clicking the submit button in Vocareum, Your solution will be run on preliminary test cases. Your code will be evaluated on the hidden test cases after the assignment deadline.
The performance of your program will be computed automatically by comparing your output sequence with the most-likely known sequence, and the matching percentage will determine the grade of your submissions.More specifically, the probability of the hidden state sequence that your solution has generated (p1), will be compared with the probability of the hidden state sequence proposed by TA agent (p2) and p1/ p2 will be used as the final score for each test case.
Score for a single test case = 2 * Probability of student’s state sequence Probability of TA agent’s state sequence
Final Score = Sum of scores across all test cases
NOTE : The Max Score for a single test case will be capped at 2 points.
The probability of the state sequence will be calculated based on the joint probability of the state sequence, along with the action sequence and the observation sequence.
In the following example [used for calculation purposes only], the score will be calculated as follows:
States List = [S0, S1, S2], Actions Set= [N], Observations set= [A, B] Initial State Probability
Appearance Probability
Transition Probability
Observation Action Sequence = [A, <N>, B,<N>, B]
If the student’s predicted state sequence is [S0, S1, S1], then
Student’s Probability = π(𝑆0)*P(A|S0)*P(S1|S0,<N>)*P(B|S1)*P(S1|S1,<N>)*P(B|S1)
= 0.5*0.9*0.3*0.8*0.1*0.8 = 0.00864
If the TA’s predicted state sequence is [S0, S1, S0], then
TA’s Probability = π(𝑆0)*P(A|S0)*P(S1|S0,<N>)*P(B|S1)*P(S0|S1,<N>)*P(B|S0)
= 0.5 * 0.9 * 0.3 *0.8 * 0.9 * 0.1 = 0.00972
Your Score for this test case = 2 * 0.00864 / 0.00972 = 1.778
|
S0 |
S1 |
S2 |
|
0.5 |
0.25 |
0.25 |
|
A |
B |
|
|
S0 |
0.9 |
0.1 |
|
S2 |
0.5 |
0.5 |
|
S0 |
S2 |
|
|
S0 |
0.5 |
0.2 |
|
S1 |
0.9 |
0 |
|
S2 |
0.2 |
0 |
