Course Review: Artificial Intelligence
What the Course Is About
"Artificial Intelligence is the art of creating machines that perform functions that require intelligence when performed by people." - Ray Kurzweil
The course covers a wide range of topics in the field of AI tracking its development up until the early 2000s. Some of the topics covered include:
- Graph search algorithms
- Simulated annealing
- Genetic algorithms
- Multi-agent Games
- Algorithmic Constraint Satisfaction
- Bayes Networks
- Machine Learning (incl. CNNs)
- Hidden Markov Models
- First-Order Logic Models
- Partially Observable Stochastic Search
- Reinforcement Learning
The course doesn't cover more recent developments including:
- Self Supervised Learning
- Generative Adversarial Networks
- Synthetic Data Generation
- Large Language Models (transformer architecture)
Personal Note
I have taken this course in Spring 2024 along with AOS. Prior to this course I have taken the following 4 OMSCS courses: GIOS, DBMS, CN, AIES.
Graded Deliverables
Here's the grading scheme breakdown:
- Plagiarism Quiz - 4%
- Assignment 0 - 1%
- Assignment 1 - 12%
- Asisgnment 2 - 12%
- Asisgnment 3 - 12%
- Asisgnment 4 - 12%
- Asisgnment 5 - 12%
- Midterm Exam - 15%
- Final Exam - 20%
Plagiarism Quiz
This is an easy 4% that you can get by simply reading the "class policies" and answering a few questions.
Assignments
The course has 6 assignments, but only the best 5 count towards your final grade. Submission for all assignments is done through Gradescope. All assignments are implemented in Python and most provide a Jupyter Notebook template. If you have access to Georgia Tech's Github, you can checkout all assignment repos
Assignment 0: Warmup
This assignment ensures that you have setup the necessary pip dependencies for the other assignments. This assignment also contains a non-mandatory part on Jupyter Notebooks, Python and Git.
Time spent: ~1 hour
Assignment 1: Tridirectional Search
You are given a graph of cities and have to implement a couple of search algorithms. The algorithms get increasingly more difficult starting with BFS, UCS, A*, bidirectional Search and ending with tridirectional search. The lectures and resources provide a good foundation for implementing these algorithms. However, the final algorithm (an optimized version of tridirectional search) is quite complex and very time consuming. It requires reading through some research papers and translating the concepts into code. I gave up after spending ~2 hours on this part as it only accounts for only 8% of the assignment grade.
Time spent: ~10 hours
Assignment 2:
The goal of this assignment is to create an AI that can play the game of Isolation. Isolation is introduced in the lectures on Game Playing. It seems like they vary the rules of the game in every semester or every couple of semesters slightly. Your AI will use the Minimax algorithm with alpha-beta pruning and iterative deepening to play the game. All of these concepts are covered in the lectures on Game Playing. Once I understood the concepts in the lectures, I found the implementation quite straightforward. The Jupyter notebook provides some really nice visualizations for the game board and the AI's moves. The grading for this assignment is based on the percentage of games your AI wins against a hidden implementation from the course staff. To get the final few percent you will need to use some more advanced techniques like tweaking the evaluation function for the variation of the game they provide. To get the full 100% you may need to get a little fancy with your algo, I got 95% by just implementing the basics suggested. I'm not sure about the effort/reward ratio of getting the final few percent. However, if you do go for it, you can enter a bonus competition against the AIs of your peers. My gut feeling tells me this would be very comptetitive.
Time spent: ~8 hours
Assignment 3
Assignment 3 requires you to build a Bayes Network and implement a sampling algorithm to answer probabilistic queries on the network.
First, you setup the network using the Python library pgmpy.
Once you figure out the library, this part is quite straightforward.
Then you answer some questions about the network by calculating probabilities with pgmpy.
The challenging part of the assignment is implementing the sampling algorithms.
You implement both the Gibbs Sampling and Metropolis-Hastings algorithms.
I found this part very challenging and almost gave up on it.
Reading through the resources and understanding the concepts was key to solving this assignment.
Also, some students posted really useful clarifications on Ed which helped me in cracking the assignment.
The implementation in Python is easy, but understading all the moving parts of the sampler is hard.
Time spent: ~12 hours
Assignment 4
In this assignment you build a multi-class classifier by implementing a random forest. First, you implement some helper functions to calculate statistical measures for accuracy, precision, recall, and confusion matrix. Next, you implement functions to calculate the given statistical measures for entropy, information gain and gini impurity. Then you implement a single decision tree and finally a random forest. Overall, I didn't find any major difficulties with this assignment. It's just about implementing the algorithms.
Time spent: ~8 hours
Assignment 5
You implement a k-means clustering algorithm and a Gaussian Mixture Model.
This assignment requires you to implement many individual functions in order to build up to the final algorithms.
Each function isn't too difficult, but there are many of them.
Furthermore, you have to implement all functions in a vectorized way using numpy, otherwise your implementation will
timeout on Gradescope.
Whenever you're writing a for loop to iterate over a matrix there's probably a more optimized way using numpy.
I struggled a lot with GS timeouts in the beginning due to my non-optimal implementation.
For the final part, you implement the Bayesian Information Criterion to determine the optimal number of
clusters for the Gaussian Mixture Model.
I couldn't get this part to work properly and had to skip it.
I found this assignment to be the most challenging of all the assignments as it was time intensive, required a fair bit of statistical background and required efficient implementations of matrix operations using numpy.
