Expert systems seem to have been left at the wayside a little bit in the 21st century. In fact, expert-systems was not even a tag on this site (until I just created it). The traditional focus in expert systems has been on rule based systems and logical resolution via, for example, 2-SAT backward chaining. Have there been attempts to integrate modern machine learning with traditional expert systems theory?
I have been interested in this question and started simple experiments. The idea is to have a basic axiom set that cannot be changed combined with dynamic extension and sensor input. Genetic algorithms add axioms. There can be many sets existing at the same time. In an additional step integration of the sets is attempted by replacing n-ary predicates with (n+m)-ary predicates.
Example: virtual rats learn ways through mazes. There are several mazes and rats can learn any of them. Each maze defines a mapping of sequences of go-left/straight/right decisions to reward/n-reward consequences. The height of the reward also depends on the number of rats that found the way before. The variation is maintained because there is competition, so when a rat variant knowing one maze becomes frequent then knowing a less commonly known maze would become more valuable. In addition to the mazes, there are regularities in the dynamics of the environment, something like: type-1 maze will change to type-2 after 100 runs and change back after 50 runs. This allows the integration of different strategies because switching after frustration is usually good thanks to the relative stability of conditions.
I did not get beyond initial steps before I had to focus completely on a commercial project again. But I would love to resume work on the idea. If you find more information, please let me know: firstname.lastname@example.org