The emphasis on rule-based systems in the mid-1980's was a mistaken notion based on experience with mostly toy systems. The work of John R. Anderson was the exception. He continued to evolve tutoring systems that worked off a combination of rules and knowledge structures.
I have been working on programs since August that use some alternate approaches to applying AI techniques. The classic architecture of "generate and test" still works. The major component of the "generate" portion are heuristic rating systems and some just plain "hacks". I quickly found that in my problem domain of choice, that I had to go beyond principles and look at performance with real world data. In conjunction with the heuristic rating systems, I also use some extremely simple machine learning. Those two together seem to be producing adequate performance. I will continue to tweak and test every day, as new ideas occur to me and I see the results of applying what I have.
