Method

• This year, I required students to read the chapter before class then used class time to have them answer questions or do short exercises on scratch paper.  The students seemed to learn more from this than lecturing, so I'd suggest doing it again.
• After trying several other approaches, I picked students to answer the questions by bringing a stack of small cards, one for each student. I'd flip over a card, announce the name, and ask the question.  When the stack was exhausted, I'd shuffle the deck and start over.
• Initially, students were embarrassed when they couldn't answer a question in class, so they read the chapters in advance as requested.  Eventually, they got used to the embarrasment, apparently, because fewer students came to class having read the chapter as the semester wore on.  I would suggest making question-answering count about 25% of the course grade.  I'd use a larger stack of cards so that each student name appear about 3 times in the stack.  As questions are put to students, I discard the card into one of three piles: question answered successfully, question answered unsuccessfully and student absent.
• Pop quizes would be also encourage reading the chapter, and would provide valuable assessment data.
• Although it seems embarrassing just to ask the definition of a concept or term, I should do that more often.  Student performance on such qualitative questions on the exams was surprisingly bad.  In general, I should spend less time walking through algorithms and more time trying to get them to see the big picture.  For instance, few students realized that the main reason that POP planning is better than situation state search is that divide and conquer usually works.  Few students could describe the essential idea separating forward from backward chaining.  Few students could say why hiearchical planning is usually better than non-hierarchical planning.
• Doing short exercises in class is good, as it gets them to think. However, it is hard to get their errors out in the open in order to discuss them.  I usually just called on a student to write his or her solution on the board.  This seemed to work better then having students switch papers and critique their partner's work.  It would have been even better if I could have had them put their papers under a camera in order to show the work to the whole class without copying it onto the board--the copying was slow, and it prevents me from having more than one error exposed.

• Because few of the students knew Lisp and Prolog, and I believe that an introductory knowledge of both is just as essential for AI students as an introductory knowledge of Search and Logic, I devoted 2 classes to teaching lisp and 2 to teaching Prolog using a laptop and a projector.  This seemed to suffice, as I did not hear many complaints yet the students finished the assignments.  However, the assignments did not require all that much knowledge of the languages.  I really wish I did not have to use class time for teaching a programming language, especially knowledge of the languages are not on the exams nor on the CS department's qualifying exam in AI.
• In order to make room for teaching Lisp and Prolog, I did not cover the chapters 13, 16 and 17 as I have in past years when I did not teach the programming languages.  I should have cut out game playing (5) and put in 16 (decision theory) instead.  If I didn't have to teach the programming languages, I'd add in two machine learning chapters (18 and 19) as those topics are covered on the CS qualifying exam in AI.
• Students needed much more practice in writing FOL expressions than they got.  Students who knew FOL before the class remained much better at it than students who learned it during the class.  A drill-and-practice program is needed for this skill.
• The prolog planning engine seems to be quite slow compared to UCPop, so it would be best not to use it anymore.
• The projects didn't seem to provide as much learning as I had hoped.  Despite writing Prolog KB and planner KBs, students still made lots of mistakes on the exams that were directly relevant to what the projects were supposed to cover.  It would be better to require 1581 (AI programming languages) as a pre- or co-requisite.  This would be even more attractive if that course had students use Prolog, Klone, Genie and other tools as well as Lisp.