HN Explorer

Ask HN: AI research project, need material

4 points by Groxx ↗ HN
For my capstone, I've decided to dive deeper into AI techniques. My main aim is to poke into neural network training techniques and emergent intelligence, especially where it can be applied in an iterative / non-deterministic manner so results can be improved with longer runs, or cut short to get decisions early.

The problem I'm having is that I know a bit about the field, but have very little idea where to go to get good materials to study from. I'm especially interested in newer books, as the field has changed pretty dramatically in the past 15-20 years, but most of the books I can find nearby were written in the early 1980s or older.

I'd be happy to pull from just about any source, be it books, magazines, research papers, blogs, etc. Can anyone point me in a good direction (or 20)?

4 comments

[ 2.8 ms ] story [ 30.1 ms ] thread
Groxx
Right now, I've got:

  The Handbook of Artificial Intelligence, volumes 1 and 3 (1981)
  Pattern Recognition and Machine Learning (by Christopher M. Bishop) (2006)
  Artificial Intelligence - Structures and Strategies for Complex Problem Solving (George F. Luger and William A. Stubblefield) (1998)
  Artificial Intelligence - A Modern Approach (Stuart Russell and Peter Norvig) (1995)
I'm about 1/4 of the way through the last one, and would love to find more in its vein; it's very readable, and has less of a focus on the math than on the logic behind the math. I'm a bit behind most AI-level math, so I'd prefer to minimize it where possible.
arebop
I did something similar for my capstone, and my recommendation is that you start a project. I also started by reading lots of papers from the ACM DL, reading books from my school library, etc. If I could do it over again, I'd start programming earlier.

It sounds like you're trying to learn about AI before you work on applying anytime classifiers in an application. My suggestion is that you interleave the studying with working, so that you can do directed study rather than breadth-first traversal through a huge subfield in CS.

Start with a very modest goal for classification and work on building a neural network for that. Guess about which particular algorithm and feature vector structure to use. Work to improve your results for a while, then go back to reading in order to pick up a new technique to plug in. Consider setting a more challenging goal. Iterate on that.

Groxx
I have a large part of a project finished already, actually. I built a graph-based hexagonal tile system + Settlers of Cataan rules, with the explicit goal of writing computer opponents for it.

I've had a neural networks course, and have written a couple for small tasks, as well as a couple genetic algorithms. What I'm interested in now is learning more of the theories underpinning things.

A longer-term goal is to learn NLP + bayesian filters, and to start fiddling around with email / content classification. It seems the sort of thing I shouldn't have to do 90% of the time, so it'd be nice to automate it.