Inductive Logic Programming (https://en.wikipedia.org/wiki/Inductive_logic_programming) find hypothesis theory H for background theory B and set of examples E. ILP algorithms and implementations usually expects, that H, B, E are logic programs - set of Horn clauses and not general FOL or HOL theories. This approach is generalized to the HOL logic programs as well, e.g. in http://andrewcropper.com/pubs/ijcai16-metafunc.pdf.

My question is - are there efforts to formulate induction/ILP for general theories. Apparently, the algorithm does exist and the problem is undecidable, but still - are there some heuristics, some approximate approaches, some more or less rigorous work for such generalization? Both - for full FOL and HOL?

E.g. references wiki articles mentions the method of inverse entailment - I see that that approach is general enough - it requires the computation of the most conscise (e.g. Occams razor principle - with the minimal Kolomogorov or other complexity) set of consequences in some depth.

Actually - ILP may be the Holy Grail of AI: 1) it can learn general policies from the specific policies and hence - implement generalization and transfer learning, e.g. in reinforcement learning; 2) it can learn the program which from background knowledge computes the set of input-output patterns (in more or less general form) and hence - solve the program synthesis task.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.