Below is an expurgated version of comments I just wrote to a student.
Details that identify the student or topic have been replaced by
brackets <....>. I hope that everyone in my AI courses might find
its suggestions useful.
- Prof. Futrelle
Your excitement about your proposed topic is palpable. But the
bottom line is that your list of "strategies" is couched entirely in
terms of <generating numbers through search, not by reasoning from
Nowhere do you discuss true knowledge-based models of <your topic>.
Even in the Wumpus World, propositional and FOL is used to reason
<not just numbers>.
What's missing is some serious exploration of <your topic>, which
returns nnnK English pages on google. <...> there are quite a number
of books about <your topic>. You need to look at the literature of
<your topic>, instead of making slight extensions of <approaches>
based on <search and numerical scores>.
My first guess about the books and websites on <your topic> are that
the great majority of them will not approach the problem at the level
of sophistication of AIMA. You need to find ones that do.
For example this google search is much more likely to return pointers
to serious work:
"<your topic>" site:aaai.org
Though you may not be able to get the various proceedings of the
meetings mentioned on the aaai site, you can at least get the
wkshp(s) one for $nn. Check out ordering online or call them on the
phone for fast turnaround.
In addition, if you google on the exact title of the <date> workshop,
you get nn hits, many of which are probably papers from the wkshp
that people have posted on their personal sites. But you should
seriously consider ordering the full proceedings.
Google Scholar looks like a good source, with "<your topic>"
returning nn hits, mostly scholarly papers (of course).
I was disappointed that in your zeal to write about your ideas, you
appear to have done no scholarly research to discover serious work on
You may well need to look toward uncertainty techniques such has
Bayes nets, the next topic in the course. These approaches allow you
to build models of uncertain knowledge directly and then use
probabilistic techniques to make decisions based on comparing
computed probabilities of knowledge of states and results. These are
numerical scores in a sense, but they are achieved by a form of
reasoning from knowledge. Check out uncertainty in AI.
Still another subdomain is qualitative <....> reasoning. Still
another is machine learning in <your topic>. You might be able to
apply some WEKA tools as a critical component in building optimizing
knowledge-based memory into your <system>.
As you find serious work, drop me a line with some attached PDFs and
pointers to books, e.g., Amazon, or comments on ones you've skimmed
at B&N or Borders or Quantum Books (Kendall Square). If you do your
research properly, I suspect that I'll be satisfied that you're be
moving beyond numbers to knowledge.
25 favorite books on your topic>
Remember, our interlibrary loan can get you books from all over the
country pretty fast. You should assume that you can get your hands
on most of the books you find interesting. See
Frankly, it's not my job to do the above scouting for you. But I do
it for selfish reasons, so I'll know more, mainly for my students,
and also in the hope that it will spur you to do a thorough and
proper job of background research and that you'll learn from it.
All this said, I tentatively approve your topic on the condition that
you immediately do some serious background research, primarily
focused on approaches that go beyond search-based and <numerical
approaches to your topic>. What's needed is knowledge and reasoning,
- Professor Futrelle
CCIS, Northeastern U., Boston, USA
PS: You'll need to standardize apart the various references to 'nn' above ;-)
<the student's lengthy proposal omitted>