Cool AI - Cool stuff for teaching AI

About this web page

This web page contains several project assignments in Artificial Intelligence for use by anyone teaching a course in AI at the college level. Each project has been used in teaching AI at the University of Southern California and has been vetted a little. The author for each project is T. Nathan Mundhenk, Laurent Itti and Michael Arbib.

For each project, we have attempted to include all the material one would need to use the project in a course. Additionally, we have included a difficulty level for each project based upon how students performed on each project. In general, these projects should be reasonable for senior level undergraduate and introductory graduate level courses. It is of course up to the instructor to decide if the project is appropriate for their class.

Most projects require programming in C++ or C. In general, we advise using GNU C++ for the projects since this is what the students used. Using other programming languages may be fine, but we cannot comment on how compatible it will be with each project.

Most projects contains three parts. The first part is a question and answer section to help prime the students on the project they will perform. The second part is the actual programming part itself. Then there is a second question part to make sure the students understand the concepts they should have learned in the programming parts. Since each project is editable, it is at the instructors discretion what to add, delete or otherwise change. Please note the GNU notice on usage of each document if you wish to change and distribute the project. In general distribution and usage is free so long as credit is given to the original authors.

1. Uninformed Search

Depth First and Breadth First Search	Acrobat Version
	Word Version
Supporting Material
Difficulty:	Easy
Original Course Used In:	CS460, CS561
Instructor:	Laurent Itti
This is a basic project in which students will create a depth first and breadth first search algorithm. The basic idea is to help Homer Simpson get from different places in Springfield to other places in Springfield. This project is not to difficult to complete. However, students seem to get confused on the matter of complexity in computing DFS vs. BFS.

2. Gradient Decent and Simulated Annealing

Simulated Annealing to solve an NP-Complete problem	Acrobat Version
	Word Version
Supporting Material
Difficulty:	Intermediate
Original Course Used In:	CS460, CS561
Instructor:	Laurent Itti
Students are asked to solve an NP-complete problem similar to the N-Queens problem using simulated annealing and must answer questions about gradient decent algorithms. This assignment is good for introducing students to the notion that a program may never completely converge to the correct solution. As such, students must learn when to stop the program. The basic story with the homework is that students must figure out how to arrange baby lizards in a nursery in a way that prevents the lizards from eating each other.

3. Logic Reasoning

First Order Logic Backward Chaining	Acrobat Version
	Word Version
Supporting Material
Difficulty:	Easy
Original Course Used In:	CS460, CS561
Instructor:	Laurent Itti
Students use backwards chaining to attempt to answer simple queries posed in first order logic. In this homework, everything is kept in horn clauses which is why this homework is easy. Students sometimes get tripped up on the if/then concept that if A then B and not A, you can still have B. As such this is a good trick to place in this homework.

4. Fuzzy Logic and Robotics

Use Fuzzy Logic to control a robot fly	Acrobat Version
	Word Version
Supporting Material
Difficulty:	Hard
Original Course Used In:	CS460, CS561
Instructor:	Laurent Itti
This is a difficult homework due to the fact that students must understand basic kinematics in mobile robotics as well as fuzzy logic. The students will create a fuzzy logic controller to move a robotic fly through simple course comprised of goal posts. One drawback to this homework is that students can "solve" the problem using other methods since the homework does not require dynamics. As such, some students try to get away with simply drawing lines. Other students will design non-fuzzy PD controllers, which is also not correct. In order to be sure students implement a fuzzy controller, we check each students code visually to make sure that did what was required.

5. Bayesian Reasoning

Creating a basic Bayesian Classifier	Acrobat Version
	Word Version
Supporting Material
Difficulty:	Intermediate
Original Course Used In:	CS561
Instructor:	Laurent Itti
Students will create a simple Bayesian classifier by taking in sample training sets and learning the best probabilistic decision boundary. In general this homework is not hard, but some students run into conceptual difficulties with probabilities. It is interesting to demonstrate that a large enough prior for a given class can allow it to squash other class which might seem like a more reasonable decision.

6. Neural Networks and Action/Perception

Basic back-propagation neural network classifier	Acrobat Version
	Word Version
Supporting Material1 Supporting Material 2
Difficulty:	Intermediate
Original Course Used In:	CS564
Instructor:	Michael Arbib
Using the training data and NSL, students train a simulated robot to react properly to a visitor based on its features. Students must learn both two and three layer designs for feed-forward neural networks trained using basic back-propagation. With this homework, students must train a robot to learn the difference between werewolves, vampires, adults and children in order to protect valuables in your mansion. The input features drive actions by the robots. For instance, vampire features should produce an impaling action by the robot.

Supporting Material

1. Auto grading with "stubby"

Most of these projects have been designed for large courses with many students. To speed up grading, we use a grading Perl script called stubby. The students are given a basic copy of the grading script in order to make sure that the project they turn in is compatible with the script the graders use.

To make sure the students comply, approximately 6% of the project grade is based on the grading script working properly on the code the student hands in. Additionally, its all or nothing for the points for working with the script since any deviation slows down grading. In general, after the first project, student learn compliance and this becomes an easy 6 points. Additionally, grading becomes easy and far less time consuming.

2. Included grading sheet

Most projects include a grade sheet with a suggestion for grading each project. Ideally, we should take a certain amount of points for each mistake, but weight this by the severity of each mistake made. However, this is of course optional and again just a suggestion.

3. Catching cheaters

We have had success using Moss to catch students cheating on the projects programming components. It should be noted that many students understand how Moss works and will attempt to circumvent it. Thus, cheating will not usually show up as a 100% match. It seems that Moss is good for bringing code to ones attention and from there the instructor should inspect the code to look for patterns.

4. Supporting books and web pages

NSL Homepage - Neural Simulation Language
Rarlab - To open supporting material
Online course material from Laurent Itti
Artificial Intelligence: A Modern Approach
by Stuart Russell and Peter Norvig
ISBN 0-13-103805-2, Prentice Hall (2nd edition).
Neural Networks for Pattern Recognition
by Christopher M. Bishop
ISBN 0-19-853864-2, Oxford University Press.
Statistics
by William L. Hayes
ISBN 0-03-074467-9, Harcourt Brace College Publishers (5th edition).
C++ Black Book
by Steven Holzner
ISBN 1-57610-777-9, The Coriolis Group.