Time spent: ~12 hours
Assignment 6
I didn't do this assignment, since I already had 5 assignments with grades above 90%. It involves Hidden Markov Models and according to some students it isn't too difficult in case you have a bad grade on one of the other assignments. Possibly, you could swap out assignment 3 or 5 for this one to reduce the overall workload. But, it could be a bit risky as you won't have any backup assignments if you don't get a good grade on this one.
Time spent: 0 hours
Bonus Assignments
The first 3 assignments have some bonus parts to get extra credit. To me these bonus assignments looked pretty interesting, but appeared as massive time sinks for the amount of extra credit they provide. If you have the time, they will probably cement your understanding of the concepts covered in the assignments. But it's definitely not required. I have personally not attempted any of these bonus parts.
Exams
The course has two take home exams which are open book, but not open internet. You have about 1 week to complete each exam. They are very fair in that they cover nothing outside of the lectures. The format rewards effort and takes out any luck factor compared to writing regular exams. In general, the more time you spend on the questions, rewatching lectures, reading the textbook, the more likely you are to get a good grade. However, they are very time consuming. Both exams had some errors which were corrected by the course staff on Ed. Also they posted a lot of clarifications on Ed, so I would definitely monitor that.
Midterm
The midterm covers lecture content from 1-6.
- Lecture 1: Search
- Lecture 2: Simulated Annealing
- Lecture 3: Game Playing
- Lecture 4: Constraint Satisfaction
- Lecture 5: Probability
- Lecture 6: Bayes Nets
I found that 15% for the midterm was a bit low when compared to a single assignment which is worth 12.5%. If you really want to ace the exam, rewatching the lectures and reading the textbook would be necessary. However, this is a lot of effort and wasn't worth it for me.
Time spent on the midterm: ~8 hours
Final
The final exam covers all content from the midterm as well as the remaining lectures:
- Lecture 7: Machine Learning
- Lecture 8: Pattern Recognition Through Time
- Lecture 9: Logic & Planning
- Lecture 10: Planning under Uncertainty
Around 45% of questions cover lessons 1-6 and are similar to the midterm. The remaining 55% of the questions cover the remaining lectures. Had I known this distribution, I would have spent more time on the earlier lectures as it gives a double benefit. which would be useful for both the midterm and almost half of the final.
Time spent on the final: ~11 hours
Grading Curve
The course grading is on a curve. Here's the distribution for Spring 24:
- A: > 90%
- B: > 76%
- C: > 61%
- D: > 46%
My Grades
- Overall Grade: 89.92% (B - so close to an A!)
- Plagiarism Quiz: 79%
- Assignment 0: 100%
- Assignment 1: 91%
- Assignment 2: 95%
- Assignment 3: 100%
- Assignment 4: 100%
- Assignment 5: 92%
- Midterm: 76%
- Final: 78%
The Course Materials
Lecture Videos & Content
You can find the lectures on Ed with your GT login. I believe some of the lectures are straight from some older MOOCs. The course has a couple of different lecturers: Peter Norvig, Sebastian Thrun, and Thad Starner. Most lectures are held by Thad Starner who is joined by a student. The lecture style between the two is conversational which I found to be engaging. The content itself is very interesting. However, the videos are about a decade old. Furthermore, the course doesn't cover any of the recent developments in AI.
Textbook
The textbook in this course is "Artificial Intelligence: A Modern Approach" by Stuart Russell & Peter Norvig. I believe this is the classic book on AI and is used in many AI courses around the world. I personally didn't read much of the book, since I found the lectures to be sufficient. The book is a great supplementary resource. However, it is definitely not required and in some cases I found it to be more mathematical than is required for the course.
Personal Opinion on AI
My Background going into the course
I have a decent background in programming, algorithms and Python. I have taken undergraduate courses in probability and statistics. However, I didn't have any experience with traditional AI as taught in this course.
I think the course is pretty self contained, but some algorithm experience & statistics will be helpful.
How Hard Was The Course
I didn't find it as hard as expected. It's certainly harder than some of the other courses I have taken including AIES, DBMS, and CN. But it's not as hard as GIOS or AOS.
What I Liked
- Choosing Python as the programming language for the assignments
- Very cool AI algorithms and concepts
- Little busy work (compared to courses like AIES)
- Possible to work ahead
- Well-designed assignments
What I Disliked
- It would be nice to have at least some pointers to resources on the current state of the art in AI.
- The final few lectures didn't have any corresponding assignments. I think it would have been good to have at least one assignment to cover reinforcment learning for example. I really enjoyed that the other lectures had corresponding assignments which kept me motivated and cemented my understanding of the topics.
- The exams needed many clarifications on Ed and had constant regrades (e.g. my final grade went from 66 to 78).
Should You Take This Class
- Are you looking for an easy course with minimal workload this semester? There are easier courses out there.
- Is it your first semester in OMSCS? This might not be the best course to start with.
- Are you looking for a course that covers a lot of ground on AI - not just the hyped stuff? Go for it!
Tips for Success
How I would approach the course if I were to take it again:
- Focus energy on the assignments
- Once you hit >90% on an assignment, time box the rest of the assignment
- Do 5 out of the 6 assignments (only the best 5 are counted). Skip assignment 3, 5 to minimize total time spent.
- Watch the lectures for high level understanding
- Focus a disproportionate amount of time on the midterm - you'll be rewarded in both midterm & final