asemanfar - a blog about programming

AI is Hard

March 26, 2008

I just finished a 10-week project course in Artificial Intelligence at my university and I've found that AI is a lot harder than it appears on the second impression. I'll clarify why I said second impression later.

This was the second AI course I've taken, the first was an intro class which turned out to be pretty easy. For this AI project course, we were asked to pick our own topic and make our project proposals and explain how this is considered to be an advanced AI topic from our textbook. My topic was basically a Roomba clone, except it was purely a simulation and no actual hardware was involved. It turned out to perform pretty efficiently compared to the Roomba's demo video. Of course theirs is real and mine is a pure simulation.

The simulated robot only was able to tell whether there was dirt underneath it or not and if any of its 4 sides were blocked by a wall/object. It then stored a representation of its known world internally and did some stuff to figure out where it hasn't been and go there. I don't know if it would work in the real world since I'm not sure how the servos and sensors work and how accurate they are.

Anyway, my first experience with AI led me to believe that it wasn't really "intelligent", just cleverly programmed which isn't far from the truth. I guess that's why it's called Artificial Intelligence, duh. So after the first AI class, I believed that the field of AI was just extrapolating everything in your domain into an evaluation function and implementing an appropriate and efficient search algorithm to search the decision trees of your domain. At least that's the basic AI stuff, there are learning algorithms that don't have an evaluation function (not sure about this) and a lot of other stuff that's beyond my degree level.

So on my second impression, which was in this project course, I discovered that the real challenge in this basic form of AI that we are doing is building an effective and accurate evaluation function as well as representing your domain world correctly and in a manner that supports your evaluation function.

In the end, I discovered AI is a lot more challenging that I originally thought. I'm sure this is also true for a lot of other fields, it's time to stop judging ... topics ... by their ... first impressions? Totally butchered that expression.