Okay, I've mucked around some more with the heuristics for learning preconditions. And have been able to (under some circumstances) actually get the thing to learn the preconditions that I wanted it to. Problem is that it eventually learns conflicting goals that can never be satisfied. For example, in the run that I just finished it learned that block 2 must be on block 3, and block 3 must be clear.
Started experimenting with learning operator preconditions in earnest today. I have a "semi-working" set of stuff. Here's what I've done:
Things kinda sorta seem to be working, but I need to examine carefully how everything is interacting by running some training trials.
What I want to work on today is getting the heuristics right so that this contraption can actually learn something useful. The current heuristics learn the color and identifier of every single object in the world before starting to examine relationships between things.
Read Aladin Ayurek's means-ends analysis paper over the weekend, and wasn't too shocked by anything in it. One thing that was interesting though was his analysis of chunking over the operator subgoals. It seemed like he'd discovered a non-contemporaneous chunk -- his problem was that two operators were getting composed; e.g., "put block a on b" and "put b on the table" were firing at essentially the same time or something. Since he was doing everything with internal representations, I'm not sure what relevance it would have to "more modern" systems where stuff is all done via external interface. (Hmm, maybe some because you still must do the planning "in your head".)
Started reading a paper by one of Doug Fisher's grad students about the relationship between problem solving and concept formation. I think I might have been "tricked by the title", though. Seems like the big thing that their system does is to examine the most useful of a variety of problem generalizations and try to cluster the most useful ones as basic-level categories. Obviously, it uses COBWEB underneath. I think what I'm looking for is something with the title "applying concept acquisition to problem solving," but I still need to finish reading the paper.
In our meeting last Thursday, John recommended that I start to read through Human Problem Solving to start thinking about some good domains where I could actually apply the precondition learning stuff. There's a lot in there -- I got through about half of the cryptarithmatic stuff (in "skim mode").
The reason that he wanted me to dig into that stuff was that He's not convinced (and I guess I'm not either, now) that blocksworld is an appropriate domain for MEA. I suppose that the obvious toy world for MEA is "robox" world, where you have a robot that's supposed to push boxes between rooms.